The most valuable thing a professional produces is not a faster analysis or a better summary. It is insight: a genuinely new way of seeing a problem, a connection that changes how they understand a situation, or a pattern that nobody has named. Think of a strategy consultant who identifies the real competitive challenge behind a client’s margin erosion, or a team member who notices a silence in the data that everyone else has learned to take for granted.

What makes this kind of discovery hard is that expertise, the very thing that makes professionals effective, also makes certain kinds of insight difficult to reach. The frame that lets them see a problem clearly also shapes what they look for — and what they stop looking for. But insights that change thinking are often found at the edges of that frame, such as in the friction between competing interpretations.

Conversational AI is already moving in this direction. A well-constructed prompt can surface competing interpretations, expose gaps, and challenge assumptions. But agentic AI — systems that are configured and directed rather than conversed with — can take users further still. Unlike a prompted conversation that is bounded by what a human supplies and thinks to ask, an agentic system holds more data and sustains analytical orientations across entire data sets without losing the thread. The discovery moves are the same; the depth is not.

Using agentic AI this way demands a professional skill different from both prompting and automation: knowing how to design systems for insight and how to make sense of what they reveal. We call it directing intelligence.

Two Ways of Working With AI

Most professionals encounter AI as a conversation: They type a question, evaluate the response, refine, then ask again. Their job is to supply the context and hold the analytical thread, and the exchange exists only as long as the window is open. The skill this demands is articulation, and specifically knowing what to ask, how to phrase it, and when to push back.

Agentic AI requires a different kind of interaction. Where prompting is reactive (a human asks, the model responds), an agent is proactive: The user configures it, and it operates. That configuration rests on three choices:

• Context (what the agent can access). Where a prompt depends on what is pasted in, an agent’s context is persistent: It’s connected to databases, documents, and records that the prompter chooses and that endure across interactions.
• Capabilities (what the agent can do). A prompted AI generates text, but an agent acts: running analyses, querying databases, comparing data sets, executing multistep analytical routines, and invoking specialized skills and tools without waiting for human input at each stage.
• Orientation (what the agent pays attention to). This moves past an instruction on how to produce a specific output and instead serves as an analytical directive by setting a purpose and trajectory that shape how the agent encounters whatever the data reveals.

The same context and capabilities, given different orientations, will surface different patterns. A single professional can direct multiple agents against the same data set and get genuinely different discoveries, not by asking different questions but by designing different systems. In practice, the professional designs not just individual agents but the system that connects them, often including an orchestration layer that compares and synthesizes across diverse AI agent outputs.

A single, well-configured agent can produce genuine insight. If an AI agent oriented toward customer behavior is given access to transaction and service records, it might discover that churn is concentrated among clients who are in their second year. The larger opportunity emerges when that agent becomes part of a system. Give the same data to specialist agents oriented toward sales conduct, onboarding experience, product usage, and service history, and each surfaces a different explanation for that second-year pattern. Add an orchestration agent configured to compare their outputs, identify where the explanations converge and diverge, and surface the contradictions that matter most, and the system produces something none of its parts could generate alone. The professional’s job is to design this system and evaluate what it reveals.

Four Approaches to Discovery With Agentic AI

Discovery rarely arrives through a single, well-aimed question. It tends to emerge from friction: from putting things in contact that are normally kept apart. Each of the four approaches that follow creates a specific kind of friction: between competing interpretations, between data and the conversations that surround it, between causes and the levels where they hide, and between categories and the reality they were meant to describe. The insight emerges from the friction itself. (See “Four Ways to Direct Intelligence.”)

Each move is also available through prompting. What agentic architecture adds, through what agents access, do, and pay attention to, is depth.

1. Use multiple lenses. The most reliable way to see past a single interpretation is to hold several at once — not sequentially, but simultaneously — in deliberate tension. When multiple well-reasoned frameworks are applied to the same situation and they contradict one another, the contradiction is itself informative. It points toward something none of the frameworks would surface alone.

Consider a strategy consultant engaged with a midsize manufacturer of engineered metal parts serving the aerospace, automotive, and energy markets. Margins have eroded for three consecutive years, and the CEO has blamed it on competitive pricing pressure across all three segments.

The consultant designs an agentic system. Four specialist agents are configured against the full engagement data set, which includes financials, competitive intelligence, customer contracts, interview transcripts, and internal strategy documents. Each agent is oriented toward a different strategic framework, with an orchestration agent configured to synthesize the specialist agents’ analyses.

A Michael Porter agent finds that aerospace is structurally attractive whereas automotive is structurally punishing, suggesting that the segments should not be treated equivalently.

A Jay Barney/VRIO (value, rarity, imitability, and organization) agent identifies a proprietary metallurgical process and aerospace relationships as genuinely rare and difficult to imitate but notes that capital expenditure is split roughly equally across all three segments. The resource advantage is real but the company is not organized to exploit it, because investment is spread evenly across all three segments rather than concentrated behind the aerospace business where the advantage actually lies.

A Richard Rumelt agent argues that the stated strategy is not a strategy at all: “Grow through diversification across end markets” is called out as a goal dressed up as a direction.

A Roger Martin agent maps the “where to play and how to win” decisions and finds them incoherent. The capabilities required to compete in aerospace directly contradict those required in automotive, and the company is attempting both.

No single framework produces a diagnosis. The friction between them surfaces a question nobody has been asking: whether the company should narrow its scope to the one segment where its genuine competitive advantage meets structurally attractive conditions.

Conducting this analysis by prompting a large language model to analyze a situation through Porter, then VRIO, then Rumelt, then Martin will generate real tensions worth exploring. The constraint with such prompting is that the consultant holds the thread. Each framework shift requires a new prompt, and the synthesis depends on the professional’s ability to carry all four interpretations forward simultaneously.

Agentic architecture changes what is possible. Each specialist agent works with the same evidence base at the same depth, so the contradictions are directly comparable. The orchestration agent reviews all four analyses and identifies where the frameworks agree, confirming what is robust (and where they clash) and pointing to what is unresolved. The professional’s role shifts from maintaining the analysis to evaluating what the system surfaces.

2. Surface silences. Some of what an organization knows about itself never makes it into a meeting, a report, or a strategic plan. Surfacing silences means systematically comparing what exists in the data with what appears in the discourse that surrounds it, and treating the gap as analytically meaningful. The insight lives in what is absent.

Consider a strategy consultant engaged with a specialized practice group within a larger professional services firm. The practice does premium, expertise-intensive work in a technical niche that demands deep domain knowledge and long-standing client relationships. The practice leader has never lost a competitive bid, personally oversees every significant engagement, and turns away more work than the practice accepts. The growth narrative is compelling: strong demand and few competitors at the same quality tier.

The consultant configures an agent with the full engagement data set, including transcripts from interviews with the practice leader and company executives, the practice’s strategic plans, market analyses, competitive intelligence, and financial records. The AI agent can systematically inventory themes across both the interview data and the formal strategic documents. Its orientation: Pay attention to what appears in one body of material but not the other.

The AI agent surfaces a silence: organizational dependency on a single individual. The practice leader, a senior expert with two decades of experience, appears as an implicit presence across virtually every topic in the interview transcripts. Growth capacity, client relationships, quality standards, pricing, talent development: Roughly a third of all substantive discussions reference his judgment, relationships, or standards, yet he is never explicitly named. Strategic plans discuss market opportunity, competitive positioning, and investment requirements, but not one of them names what happens when every advantage the practice holds is embodied in one person.

The agent also flags other potential gaps. An apparent absence of succession planning turns out to be addressed in firm-level documents outside the agent’s data set; silence around competitive risk reflects a data gap much more than an organizational one. The consultant investigates and sets both findings aside to focus on the one that holds up: the silence around talent dependency.

Everyone in the company knows that the practice leader is exceptional. But knowing it and naming it as the central strategic constraint are different things. The strategic conversation is organized around market opportunity, but the operational reality is organized around one person. This changes the question from “Can this practice grow?” to “Can it grow without first solving the problem nobody has put on the table?”

Prompting an AI to compare interview themes against strategic plan priorities will flag disconnects, and a skilled professional can push toward what seems absent. However, after weeks or years of seeing things in a particular way, the professional’s perspective becomes constrained in that way, which also shapes how they prompt AI.

The AI agent does not share that immersion in a particular context. It compares the full interview corpus against the full set of strategic documents, without having absorbed the team’s framing. And because the comparison is exhaustive rather than selective, the pattern’s pervasiveness is verifiable rather than impressionistic. The AI can more easily notice what familiarity has made invisible.

3. Bridge levels. When something goes wrong in an organization, the explanation is usually found at the same level where the symptom appeared. For instance, a portfolio problem gets a portfolio explanation, or an operational failure gets an operational diagnosis. But causes don’t always live where symptoms surface. Bridging levels means tracing causal chains across scales of analysis, connecting micro decisions to macro outcomes, and linking macro patterns to the specific behaviors that produce them. The insight comes from traversing the levels that are normally examined in isolation.

Consider a strategy consulting team that is working with a diversified industrial corporation that comprises three divisions and is seeing a declining return on invested capital (ROIC). The corporate narrative is a familiar one: The company is facing market headwinds and competitive pressure. The quarterly review focuses on portfolio-level metrics. Everyone nods.

The lead consultant configures an AI agent with data from three levels of the business (corporate, divisional, and operational) simultaneously, including corporate portfolio metrics and capital allocation records, divisional P&Ls and competitive positions, and operational data within each division. The AI can trace statistical relationships across all three levels. Its orientation: Look for places where a cause and its symptoms sit at different levels.

The agent traces the ROIC decline downward. It is concentrated in Division A, but not because Division A is underperforming within its market. A capital allocation formula implemented three years ago weights recent revenue growth when distributing investment. Division A is the fastest-growing division, but it is also growing in a commoditizing market with declining margins. Division B, slower-growing but carrying the strongest competitive position and the highest margins in the portfolio, is being systematically starved of capital. Its competitive edge is eroding quarter by quarter.

When the ROIC is traced upward, it becomes clear that the capital allocation formula was designed to invest behind growth. At the corporate level, this sounds entirely rational. But at the divisional level, it overinvests in deteriorating competitive dynamics and underinvests in defensible advantage. The ROIC decline the CEO has been attributing to the market is actually being produced by a corporate policy operating exactly as designed.

So here, the symptom lived at the portfolio level but the cause lived in a corporate policy. And the intervention — redesigning how capital is allocated — sat at a level that neither divisional management nor the CEO’s market narrative was examining.

Prompting an AI to explain how corporate capital allocation might be affecting divisional performance, or whether investment patterns align with competitive positioning, will surface plausible hypotheses worth investigating. The constraint is that the prompter has to suspect the connection before they can ask the question.

The agent doesn’t need that prior suspicion. With access to all three levels at once, it can follow connections the consultant didn’t know to look for, because it was configured to work across the levels rather than to test a hypothesis already in hand. The diagnosis was there in the data all along; it was simply invisible from any single level. And the finding can feed the next step: A stress-testing agent can check whether the formula’s measure of revenue growth actually tracks competitive strength in each division or whether that category is itself part of the problem.

4. Stress-test categories. Every organization runs on categories: the classification systems that sort customers, failures, costs, and behaviors into named buckets that route decisions to the right people. The problem is that categories are designed to reflect how an organization thinks, not necessarily how the world behaves. Stress-testing categories means comparing your classification system against the operational reality it was built to describe, and treating the divergence as the finding.

Consider a data analyst at a ready-mix concrete company investigating a persistently high rejected-load rate, where trucks arrive at job sites and are turned away. The company classifies rejections into five categories: wrong mix design, late delivery, quality failure, customer change, and over-order. Each category connects cleanly to a specific department. Wrong mix and over-order go to sales, late delivery goes to dispatch, quality failure goes to the plant, and customer change is marked uncontrollable.

The analyst configures an agent with the formal rejection classifications alongside the full operational record, including order tickets, dispatch logs, GPS tracking data, batch plant records, driver comments, weather data, and customer communications. Its explicit job is to run a pattern analysis across the entire rejection data set and attend to where the formal categories and the actual causal structure diverge.

The agent finds that the categories are both imprecise and actively misleading. Many late-delivery rejections cluster on specific days and weather conditions. Traced backward, they are downstream effects of a batch plant aggregate hopper that slows under adverse weather conditions, delaying loads that then arrive outside the pour window. Categorizing the rejection as late sends the investigation to dispatch, but the cause lives in the plant.

A cluster of customer-change rejections follows a different pattern. They’re concentrated in jobs where a general contractor placed the order but a subcontractor controlled the pour schedule. The label marks those rejections as uncontrollable when they are, in fact, a predictable coordination gap that the company could address. And then there is a pattern for which the formal system has no category at all. Driver comment fields across hundreds of loads record the same informal notation: Site not ready. The truck arrived, but the job site could not receive the pour. Drivers waited, returned, or were rerouted, and dispatchers coded the rejection into whatever category seemed fitting. Here, a recurring root cause existed only in the margins of the data because the classification system was never built to see it.

Pulling a sample of rejection records and asking an AI tool whether the categories are capturing real causes will surface plausible candidate causes and flag obvious mismatches. The constraint is sample size and selection: What the analyst pulls inevitably reflects what already seems worth investigating.

The agent, however, tests every category against the full operational record, across thousands of loads, with no prior assumption about which categories are accurate. It zeros in on weather patterns because it examined everything rather than only a representative slice. It surfaces the “site not ready” pattern because it has no deference to an existing taxonomy. Once those mismatches are surfaced, the broader agentic system can pursue them. A bridging agent could trace the “site not ready” pattern upward from driver comments to dispatch costs to fleet utilization to portfolio-level financial impact, quantifying a problem the formal system had rendered invisible.

How to Direct Intelligence Skillfully

Gaining the most useful results from directing AI agents to surface new insights requires practice, like any new skill. The following are some guidelines to keep in mind.

Configure for discovery, not answers. Start with how you set the system up. The instinct, especially for professionals trained to specify deliverables clearly, is to tell an agent what to find. Resist that instinct. An agent configured to confirm a competitive advantage will confirm it and, in doing so, won’t discover that the advantage is real but being systematically diluted by incoherent investment. A system of agents configured to hold four competing strategic frameworks in tension will produce something harder to digest and considerably more valuable: a contradiction that points at a question nobody was asking. This is problem-setting rather than problem-solving. Humans define what agents should pay attention to and not what they should conclude.

Treat the unexpected as signal, not error. What you do when agents produce something you didn’t expect matters as much as how you configured them. The natural response is to treat the unexpected as error. In efficiency mode, divergence from the anticipated output is waste, and the job is to correct that divergence. In discovery mode, that instinct runs exactly backward. An unexpected output is the most valuable signal the system can generate, because it tells you something about your own assumptions or about the actual structure of the problem you thought you understood. The “site not ready” pattern had existed in the driver comment field for years. The talent dependency permeated a third of the interviews conducted with employees, yet it appeared in none of the professional services firm’s strategic plans. In both cases, the discipline for anyone directing intelligence is to resist the urge to explain that kind of surprise away before investigating what it points to.

Evaluate proposals, not conclusions. The discipline to investigate rather than reject the unexpected carries directly into how you interpret what agents produce. Patterns surfaced by an agentic system are proposals, not findings. They open inquiry rather than close it. When the bridging agent traced ROIC decline to the capital allocation formula, the right response was to test that pattern against alternative explanations and assess whether the effect was large enough to matter. When the Porter and VRIO agents contradicted each other on the strategic situation, the productive move was not to decide which agent was right but to recognize that the contradiction was pointing at a question the company had never confronted. And when the “site not ready” pattern came through, it required verification against GPS time stamps and driver logs before it became actionable. Machine-generated patterns become knowledge through the professional’s judgment, not in spite of it.

Track what you advance, and what you reject. Over time, tracking which proposals you pursue and which you set aside builds something useful: a record of your own analytical instincts, which moves generate value in which contexts, and where your judgment tends to foreclose inquiry before it should. Your rejections reveal as much as your discoveries.

The agentic systems described in this article already involve orchestration: specialist agents whose outputs are synthesized, compared, and set in tension through an orchestration layer that the professional designs. As these systems mature, the orchestration will deepen. A silence-surfacing agent could autonomously trigger a lens-multiplying agent to investigate what it finds; a bridging agent could hand a pattern to a category-testing agent to probe for boundary cases, without the professional initiating each step. The professional’s role could shift from designing individual configurations to shaping the conditions under which agentic systems discover productively. Such moves will compound in ways that are difficult to fully anticipate now.

What won’t change is the underlying skill. Directing intelligence draws on capacities that have always distinguished exceptional professionals: the ability to frame a problem so that new things become visible, and to hold contradictions open long enough to learn from them. Agentic AI does not replace those capacities. It gives professionals a more powerful medium through which to exercise them. The professionals who thrive will not be the ones who automate the most. They will be the ones who understand that discovery is the highest-value thing a professional produces, and who have the judgment to direct intelligence toward it deliberately.

Since early 2025, organizations seeking to curry favor with the current U.S. administration have disavowed earlier commitments to workforce diversity, equity, and inclusion (DEI) and proclaimed that their hiring practices are therefore meritocratic. But framing DEI and merit as antithetical dichotomies, as various executive orders have done, does not mean that abandoning the former results in the latter.

Good DEI practice is designed to broaden the pool from which companies seek to recruit top talent rather than, say, limit their searches to members of reliable old boy networks. Good DEI practice does not replace hiring criteria like demonstrated talent and achievement with demographic criteria. And supposedly meritocratic hiring, in practice, does little to ensure that individuals are selected based on their likelihood of success on the job. Indeed, unbeknownst to many proponents of meritocracy, merit-based practices can be both elusive and illusory.

Merit is a compelling principle. However, many so-called meritocratic practices are fraught with error, even if done with the best intentions. Meritocrats often advocate for practices that they genuinely believe are merit-based but are not — what I call the merit mirage. They may also disparage practices that they believe are not meritorious but are (that is, they demonstrate merit skepticism). Understanding where many attempts at meritocracy go wrong will help leaders better navigate the alleged conflict between DEI and merit.

What Validity Means in Hiring, and Why It Matters

What laypeople call “merit” is referred to as validity by industrial-organizational psychologists who are experts in the technical science behind employment selection. Validity means that the selection practice is job-related: It cannot be assumed; it must be demonstrated according to accepted validation methods. A truly valid hiring process can select higher-performing employees. In contrast, face validity refers to hiring practices that appear valid (or job-related) to the layperson but provide little direct benefit to the organization beyond the appearance of fairness.

Some assessments that lack empirical proof of their validity have face validity and might be passionately supported by individuals who believe in merit due to the merit mirage. Many managers who are too busy to properly vet these types of measures fall prey to the use of such assessments. In addition, applicants who perform well on assessments that have little predictive value truly see the outcome as evidence of their merit, even though such measures are not valid. These types of assessments often advantage majority groups.

Other examples of the merit mirage include the following practices:

• Using educational requirements that are assumed, but have not proved, to predict job performance.
• Relying on “years of experience” requirements but failing to recognize that quantity is not the same as quality, or that after a certain number of years, there are typically significant diminishing returns on years of experience.
• Assessing qualities that are not job-related or are relatively unimportant. Common examples include overvaluing first impressions, nonverbal communication skills, and professional dress. Polished interview skills often do not translate to strong job performance.
• Confusing rigor for relevance. A challenging assessment process is no substitute for one that is more relevant and more comprehensive. Holistic assessment is often more effective than arduous assessment.
• Evaluating group characteristics instead of individual characteristics, such as identity-based stereotypes (including positive ones), or selecting based on the reputations of previous employers or educational institutions attended. Individual merit must be based on individual qualities.
• Using subjective assessments of motives or level of motivation. This is easy to fake; further, motivation to get a job is often incorrectly assumed to predict motivation to perform a job.
• Assuming that a selection practice that is frequently used by others or has shown to be valid for other jobs or organizations is valid for the job you are using it for.
• Incorrect assumptions about what is being measured. A common example is the use of credit checks. One typical belief is that people who have good credit are responsible and those who have bad credit are irresponsible or are more likely to engage in unethical behavior. There is no empirical justification for either assumption. Valid measures of conscientiousness and integrity would better assess responsibility and ethics, respectively, and would be more meritorious compared with credit checks.

The Merit Mirage in Practice

To achieve true merit-based hiring, it is important to use assessments that have proved to be valid/job-related, standardized, and reliable. However, this requires a level of human resource management expertise that most hiring managers do not possess; in fact, even many HR professionals err.

For example, a large organization developed and validated a rigorous and thorough exam for entry-level selection. However, when it started using the exam for making selection decisions, it did so in a manner that differed from how the exam was used in the validation study; the assumption was that it would make no difference because the exam itself had not been modified. After using the exam for 12 years, the organization finally investigated its validity as used and found that it was not valid, because the correlation with job success was not statistically significant. It had wasted more than $17 million administering this rigorous but invalid exam to thousands of applicants. The organization (and thousands of selected applicants) thought selections were based on merit, but they were essentially random — it was a merit mirage.

Perhaps the biggest cause of the merit mirage is a stubborn reliance on intuition and subjectivity, and the incessant use of haphazard selection procedures. For more than four decades, there have been significant advancements in developing selection procedures that can achieve both merit and diversity. Yet these tools are not widely used. One reason is lack of awareness of the tools or how to properly use them. However, another reason is that many recruiters and hiring managers prefer to rely on their own discretion and intuition. They rely on the myth of selection expertise, believing that their knack for identifying merit is more effective than standardized assessments, despite widely available research to the contrary. For example, many interviewers (whether managers or HR professionals) understand that unstructured interviews are ineffective and that structured interviews (which are more valid, reliable, standardized, legally defensible, and less prone to bias) are far more effective. Yet, paradoxically, they are relieved to know that semi-structured interviews (which are moderately effective and allow for some discretion) are another option, and they are strongly inclined to use this option even though it is less effective compared with structured interviews.

The Roots of Merit Skepticism

Other selection methods may be shown, through scientific empirical studies, to possess validity and yet may lack face validity and thus appear invalid to the layperson. It can be difficult to convince managers to adopt these types of measures or to communicate their value to applicants who perform poorly on them. These measures help organizations hire the most qualified applicants, even if such approaches may be opaque and even upsetting to applicants, or viewed with skepticism by managers. These types of assessments typically lessen the majority group’s advantage (which can be misinterpreted as favoring the minority group) and often result in demographic parity.

Merit skepticism often seeks to dismiss personality assessments and biographical questionnaires that surface candidates’ life experiences — but some organizations instead choose to embrace them. For example, a large global company that is very successful, has a strong culture, and is ranked highly on several “best places to work” lists, relies, in part, on personality, values, and strengths assessments in its entry-level selection process. Applicants who are not hired frequently groan about the long selection process and complain about the portions of it that they perceive to be irrelevant or unrelated to the job or to pertinent experience and education. Even though the selection process lacks face validity to applicants, it is valid and instrumental to the organization’s effectiveness. It is consistent with merit principles, even if applicants do not perceive it to be.

Many HR professionals and corporations recognize the value of personality and biographical measures for employment selection because such measures have a long history and robust empirical support. However, the U.S. government has had challenges embracing them. In 2014, the Federal Aviation Administration began using a biographical questionnaire for air traffic controller hiring, but that led to a class-action lawsuit. The validity of the measure is yet to be determined because the case is still pending, but it is notable that most of the critics of the biographical questionnaire attacked it based on its lack of face validity (because laypeople could not see a rational relationship between the items and job performance) rather than its validity (or lack thereof).

More recently, in September 2025, the U.S. Army ended its use of the Command Assessment Program. This was a rigorous, holistic approach to making promotion decisions and included a variety of assessments, including psychological measures. Secretary of Defense Pete Hegseth terminated the program, calling it a failed “woke” experiment. He declared that promotions will now be based only on merit and performance, and the Army reverted to the Centralized Selection List process for making such decisions. That process was first adopted in 1975 and consists of a review of personnel files that lasts only a matter of minutes. In sum, merit skepticism can result in lawsuits and less meritorious solutions.

It’s critical for leaders to promote the use of proper validation methods so that their organization’s candidate selection practices yield genuine merit rather than the merit mirage. Such a rigorous approach may give companies a stronger position to defend against charges of discrimination than flawed merit or diversity initiatives. It requires investment in developing greater HR management expertise so that, for example, assessments can be thoroughly vetted for validity. Perhaps, more importantly, managers need to possess (1) the humility to recognize that their discretion is likely flawed, (2) the openness to use alternative methods, and (3) the wisdom to rely on empirically supported practices.

Many of the selection methods that achieve true merit (such as valid structured interviews, reviews of work samples and biographical data, and some assessments of integrity, and some personality traits) also yield diversity. Thus, it is possible to both maintain DEI commitments and hire based on merit: What some present as two opposite goals can in fact coexist.

Marketing capabilities — the complex bundles of skills, processes, and organizational know-how that enable companies to implement customer-related activities and adapt to marketplace changes — are rated by marketing professionals as important to business success. At the same time, artificial intelligence is rewriting the rules of content creation, customer targeting, and performance measurement. Responsibilities such as managing generative engine optimization (GEO) are emerging as vital online capabilities that did not exist even two years ago.

Eighteen years ago, one of us (Christine) launched The CMO Survey, polling marketing executives about the state of the industry and their organizations. In January 2026, we conducted the 35th edition of the survey, which received 308 responses from marketing leaders at for-profit U.S. companies. The results are troubling.

It’s our assessment that two critical things are happening right now. The first is that the requirements of effective marketing are shifting faster than at any point in The CMO Survey’s history, led by the need to figure out where and how to incorporate AI capabilities. The second is that the state of the marketing profession is not ready for this moment. Instead, marketing teams are systematically undermining their own ability to build the capabilities important to their success.

This is not a minor inconsistency. We see a pattern — visible across budgets, hiring, organizational behavior, and strategic priorities — that raises fundamental questions about how companies today are making decisions about investing in marketing capabilities. Our prediction is that this soft commitment to what we believe is critical marketing know-how will cost organizations competitively if it is allowed to continue as it is today.

Underinvestment Looms Large

Nearly 60% of marketing leaders reported that their primary approach to building marketing capabilities is through internal training and hiring rather than via external partnerships or acquiring other companies. This preference hasn’t changed since we last asked about it in 2020, despite six years of dramatic shifts in what successful marketing requires. And despite this reliance on internal resources — and surveyed leaders’ belief that marketing capabilities are important to business success — getting financial support for capability building remains a challenge.

In particular, when we look at what companies are actually doing to support that stated commitment to training and hiring, the details tell a story of underinvestment. Training and development budgets have declined steadily for years and now stand at just 3.8% of marketing spend — down from a pre-pandemic high of 5.8% in 2019. Marketing head count growth has dropped sharply, falling more than 50% from last year’s rate. And when asked what capability is most lacking in their organizations, the most common response from marketing leaders was not a skill deficit but inadequate resources: not enough people, time, or budget to make existing capabilities function effectively.

Strategic intent and resource allocation point in opposite directions. Companies say that they build their capabilities through people, but they are systematically reducing or slowing their investment in those same people. We call this disconnect the marketing capability paradox.

A closer look at agility and skills investment illustrates this contradiction. Seventy-one percent of marketing leaders in our survey said that agility is key to their organization’s success. Marketers reported performing reasonably well at this, able to quickly revise priorities and shift resources in response to change. But at the same time, they reported that their weakest-rated activity across all agility dimensions is “building the capabilities that facilitate agile marketing actions.”

Companies see themselves as good at reacting to change, but they are weaker at building the organizational foundation that would make those reactions less costly and more effective. This distinction — between responding to the present and investing in the future — is a troubling signal that runs through every dimension of how marketing capabilities are managed in organizations.

Seven Barriers to Capability Building

Cuts in training and declines in head count are the most visible symptoms of this paradox, but other data hints at larger structural challenges. From our survey results, we identified seven interconnected forces that are working against capability development. Understanding them together reveals why this problem is so resistant to easy solutions.

There is a serious gap between the adoption of technology and marketing teams’ preparedness for using it. Technology adoption is outrunning organizational readiness. Companies are investing in technology faster than they are building the human capacity to effectively use it, which means each new tool widens the capability deficit rather than closing it. One survey respondent succinctly described their biggest barrier to maximizing the impact of marketing technology as “attracting the right talent, retaining them, and keeping them up to speed on changes happening.”

Asked to rate elements of their marketing technology activity on a 7-point performance scale (where 1 is going “poorly” and 7 is doing “very well”), survey respondents didn’t give even one activity a score above 5. Performance levels have not improved over the past two years, even as marketing departments have scaled up their deployment of AI tools. The two lowest-rated activities — hiring employees to manage marketing technologies (3.7) and training employees on emerging marketing technologies (3.9) — are the very ingredients needed to close the gap between deployment and effective use of technology.

This is not a minor issue. AI use in marketing activities has nearly doubled since 2024, from 13.1% to 24.2% today. Respondents projected that AI will account for more than 50% of all marketing activities within three years as it increasingly delivers measurable improvements in sales productivity, customer satisfaction, and marketing overhead costs. Gains from AI are rising year over year, but the people power to keep pace is not.

Too many marketers have a structural orientation toward the present. Every year since 2019, marketers have reported in our surveys that they devote roughly 68% of their time to managing the present and 32% to preparing for the future. This ratio has held constant across the COVID-19 years, the digital transformation era, and now the AI revolution. This tells us that the ratio is not a situational response to any particular economic pressure but a structural orientation that is remarkably resistant to change.

But capability building is inherently a future-oriented investment. It requires sustained attention, a multiyear commitment, and a willingness to accept near-term costs for long-run returns. In an environment where organizations have consistently prioritized the present over the future for the past seven years, that kind of investment cannot take root. As one respondent noted, “There is little time for future thinking. We have a strategy, but our actions are tactical and short term, making it difficult to prove effectiveness.”

Current economic pressures are intensifying a condition that predates them and have pushed 70.6% of marketing leaders toward short-term impact over long-run gains and 26.8% toward explicitly emphasizing spending over building capabilities, according to our survey data.

Marketers aren’t highlighting strong data around impact and retention. Our survey found that marketing’s impact on customers is growing more durable, meaning that payoffs are being felt further out in time than previously was the case. The median duration of marketing’s impact has lengthened from several months in 2022 to six months in 2026, with a meaningful shift toward effects lasting one year or longer.

In addition, customer retention, the metric most directly linked to the depth of marketing capabilities, is growing at 12.8%, well above the growth of customer acquisition, at 7.4%.

In spite of this, marketers continue to focus on short-term impact. This is a missed opportunity: The cumulative value of sustained marketing investment may be greater than short-term measurement approaches currently capture.

Ties with the C-suite are weak. Building and sustaining marketing capabilities requires organizational buy-in and support that extends beyond the marketing function. Yet the partnership between marketing leaders and CFOs, measured on the 1-7 scale, stands at just 4.5 for building a business case for marketing spending — barely a change from 4.3 in 2021. Without a strong CFO relationship, marketing leaders will find it difficult to make a credible internal case for capability investment. This is key to why training budgets and head count suffer.

The result is a vicious cycle. A weak CFO partnership leads to underinvestment in capabilities, which weakens marketing’s ability to demonstrate value, which further erodes the CFO relationship. This deficit also shows up in weak collaboration with technology leaders: Only 55.7% of marketing leaders reported collaborating with their CIO or CTO on digital activities. That leaves the rest without the technical partnership that adopting AI tools increasingly demands, especially if they’re to be adopted at scale.

A rigid build-versus-partner mindset stifles development. AI-related capability gaps spanning analytics, generative AI, GEO, and demand generation were cited by a combined 35.7% of marketing leaders as their most pressing unmet needs. These are precisely the areas where external partnerships and acquisitions might accelerate capability building more effectively than internal development. Yet the build preference persists, even as the resources available to execute it continue to decline.

It’s genuinely surprising that despite six years of dramatic change in what marketing capabilities are required (particularly the growing centrality of AI, analytics, and technology skills), how companies approach capability development has remained unchanged since we started conducting our survey in 2020. The overwhelming preference for building capabilities internally through training and hiring — the choice of nearly 60% of our surveyed marketing leaders, as mentioned earlier — was a reasonable strategy when capability requirements were relatively stable. It is a much more problematic strategy when the technology capability landscape is shifting as rapidly as it is today.

The foundation is too weak to sustain new initiatives. Perhaps the most telling finding in this year’s survey is that the most cited capability gap is the inadequate resourcing of existing capabilities — not enough people, time, or budget to make what companies already have function effectively. Companies are not failing to build the right capabilities; they are failing to sustain the ones they already have. The implications are significant: Even if companies were to invest in the right new capabilities, they would be building on a foundation that is already crumbling.

Marketing’s essential purpose has a framing problem. How marketing leaders think about the value of skill building may be contributing to their underinvestment. Nearly 4 out of 5 surveyed marketers (78.2%) said capabilities matter primarily because they deliver higher ROI for every marketing dollar spent. Fewer named reasons that are more strategically durable, such as increasing the effectiveness of managing customers (cited by 52.1%), making it more difficult for competitors to imitate them (37.0%), or attracting and retaining top talent (33.3%). When capabilities are justified primarily on short-term ROI grounds, they will always lose the budget argument to investments that demonstrate returns more quickly and more visibly.

The deeper strategic case for investing in marketing capabilities, including the argument that they build organizational resilience, create competitive barriers, and compound in value over time, is largely absent from how marketing leaders think about capabilities and communicate them to others. This framing problem affects more than the conversations marketers have with CFOs. It shapes how marketing capabilities are more broadly prioritized internally, how they are resourced, and how quickly they are cut when financial pressure mounts. Leaders who cannot articulate why capabilities matter beyond their immediate financial return will find it difficult to protect them.

What Companies Should Do

What makes this situation particularly concerning is that the seven forces are not independent. They reinforce one another in ways that make the capability deficit self-perpetuating. Short-termism reduces the organizational appetite for capability investment. Reduced capability investment weakens marketing’s ability to demonstrate value. A weakened ability to demonstrate value increases pressure from CEOs, boards, and CFOs — which drives more short-termism. Meanwhile, technology adoption accelerates the demand for corresponding capabilities; the CFO partnership remains too weak to fund a response; the build strategy persists even as its resources erode; and the framing of capabilities as an ROI tool rather than a strategic asset ensures that they will always be outcompeted for budget by investments with more immediate and visible returns.

Our survey data suggests three priorities for marketing leaders and their organizations.

First, executives across the C-suite should wake up to the need to decouple capability investment from short-term financial pressure. Marketing training budgets and staffing are currently treated as variable costs that can be cut when profits disappoint. Indeed, our surveyed executives reported that marketing expenses are cut 45.4% of the time — more frequently than other areas of the company — when executives face profit shortfalls. But treating capability investment like discretionary spending mischaracterizes its nature entirely. Capability investment is analogous to R&D: Cutting it saves money in the short run while compounding the capability deficit in ways that are costly to reverse. Companies that protect capability investment through economic cycles will be better positioned when conditions improve.

Second and relatedly, marketing leaders need to reframe the case for capability investment. Marketing leaders who justify it primarily on ROI grounds are making their own investment case harder to win. The more compelling and durable argument connects capability investment to competitive advantage, organizational resilience, and the ability to attract and retain the talent that the AI era requires.

Third, markers must revisit the build-versus-partner assumption. The stability of the build preference across six years of surveying in an age of dramatic change in marketing’s requirements is striking. For AI-related capabilities in particular — where the speed of change and the specialized knowledge required make internal development especially challenging — a more aggressive partnering strategy is likely to be more effective. This does not mean abandoning the build approach, but it does mean subjecting it to the same scrutiny applied to any other strategic assumption that has not been tested in years.

The capability paradox is not just an operational problem but also a strategic one. Companies are underinvesting in the very capabilities that the evidence suggests are generating their most durable and valuable results. In a competitive environment being reshaped by AI, where the capability gap between leaders and laggards will widen rapidly, that is a costly mistake to make and a difficult one to reverse.

Over the past few decades, there has been burgeoning interest in how leaders and organizations can support employees in the aftermath of traumatic events.1 But there is another pervasive and often hidden source of trauma that originates well before individuals enter the workforce. Childhood adversity, referred to as adverse childhood experiences (ACEs), is widespread and affects the well-being of a large portion of the workforce.

Recent estimates suggest that 60% of adults worldwide have faced one ACE, and 16% have faced four or more.2 ACEs shape employees’ self-perceptions, job attitudes, behaviors, and workplace relationships, typically in ways that undermine their well-being and functioning. In light of this, we are sharing a new, evidence-based toolkit for organizations and leaders seeking to design healthier work environments that can attenuate the enduring effects of childhood adversity for their employees.

How Childhood Adversity Can Resonate in the Workplace

ACEs encompass a range of early-life hardships. They include interpersonal problems, such as abuse or neglect; household dysfunction, such as divorce, domestic violence, criminality, or substance abuse; the death of or separation from family members; and witnessing or being the victim of community violence. ACEs also include physical or mental health issues in the home and economic adversity. In short, ACEs cause both direct and indirect harm to children. Such situations are typically experienced as harsh, unpredictable, dysregulating, and shaming, and are characterized by a lack of agency and control.

ACEs lead to psychological and physiological changes that persist into adulthood. They include blunted cortisol responses to stress, or the body is not releasing cortisol as needed; heightened reactivity of the amygdala, which leads people to perceive threats where there are none; and diminished hippocampal gray matter volume, which is linked with memory impairment, cognitive and emotional difficulties, and mood disorders.3

The workplace implications of these changes mean that ACE survivors are more likely to experience low self-worth, risk aversion, turnover intentions, emotional distress, job strain, low job performance, and dismissal and are often the targets or perpetrators of workplace mistreatment. These outcomes can leave adults feeling chained to their traumatic pasts.4 As American author William Faulkner wrote, “The past is never dead. It’s not even past.”

Despite the potential for poor outcomes for ACE survivors, neuroscience research has found that even adult brains can change as people learn and grow — a phenomenon known as neuroplasticity.5 While scientists once thought that the brain was largely fixed after childhood, they have since determined that the adult brain remains dynamically adaptable and is capable of forming new neural connections, strengthening or weakening existing pathways, and reorganizing functional networks. The brain can update prior learning through repeated, meaningful adult experiences that counter earlier, adverse experiences. Importantly, this means that ACE survivors’ professional experiences can themselves offer a platform for recovery. How work experiences might facilitate neuroplasticity in the brain to help people overcome the deleterious effects of ACEs is a key question we hope to answer for leaders concerned with employee well-being.

Critically, whether neuroplasticity occurs depends on the quality of an experience. It is most strongly promoted by situations that are emotionally meaningful, have a social/relational component, and repeatedly reinforce new learning, thinking, or habits. Research also shows that predictable routines, supportive social relationships, opportunities for autonomy and mastery, and experiences that reduce chronic stress all facilitate adaptive neural change. They do so by dampening hyperreactive threat responses, strengthening regulatory and executive functioning, and restoring the healthy neuroendocrine functioning needed to regulate stress and mood.

Over time, these healthy conditions enable individuals to replace survival-based responses learned in adverse and threatening environments with more flexible, regulated, and agentic patterns of thinking, feeling, and behaving. Neuroplasticity is thus a practical mechanism through which the workplace, often one of the most stable and recurrent environments in adulthood, can meaningfully contribute to recovery, resilience, and long-term well-being among ACE survivors.

It is important to acknowledge that claims about neuroplasticity are sometimes met with skepticism, particularly when they are interpreted as suggesting that the brain can be easily or rapidly “rewired” or that early adversity can be fully undone through later experiences. The evidence does not support such a simplistic view. Rather, neuroplasticity reflects a gradual, experience-dependent process that unfolds over time, varies considerably across individuals, and is shaped by factors such as chronic stress, developmental timing, and the consistency of environmental input. Importantly, plasticity does not imply erasure of earlier learning but, rather, the capacity for new neural pathways and regulatory patterns to emerge alongside the influence of earlier adaptations — and, in some cases, attenuate them. Framing neuroplasticity in this way situates it not as a promise of quick transformation but as a scientifically grounded mechanism through which sustained, supportive experiences, including those in the workplace, can incrementally support adaptive functioning among ACE survivors.6

How Healthy Workplace Experiences Lead to Positive Change

Neuroplasticity is facilitated by experiences that contradict what ACEs have led survivors’ brains and nervous systems to expect. Benevolence and stability can shift the core belief that others, especially authority figures, are harsh and unpredictable. Relationships through which survivors learn to better manage their emotions and behavior, and through which they receive validation of their feelings, can counter emotional dysregulation and shame. Those who have felt helpless and unable to act can gain confidence by being empowered to claim agency and control through their work.

Importantly, the kinds of experiences that may be seen as therapeutically corrective are recognizable hallmarks of organizational cultures that many healthy companies now strive to foster. What promotes healing and recovery for survivors of ACEs is also increasingly seen as promoting employee well-being broadly, leading to greater engagement and higher performance. Leaders and organizations play a crucial role in all this. In what follows, we’ll explain specifically how and why such positive workplace experiences can be helpful.

Reduce anxiety, fear, and defensiveness. Children who are targets of or witnesses to violence or conflict in the home may have caretakers who lack healthy emotional regulation and self-control, and they may be punished when they voice their own emotions.

Experiences with this type of caretaker as a child can foster many negative outcomes in adulthood, including a greater likelihood of being the victim or perpetrator of workplace violence or physical abuse. Survivors may lack the skills needed to process feedback from, or even simply disagree with, others in their organizations. Many adults with histories of childhood trauma struggle to regulate complex thoughts and emotions, especially anger or hurt, that were not safely processed earlier in life. As a result, they may respond defensively, reactively, or with hostility when they do not understand others’ decisions or opinions. Feedback — even when it is constructive, helpful, and/or warranted — can be threatening and difficult to process.7

Such responses originate in changes to the brain caused by repeated exposure to volatility, conflict, or punishment. The neural systems responsible for detecting threats, particularly the amygdala, are sensitized, while regulatory systems in the prefrontal cortex that support reflection, impulse control, and emotional regulation are weakened. In parallel, chronic activation of the body’s stress response system alters cortisol signaling in ways that bias the brain toward hypervigilance or shutdown. Over time, these adaptations can become default responses, such that even neutral or mildly challenging interpersonal interactions in adulthood are unconsciously experienced as dangerous, triggering fight, flight, or freeze reactions rather than thoughtful engagement.8

When adults later encounter environments that are regularly calm, stable, and emotionally consistent, these same neural systems are exposed to a markedly different set of inputs. Repeated experiences of kindness and clarity, such as measured feedback, transparent expectations, and respectful disagreement, reduce the frequency with which the brain’s alarm system is activated, allowing prefrontal regulatory regions to exert greater influence over emotional and behavioral responses. Over time, this shift facilitates neuroplastic change: Stress-related circuits become less dominant, while neural pathways associated with safety, trust, and cognitive control are strengthened. Importantly, these updates occur not through isolated moments of support but through sustained patterns of predictable and compassionate interaction that enable the brain to revise earlier threat-based learning.9 Thus, these relationships between colleagues at work are critical to helping ACE survivors learn to more effectively process interpersonal interactions, particularly those involving criticisms or conflict.

Foster self-worth and healthy boundary-setting. Children who grow up with emotional abuse or neglect often lack support for developing healthy boundaries and blame themselves for what they have suffered, which erodes their sense of self-worth. When distress is ignored, punished, or met with criticism, neural circuits responsible for soothing arousal and integrating emotional experiences are underdeveloped or inconsistently engaged. As a result, emotional states may feel overwhelming, confusing, or threatening in adulthood, often triggering excessive self-blame, heightened sensitivity to feedback, or compulsive efforts to seek reassurance.10

In adulthood, this can give rise to a constant need for external validation from others in the workplace, such as managers, coworkers, or customers, and individuals’ inability to separate their self-worth and self-esteem from their job. Often, these employees have trouble setting boundaries or saying no at work, which can lead them to take on more than they can shoulder. In turn, they are left feeling overwhelmed or experience sudden burnout.11 Their sense of self is fragile, and they tend to place the needs of others above their own to an unhealthy degree.

A different pattern emerges when individuals repeatedly experience co-regulation and healthy validation in relational contexts, including at work. When colleagues or leaders respond to distress by acknowledging emotions without amplifying or dismissing them, the individual’s brain is afforded repeated opportunities to practice regulation in connection with others. Over time, these empathetic interactions strengthen neural pathways that integrate emotional awareness with cognitive control, supporting individuals’ greater tolerance of internal states and allowing them to engage more adaptively with boundary setting, uncertainty, and interpersonal demands at work.12

Restore agency and control. When children who experience adverse events learn that they lack control over their situation, their motivation to try to exercise control diminishes — a conditioned response called learned helplessness.13 This pattern can continue into adulthood, even showing up in situations where control is possible. Adults who have survived ACEs may struggle with seeing clear paths forward when encountering challenges at work and are more likely to demonstrate cognitive distortions psychologist Martin Seligman dubbed the “3 P’s” of learned helplessness: personalization (blaming themselves for setbacks rather than taking into account external factors); permanence (believing that the setbacks reflect an unchangeable reality); and pervasiveness (generalizing from one negative outcome to predicting failure in all parts of their lives). Brain changes associated with lack of control or power can bias an individual’s attention toward potential risks, errors, and negative outcomes, promoting caution, withdrawal, and behavioral freezing.

These patterns are not fixed. When workplaces offer meaningful autonomy, amplify employee voices, and give them influence, the brain receives repeated signals that effort can shape outcomes. (A considerable body of research also links autonomy on the job to intrinsic motivation and employee well-being generally.) Under these conditions, inhibition systems quiet, stress hormone levels decrease, and neural resources shift toward approach-oriented networks that support motivation, learning, and adaptive problem-solving. When an individual has sustained experiences of agency, neuroplastic changes recalibrate the balance between vigilance and engagement, reinforcing a sense of personal efficacy and control that was often unavailable to them earlier in life.14

Management Imperatives for a Healthy Workplace for All

Managers are unlikely to have insight into employees’ childhood experiences (nor should they ask anyone to disclose them), so the suggestions we offer below are not targeted interventions. Rather, they reflect healthy work design and leadership principles that we argue can help the many people whose well-being has been affected by ACEs. We offer five recommendations.

Establish norms around respectful interaction. Leaders can set expectations for how people in the workplace interact by consistently modeling respectful engagement: being present, listening attentively, and communicating in ways that convey respect and regard for others’ worth. Furthermore, they can explicitly articulate that this behavior is expected and make consideration and mutual respect a dimension of performance.

Design work to help others build agency and influence. Regularly soliciting ideas and feedback from employees can not only foster better decision-making but also help employees build self-efficacy and a sense of influence. Leaders can further reinforce agency by ensuring that each employee has ownership over a clearly defined project or area of responsibility.

Model emotional steadiness. Leaders can shape emotional climates by demonstrating calm, measured responses in high-pressure situations. When leaders manage their emotions effectively in public settings, they demonstrate that stress can be navigated without toxicity or blame, creating the conditions for co-regulation.

Show appreciation and respect publicly. When people feel as though their contributions are significant in the eyes of others, their sense of self-efficacy and performance increase. Leaders can convey worthiness and affirmation by highlighting others’ contributions, strengths, and responsibilities in visible and specific ways.

Use strengths-based feedback to expand employees’ sense of possibility. By taking a developmental approach to performance that builds on employee strengths rather than seeking to correct shortcomings, leaders can help employees shift to describing themselves with more positive, empowering, and secure internal narratives.

Making change does not necessarily require grand gestures or large-scale initiatives. Rather, small efforts — micro-behaviors embedded into everyday organizational functioning and repeated over time — are the key to building a healthier workforce. Engaging with employees empathetically, respectfully, and supportively requires awareness and sensitivity but not time or money. For survivors of childhood trauma, these everyday interactions can play a critical role in rewiring stress responses and steadily repairing the lasting effects of adversity.

The online revolution enabled frictionless access to virtually every type of content — including adult content. Generative AI has added a whole new, and potentially more dangerous, dimension to the boom in explicit content. In the absence of strong safeguards, AI users can transition from consumers to creators of their own explicit media with shocking ease and speed. National governments around the world as well as the U.N. have issued grave warnings about wide-scale production of AI-generated illegal deepfakes, including images involving minors.

Yet Big Tech does not seem to be responding with the same level of seriousness. Earlier this year, Sam Altman, CEO of OpenAI, announced his intention to introduce an “adult mode” within ChatGPT that would allow users to carry on X-rated conversations with the chatbot. The ensuing furor led to the idea being sidelined — for now. Yet Altman’s moment of candor signals that tech leaders are potentially still toying with adult content as part of AI’s value proposition.

Our ongoing research suggests that the human consequences of such a decision may be greater and more direct than the “tech bros” expect. Opening the floodgates to AI-generated adult content could, in fact, invite sharp increases in severe sexual criminality.

We took advantage of a natural experiment in response to a schism in the scholarly literature. Some psychology researchers theorize that fantasies in online environments have a cathartic effect for potential criminals, keeping their transgressions confined to the digital sphere. But an opposing strain of social learning and stimulation theory has it that online fantasies stoke rather than satiate the desire to commit sexual crimes.

We had a unique opportunity to explore this question due to the fact that, unlike many other countries, Japan’s National Police Agency (NPA) publishes detailed monthly crime statistics, including categories of offenses and ages of victims.

We accessed NPA statistics for the period January 2022 to October 2024. Smack in the middle of this time frame was the mid-2023 Japanese launch of ChatGPT’s mobile application. We treated that launch as an external shock that materially reduced access frictions in a mobile-first market. (In 2023, 72.9% of internet users in Japan accessed the web via smartphones.) That same period also saw major improvements in image-generation tools, including a July 2023 Stable Diffusion update that made high-quality image creation easier for novice users. So it made sense to use mid-2023 as a technological turning point for before-and-after comparisons.

Our analysis found that Japan experienced a significant increase in reported rape cases after mid-2023, with the monthly national average rising from 148 to 265 between the first and second halves of our observation period. Using difference-in-difference techniques designed to create clean comparisons between rape as the “treated” offense and murder as a “control” offense (chosen because both are classified by the NPA as serious violent crimes and follow broadly comparable institutional processes), we found that the increased incidence of rape consistently outpaced that of murder over the same period by approximately 12 cases per month. Before mid-2023, there was no statistically significant difference in the monthly fluctuations of the two types of offense. The same basic pattern held when we compared rape to the other offenses Japan classifies as serious violent crimes that are not necessarily sexual in nature, such as arson and robbery.

Still, it could have been a case of correlation rather than causation. Viewed purely in terms of the timeline, the unfortunate spike in rape cases could have been related to all sorts of social factors unobservable in our data. To strengthen our analysis, we obtained data from one of Japan’s largest and most popular online art communities, where users can upload, view, and vote on publicly shared artworks. The community attracts approximately 3.6 million monthly visitors and roughly 2.9 billion monthly pageviews. The platform’s primary language is Japanese, and external web traffic estimates indicate that roughly two-thirds of visits originate from Japan (whose population numbers approximately 130 million). Given this scale and concentration of Japanese users, the platform provides a meaningful window into national-level shifts in online behavior. Importantly, the online community does not prohibit images that are adults-only and/or algorithmically generated, but community members are empowered and encouraged to tag images as such, whether or not they created or uploaded the image.

We discovered that periods of peak engagement (likes, bookmarks, and views) with material tagged as AI-generated and adults-only were also those in which the increases in rape relative to murder were particularly high. This statistical relationship manifested only after mid-2023, when generative AI became broadly available in Japan.

The totality of this evidence points to a scenario in which AI-generated explicit content directly contributed to a rise in sex-related crime. Rape is one particularly heinous example of such a crime, but there is no reason to expect that this scenario would not also apply to other sexually motivated offenses.

The implications of our results became especially alarming when we narrowed the focus to content and criminality involving minors. Engagement with explicit images that were tagged by users as featuring minors was associated with an even larger increase in rape relative to murder. When we zeroed in on rape cases with victims under the age of 20, the estimated amplification was larger still. Although these results do not establish a one-to-one pathway from specific images to specific offenses, they are consistent with a disproportionate increase in offenses involving younger victims. Subsequent studies clarified that the additional crimes mainly consisted of adults offending against minors rather than minor-on-minor crimes.

Our Japan-specific findings appear to dovetail with patterns noted elsewhere. For example, observers were shocked and stymied when the CyberTipline, a reporting system for online child sexual exploitation run by the National Center for Missing and Exploited Children in the U.S., announced that in 2023, reports had gone up 12% over the previous year. The figure’s correlation with the timing of the diffusion of AI tools for the general public is highly suggestive of a connection in line with our findings.

To be sure, our research merely documents an apparent causal connection. Our findings do not reveal exactly how the dissemination of, and engagement with, AI-generated adult images fueled criminal behavior, nor can we state with confidence that these images were solely responsible for the increase. Certainly, we do not claim to have definitively settled scholarly disputes over whether explicit online content acts as a catalyst for criminal behavior or facilitates an essentially harmless cathartic release.

However, the context-specific connection we have identified is serious enough to justify concern about Big Tech’s inclination to integrate explicit content into popular AI models. Against the seeming rationality of Altman’s avowed aim to “let adults be adults,” we must juxtapose the probable impact on real-world sexual crime, especially as it pertains to the most vulnerable among us: underage victims.

So far, the most aggressive regulatory attention in this area has targeted “revenge porn” deepfakes, as is the case with the TAKE IT DOWN Act signed into law by President Donald Trump in 2025. When it comes to AI-generated explicit content that does not depict identifiable individuals (including content involving minors), current legal frameworks have not caught up to the generative capabilities of AI. As of this writing, for instance, European Union legislators are mired in debate, weighing online privacy concerns against the obligation to confront the sexual abuse of minors, the result of which is an enforcement impasse. While societies work out how to respond to newly emerging threats, websites and apps for the express purpose of producing explicit content have proliferated.

Authorities and policy makers are not helpless, however. Bare-minimum, relatively uncontroversial safeguards that could be brought to bear in the short term include these options:

• Watermarking and provenance tracking, which could assist in holding providers accountable by linking content to the platform or tool that created it.
• Age verification, which, if rigorously applied, could at least restrict minors from taking part in the production of explicit material.
• Tighter enforcement of laws that have already been adopted by 45 U.S. states, such as those criminalizing online child sexual abuse material.

Steps such as these would be an appropriate starting point. But the scale of the issue — and the real-world implications for criminal behavior revealed by our research — demands that lawmakers redouble their efforts to close legal loopholes that allow AI-generated explicit content to propagate unabated. And companies would be wise to consider the reputational harms of being seen as part of the problem.

“The ChatGPT moment for robotics is coming,” declared Nvidia CEO Jensen Huang at the Consumer Electronics Show in January 2025. It’s a widespread expectation: that humanoid robots will follow the same explosive adoption curve as generative AI. Our research suggests the opposite. Humanoid robotics will be adopted unevenly, across diverging use cases and geographies. Leaders need to reset their strategies for a more complex, jagged path to scale.

Humanoid robots are developing at pace. Engineering advances and new powerful models of physical intelligence, including vision-language-action models and improved locomotion systems, are converging to create a robotics super-cycle. But leaders who expect the path of humanoid adoption to mirror that of generative AI are misjudging the strategy required.

A Less Predictable Pace

Our research shows that three fundamental forces are shaping a very different adoption curve.

Force 1: No One-Size-Fits-All Model

Some humanoid manufacturers envision general-purpose robots working across multiple activities and industries. But for the foreseeable future, humanoid robots will not be commercially deployed as universal productivity tools.

Consider the multitude of roles that humanoids could theoretically take on: carer, guard, soldier, inventory picker, assembly line worker, hospital assistant, or concierge, to name but a few. Each role, and often each separate task of that role, imposes distinct and often incompatible requirements.

A warehouse robot, for example, may need to lift up to 132 pounds, requiring sophisticated actuators — the equivalent of human muscles — to complete its task successfully. Those actuator systems will make up 40% to 60% of its cost. But a care assistant robot requires different capabilities: subtle facial expression, fine motor control, and emotionally sensitive interaction. In this case, the technology build will tilt more heavily toward perception and haptic technologies. The diversity of humanoid roles implies fundamentally different technology build specifications, operating requirements, and ultimately, business cases for deployment.

This divergence extends beyond the physical build into software and connectivity. A security robot will use low-latency edge-computing data processing close to its physical location — so it can rapidly respond to environmental changes, such as an incursion by an unknown person or vehicle. By contrast, a humanoid hospital assistant may rely on large graphical models to navigate hospital facilities and create 3D images of patient charts and X-rays. Each scenario requires different computing architectures. This specialization, in turn, supports an array of humanoid variations, such as fully mobile bipedal robots or wheeled torsos, reflecting the different physical demands of each role.

The result is the de facto specialization of competing humanoid manufacturers, each prioritizing different market niches. Analysis of these companies shows heavy focus by market segment — manufacturing, warehousing, care, service, and home — in market strategy and customer cases.

Humanoid robots will therefore come in a variety of shapes and specifications. They are being built as specialists, not generalists, and will be scaled role by role.

Force 2: Geographic “Islands of Viability”

Robot costs are dropping steadily, but every humanoid deployment must be justified with a clear return on investment — a determination that will vary sharply by location due to workplace regulations, labor market dynamics, and demographics, among other factors.

Consequently, early humanoid business cases will be concentrated in a few “islands of viability” where conditions are ripest for first deployments. ROI will be strongest in locations that have high labor costs and acute workforce shortages, or that have existing infrastructure and potential for immediate mass-market scale. There will be fewer compelling business cases in parts of the world with abundant and cheap labor, or limited humanoid manufacturing ecosystems.

Japan is arguably the world’s most pronounced island of robot viability for care and manufacturing tasks. Facing a projected 11 million worker shortfall by 2040, the Japanese Ministry of Economy announced new procurement targets to drive robotics adoption and recently announced it would provide nearly 400 billion yen (approximately $2.4 billion) in aid to build an AI system to control robots. South Korea faces similar demographic pressures, with a rapidly aging population and the world’s lowest birth rate. Both countries are seeing rapid humanoid investment. For instance, Hyundai plans to invest $6.3 billion to build a new robotics manufacturing complex in South Korea as the company shifts its strategy from automotive manufacturing.

Other islands of viability are centered around manufacturing. Germany has a documented skills shortage, while Singapore has a limited domestic labor pool and strict foreign labor quotas — conditions in which humanoid deployment can generate returns for companies deploying them.

China presents a different dynamic. Its full-stack supply chain, competitive manufacturing base, and mass-market scale allow manufacturers to build humanoids at a fraction of the costs faced by competitors. It is estimated that China built 90% of the humanoid robots produced globally in 2025. Two companies, Agibot (based in Shanghai) and Unitree Robotics (in Hangzhou), together shipped over 10,000 units in 2025, while U.S.-based peers like Figure AI and Tesla remained in the low hundreds for actual customer deliveries. The Lunar New Year celebrations in 2026 produced by China Central Television, featuring dancing Unitree robots, was a national declaration of intent for this sector, reflecting long-term government investments.

Regulation adds another layer of geographic complexity, as countries take divergent approaches to safety, privacy, and liability, especially in sensitive settings such as health and social care. As with semiconductors and AI software, humanoid robotics will likely be entwined with the geopolitics of technology, which could result in parallel humanoid ecosystems in the U.S. and China.

The message is clear: For years to come, where you locate humanoids will be fundamental to their viability.

Force 3: The Human Factor

The third factor shaping humanoid adoption is the least understood and potentially the most consequential. People are physiologically wired to react to the human form. When a humanoid enters a workspace, care home, or shop floor, its human resemblance triggers instinctive and visceral responses that other technologies don’t.

Human reaction to humanoids will be an unpredictable but crucial factor in the success of their rollout. Most attention focuses on the potentially negative user reactions of humanlike robots. The uncanny valley effect, first identified in 1970 by Japanese researcher Masahiro Mori, captures how encounters with humanlike robots can leave people unsettled and uneasy. Yet the full picture is more complex.

A growing body of research investigates the diverse and often polarized reactions that humanoids engender. For example, in care settings, a 2025 University of Cambridge study found informal caregivers became increasingly comfortable sharing their emotional struggles with a humanoid robot, and subsequently experienced measurable improvements in loneliness and mood. Some research points to more ambivalent reactions; one study in the hospitality industry found that while two-thirds of guests positively reviewed experiences with robot servers, 28% reported discomfort — with little middle ground.

Human personality types also affect the quality and effectiveness of human-robot interactions. Emerging research suggests that humans tend to prefer humanoid robots with similar personality types to their own. Put simply, extroverts will likely prefer humanoids with more social personalities or exaggerated behaviors, while introverted individuals may be more skeptical or less trustful of humanoid interactions.

Recognizing this complexity, researchers in the emerging field of human-robot interaction are working on how to build trust between humans and humanoids. One focus area is embedding empathy in robots. For example, the whEE (when and how to express empathy) framework is a new approach for adjusting the empathy level based on the situation. Another focus is how to effectively blend different personality types to achieve the best outcomes.

While much remains to be understood about human-robotic interactions, it is clear that humanoid robots must work with the full spectrum of human personalities, as not all users will react in the same way. For business leaders creating or deploying humanoid robots, this means being aware of how easily the experience for employees and customers can shift from positive to negative.

Preparing for the Humanoid Wave

None of these factors will stop the rise of humanoid robots, but these three forces suggest a near future in which humanoid adoption will be uneven and dependent on their role and location, and our complex human responses to humanlike forms.

Few organizations are prepared for this asymmetric adoption. Most risk either accepting visionary promises of productivity that prove unrealistic or missing out altogether on the potential business benefits of humanoids. To succeed, forward-thinking leaders across sectors need to engage on three fronts:

Develop a role-based, location-centric portfolio. The first priority is to identify and prioritize a tight portfolio of high-potential humanoid roles within the organization. Ask which roles could benefit from humanoid involvement. Give each role its own business case, with separate outcomes, risk profiles, and scaling paths. The case for each will not be just about pure labor substitution gains but also benefits such as safety, auditability, and availability.

Determine a geographic market. The geographic variability in workplace regulations, demographics, and supply chains means that focusing on only one country is risky. Your humanoid strategy needs to build in geographic options and partnerships. This might involve running a pilot in a high labor cost market, such as Germany, where the ROI case is strongest, while investing in a learning partnership with a university or research lab in Silicon Valley or a manufacturing operation in Hangzhou. Use these interactions to accelerate learning, even if primary operations sit elsewhere.

Your partnership shortlist should span humanoid-native companies and adjacent sectors such as automotive or consumer electronics ecosystems. Companies in these sectors are already wrestling AI into mass-market physical products. Even if the business challenges of AI and robot implementation differ, you can tap into the accumulated expertise within these sectors for transferable deployment lessons.

Consider the human reaction. Forget about the long history of disappointing and flaky robots. Plan, design, and test for the new chemistry of humans and humanoids — the frontier of advanced AI adoption.

Humanoid deployment isn’t a purely technological transition. It must also involve preparing humans to understand and benefit from this new wave of humanoid counterparts. Above all, articulate the benefits of the next generation of these robots: always available, fully aware of their surroundings, and fluent and empathetic in communication. Throughout the deployment, measure and closely monitor the impact on trust among your employees and customers, and be prepared to adapt rapidly.

The humanoid wave is gathering force. Advanced robotic capabilities exist now, even if more remarkable and superhuman capabilities seem distant. As a result, humanoids will disrupt markets faster than many leaders expect — and more unevenly than anyone is planning for. Success will belong to those who start on their own path to adoption now.

The executive team at Medita Holdings was overstretched. Leaders struggled to decide where to focus their attention — and the company’s shrinking resources — across a patchwork of unrelated ventures. Rivals were pursuing growth opportunities and creating new partnerships, but Medita had stalled.1

During team meetings, leaders were fully engaged and appeared to get along well. But the head of HR had noticed that leaders tended to engage in unproductive discussions, such as questioning earlier decisions without resolving any of the issues they raised, or allowing conversations to drift away from the topic at hand. She began to suspect that the core problem was not the strategy at all. Was the CEO a weak leader? Or was there some other explanation for the team’s chronic inability to focus and follow through?

It turned out that the endless debates sprang from a personality trait the group had in common: strong intellectual curiosity. When teammates struggle to collaborate and make decisions, their behavior can often be traced to similarities and differences in their personalities.

But systematically paying attention to how different personality types affect how teams interact and perform is rare. In fact, the structure and work processes of modern teams make it harder for teammates to get to know one another well and build rapport. Companies increasingly form ad hoc teams as new challenges and opportunities arise. Virtual meetings, now routine, tend to be task-focused and offer few opportunities for team members to connect on a personal level. In this context, “slogging your way to a better relationship with colleagues” through trial and error is painful and inefficient, as Ian Roberts, the CTO at Swiss plant equipment manufacturer Bühler, told us.

Over eight years of working with thousands of executives on hundreds of teams, we developed a way to improve team dynamics more easily using a familiar tool: a personality assessment. Specifically, we used the widely studied and accepted Big Five personality test to help reveal the psychological differences that can sap collaboration and team productivity. The assessment highlights individual differences in how people think, act, and feel across 30 personality traits in five broad dimensions: need for stability, extraversion, openness, agreeableness, and conscientiousness. This framework has dominated personality research because of its high validity and robustness over time.2 (See “The Research.”)

The advantage of using personality assessments as diagnostic tools is that they open up rich conversations. Through facilitated discussions, participants can connect their collective personality traits with team tensions or difficulties — whether latent or overt — and defuse them in the process. And for newly formed teams, the same exercise can provide insights that accelerate cohesion. Whether team members collaborate effectively is often rooted in their ability to work through, accept, and capitalize on their differences.

Know Your Team DNA

When we consulted with Medita Holdings, we suspected that the reasons for executives’ indecisiveness went deeper than weak leadership. To develop a more rounded picture of the team dynamics, we read the most recent 360-degree feedback team members had received, interviewed them individually, and asked each of them to complete a personality assessment. We evaluated this information to determine whether their self-perceptions matched how their colleagues saw them and to identify personality traits that might not be apparent to themselves or others.

Next, we set up a workshop to discuss the results. To prepare, we plotted the personality scores of the five team members, plus the leader, on a single sheet of paper, like constellations. We used a different color for each person but did not identify them by name. The illustration revealed the traits where team personality preferences strongly clustered or diverged, giving participants a sense of what we refer to as the team DNA. Each pattern carries risks and benefits. (See “Visualizing Team DNA.”)

Take extraversion, for example. When a team clusters at the low end of this dimension, their discussions tend to remain structured and everyone is likely to be heard. But their meetings may lack energy. Conversely, when a team clusters at the high end, members may enjoy lively interaction, but they may talk over one another and listen less. Further, when a team is split across the extraversion spectrum, it can potentially combine reflection with action. But the team may struggle with tension between the quieter and more outspoken subgroups, who might see each other as passive or domineering.

The Medita team DNA plot revealed a striking cluster at the high end of the intellectual curiosity continuum, indicating that the team members were extremely open to exploring new ideas. Observing a lower-scoring outlier on every trait related to agreeableness initially (and naturally) prompted them to speculate about who it was. But as we delved more deeply into the specific traits within each dimension, they turned to discussing the challenges for this still-anonymous individual. They recognized, for example, that their colleague might grow exhausted in their role as the lone challenger of the consensus and stop participating. As one team member put it, “You begin to sound like a broken record.”

Further, we explored how individuals scoring very high or low on specific traits may help the team. Participants noted that their colleagues who scored low on conscientiousness may be more flexible when business objectives shifted and that those who scored highly on anxiety may be quicker to detect emerging risks.

Once a team has explored its dynamics in the abstract, we consider revealing individuals’ identities — provided that the members all agree to it. When a team opts for disclosure, the facilitator must invest more time upfront setting the ground rules and boundaries for a respectful discussion. Concealing and revealing people’s identities can both generate rich insights, as the examples that follow show.

Three Patterns to Look for

Once the team understands its profile, members can reflect on the implications of the most extreme patterns — specifically, which perspectives they might be overlooking and where they can expect disagreements.

Three patterns warrant particular attention: a skewed cluster, where scores are tightly grouped at one extreme; a dispersed distribution, where the team diverges widely on one trait; and a grouping with a strong outlier. Left unmanaged, these configurations can create blind spots or tensions, leading to three traps that can derail teams.

1. The like-mindedness trap. The Medita team was proud to see that every member scored high on intellectual curiosity, feeling that it reflected intelligence, creativity, and a learning orientation. But when we started probing the consequences of their collective openness to ideas as indicated by the skewed cluster, they realized that it was the source of their core problem: their frequent digressions and endless discussions.

There was no one to counter their exploration of any decision. “Each new comment opens a Pandora’s box,” noted the CEO. “It’s sometimes impossible to just put a full stop on one item and move on to the next agenda point.”

Because leaders choose team members with qualities they especially value in followers, they may inadvertently drive that convergence of scores. But like-mindedness can lead to blind spots that are difficult to detect until the team fails to make decisions or makes poor ones. The blind spot for the Medita team was members’ inability to see when their free-ranging discussions became unproductive. After surfacing this missing factor, the team designated a rotating “focus keeper,” whose role was to signal when the conversation drifted and get them back on track.

2. The synergy trap. On some teams, personalities may be polarized, with two or more people scoring at opposite extremes for a specific trait. These widely dispersed distributions reveal a potential for synergy, where better decisions arise from integrating diverse perspectives. But that does not occur automatically. As long as personality differences remain unexplored, they are hazardous. Diverging on a single trait can create irritations that ruin relationships and ripple through the team rather than bringing people together. The group may fragment, and its performance may decline. In the worst cases, polarization leads to conflict and turnover.

In a Nordic food processing company we advised, the top team was perturbed by tension between the two most powerful figures below the CEO: the chief strategy officer (CSO) and the CFO. The distress call we received from the CEO went something like “Please do something before I fire them both! They are killing our team meetings.”

When we looked at their personality profiles and listened to how each executive viewed their relationship, a plausible explanation for their tension quickly emerged. The CSO had a high score for gregariousness, while the CFO’s same score was low. “I can’t stand his constant talking,” the CFO told us. “He thinks out loud, and whenever there is a small thing to say, he wants to meet.” From the CSO’s perspective, the CFO was “secretive” and “unwilling to share” because he was often silent and unresponsive. “He just wants to make decisions on his own,” the CSO complained. Their simmering frustrations seeped into team meetings, where the pair pointedly avoided eye contact or direct exchanges. The discomfort radiated outward, dividing colleagues and putting everyone on edge.

Simply pinpointing the source of their differences was a revelation. They pledged to stop venting about each other to colleagues, and they chose to focus instead on their similarities — notably, their high conscientiousness scores. They were both strongly motivated to achieve their goals and confident in their abilities.

A whole team can be polarized too, which can lead to a deep fault line. This happened at a European engineering company. The board chair called us in after informal chats with members of the top team led him to conclude that some executives were disengaged and at risk of leaving.

Another board member described the problem to us. “Everything gets decided by the CEO and these two people,” she said. “It’s like the other four have nothing to contribute.” Worse, she sensed that the CEO’s two “lieutenants” were not the most capable members of the team.

The CEO’s extremely low score on measures of moodiness indicated that he maintained a consistently positive outlook, which cleared the way for broaching this issue. He described himself to us as an incurable optimist, and his favorite team members’ mood scores most closely matched his. The rest of the team was loosely grouped at the higher end of the mood scale, meaning they tended to be more measured in their outlook.

Pointing out the CEO’s fixation on positive energy enabled us to discuss his unconscious bias against the four less upbeat team members and the risks of unchecked optimism. It also clarified why the four seemed increasingly disengaged. He had assumed that their withdrawal had led him to favor the other two team members, but in reality, the four participated less because he favored the other two. Recognizing this helped the CEO to appreciate their value as counterweights to his exuberance and led him to plan an offsite to reset the relationships.

3. The outlier trap. Teams fall into the third trap when the entire team is unified on a single trait except for one extreme outlier. This outlier may be ignored or dominate the team, in both cases preventing the group from making full use of its competencies.

In the first scenario, the outlier’s dissenting voice, though valuable, is disregarded by the rest of the team. Consider the case of a self-managed project team struggling to make progress on a strategic consulting task. After we debriefed each individual on their personality profile, we asked the group to articulate some rules of engagement and explore their team dynamics without a facilitator. (We wanted to see whether teams could run the exercise themselves; we talk more about that below.)

They quickly homed in on their starkest differences and figured out how they affected their interactions. All of the team members were working on the project alongside their usual responsibilities, and they clashed repeatedly over how often to meet. One person insisted on frequent and regular meetings. The more she pushed, the more they resisted, dismissing her as a “pain” and exacerbating her frustration.

“And this is why,” they said, pointing to their scores for conscientiousness. While most of the team had scored somewhere in the middle on this set of traits, the woman who wanted more meetings had scored extremely high on both achievement-striving (that is, focus on reaching her goals) and self-discipline.

This insight helped them all realize the source of her deep concern about scheduling and progress. They came up with a compromise that they could all commit to: Instead of scheduling more meetings, they would hold brief end-of-week check-ins to stay aligned.

Beyond ignoring the “odd one out,” the team can also become “odd one led,” notably when the outlier is either the formal leader or the opinion leader in the group.

Take the case of a high-profile but underperforming European opera house. To turn it around, the board had appointed a former accountant as director general. While the new director general brought commercial discipline to the operation, the opera house’s artistic productions continued to underwhelm critics and audiences. The four-person executive team seemed unable to break the cycle of turnover among directors, designers, and performers.

Interviews with members of the executive committee and other senior leaders revealed that the director general had reinforced the existing siloes between the artistic and commercial functions and that the financial focus had come to dominate their discussions.

The personality profiles of the four top executives revealed that the director general had a strikingly low score on adventurousness, whereas the scores for the rest of the team were bunched in the middle. The director general’s preference for the habitual seemed to curtail efforts by the rest of the team to introduce new thinking and explore how other performing arts venues were reviving their fortunes.

His score opened a wide-ranging discussion about the opera house’s need for fresh ideas. Later, the executive team was restructured to include the artistic side of the organization, giving it a greater voice in decisions.

Although it’s too early to know whether those steps will improve the opera house’s financial and artistic health, the changes have reduced frustration within the organization and brought forward some fresh initiatives.

The Ultimate Icebreaker

Personality is not the sole cause of team dysfunction and tension, but it plays a significant role. It is a guaranteed differentiator, even in the most homogeneous teams. This makes a workshop focused on personality types a rich and intriguing entry point for discussing team dynamics.

“Talking about personality kicks off conversations and lets us see each other’s struggles and underused superpowers,” Medita Holdings’ head of HR said. She noted that extreme scores might not surprise anyone but that discussing them candidly enables participants to find value in traits they otherwise see as frustrating.

Further, many executives have experience using the Big Five personality test for their personal development, so they understand what they can learn from it. The International Personality Item Pool (IPIP) is a free, open-source version of the instrument that includes a comprehensive report explaining the scores and the option to plot comparison charts. Additionally, the IPIP can be adapted freely and does not need to be administered by a certified facilitator.3

Debriefing the exercise is straightforward for any facilitator, coach, or consultant who is familiar with Big Five assessments and can set and enforce behavioral boundaries, help teams reflect, and keep discussions on track. The team does most of the work, by identifying the dimensions where they struggle or that resonate most with them.

Team leaders can facilitate the process themselves, role-modeling and enforcing the behaviors required for an open discussion. To succeed with this approach, however, they should disclose their own personality assessments and participate in the discussion, as happened with the self-facilitated project team mentioned earlier. Disclosure allows team members to question the leader’s preferences and blind spots, connect their insights to observed behavior, and make practical adjustments.

However, even when team leaders are willing to share their assessments, other team members may hesitate, concerned that revealing their own preferences will expose them to judgment or misinterpretation. For this reason, the exercise often works best when facilitated by someone from outside the group who can credibly reassure participants about the purpose and process of the discussion. It can take time to grasp that there are no “bad” profiles and that individuals can safely reveal their identities. An external facilitator can also point out dynamics involving the leader without triggering defensiveness.

Before starting the exercise, the facilitator must lay the groundwork to help participants cocreate the rules for interaction, understand the payoff, and buy into the process. A half-day session allows for this preparation plus two or three hours of discussion and reflection. It is important not to rush any step, so that everyone has a chance to contribute. Researchers find that conversational turn-taking boosts psychological safety and team performance.4 (See “Guidelines for a Productive Discussion.”)

Mutual agreement on the ground rules creates clarity and trust by addressing three key questions: (1) What will be important to achieve in this conversation? (2) What topics or behaviors are off-limits? and (3) How will we call out unacceptable behavior? Facilitators can foster candid and respectful discussions by requiring participants to focus on a common goal, disagree respectfully, stick to facts and evidence, acknowledge their biases and knowledge gaps, and ensure that everyone has a voice.5

Teams do not need to wait for problems to arise to engage in the process. We know a couple of leaders who shared their personality profiles unilaterally with teams they had inherited, to help their new followers adapt faster to their quirks and leadership styles. Inspired by this proactive approach, we expanded our research to new teams and teams with new leaders. We found that personality assessments can help them jell more quickly.

Using this exercise when team members first begin working together has several benefits. It provides them with a common vocabulary to talk about their individual and collective strengths, sensitizes them to the limits of those strengths, and highlights the potential value of having contributors with unusual traits. More broadly, it jump-starts team collaboration by getting colleagues used to disclosure, information sharing, and mutual respect.

After we ran this exercise with the top team at a French fintech group and an internal coach, they decided that the entire staff would take the Big Five personality assessment to prepare them for joining any new team. “Discussing hidden differences as a team … changed our view of colleagues and the way we talked to each other,” the CEO said. “It saved a lot of time and misunderstandings.”

Reducing the Guesswork in Teamwork

Whether the aim is to fix problems or prevent them, comparing personality assessments benefits individuals and their teams. Enhanced self-awareness makes people better team players, while team awareness improves group interaction.

Even if someone has received personality feedback before, when they compare their results with those of their teammates, they see how distinctive they are. Enhanced self-awareness enables individuals to better recognize how their traits work for or against them in different situations.

For example, a Danish business leader we coached scored lower than he expected on altruism — specifically, on his consideration for others. Yet, as an accomplished manager, he knew that he had to take care of the people on this team. Because he lacked the intuition to guide his interactions, he developed a habit of checking in regularly with his direct reports. As a result, they viewed him as considerate of their needs, despite his underlying self-absorption.

Although it’s difficult for people to modify their traits, anyone can stretch their repertoire of behaviors, as the Danish leader did. Individuals can develop routines to keep their traits in check, such as resolving to speak last if they score high on assertiveness and tend to dominate conversations.

Another option is to find others with complementary profiles. For example, in a Financial Times interview when he was CEO of VMware, Paul Maritz was candid about his Achilles’ heel. “I don’t like confrontation,” he said. “I need other people to keep me honest on that front.” He appointed an “enforcer” to his team to keep him focused on tough decisions he might otherwise dodge.6

Self-awareness is only half the battle, however. Realizing that colleagues have different ways of thinking, doing, and feeling is the other half. The team DNA chart highlights areas where the entire team needs to be vigilant. A team that is aware of its collective tendencies can take steps to compensate for its potential vulnerabilities.

In one exercise, the CTO of a watch manufacturer was surprised to see that the entire team, including himself, had midrange scores on imagination. He had perceived himself as highly innovative, but when he challenged the validity of the results, his team members disagreed. “No, you have initiative,” a colleague said. “But that’s different from innovation. You start things, and you give some impulse, but then your ideas are not … .” The sentence trailed off.

That feedback was tough for the CTO to take, because he oversaw innovation. But it prompted an overdue discussion about how the company could resuscitate its innovation capability and get out of its slump.

With teams in a state of constant flux, leaders need better tools to bring new members up to speed, facilitate productive exchanges, and quickly create rapport. As the chief operating officer of Medita Holdings observed, “They’ll find out about each other anyhow. May as well do it upfront” while establishing principles and good habits for communication.

Exposure to the range of personality differences within a team helps executives better understand where colleagues are coming from. It makes it easier to value and solicit their input and to anticipate their quirky reactions. Antagonists come to realize that a teammate’s thinking is not stubborn or wayward but predictably different — and they can stop taking the opposition personally.

Paradoxically, knowing colleagues’ personality traits depersonalizes discussions. Above all, it boosts people’s appreciation for what each party can bring to the table. As Maritz memorably observed, “Really great teams have team members who know who they are and who they’re not, and they know when to get out of the way and let the other team members make their contribution.”7

Erik Brynjolfsson has a challenge for anyone worried about artificial intelligence: Stop asking what AI will do to us, and start asking what we will do with AI. In this episode of Me, Myself, and AI, the Stanford University economist explains why technology isn’t the biggest barrier to progress — people, organizations, and institutions are. Drawing on new research into AI’s impact on jobs, productivity, and economic growth, he argues that the future isn’t predetermined: It will be shaped by the choices we make today. This is a timely conversation about human agency, shared prosperity, and why the most important AI breakthroughs may have less to do with technology than with how we use it.

Subscribe to Me, Myself, and AI on Apple Podcasts or Spotify.

Transcript

Allyson Ryder: Today’s guest has a bold provocation: AI isn’t being held back by the technology. It’s being held back by us. Curious how? Find out now.

Erik Brynjolfsson: I’m Erik Brynjolfsson at Stanford, and you’re listening to Me, Myself, and AI.

Sam Ransbotham: Welcome to Me, Myself, and AI, a podcast from MIT Sloan Management Review exploring the future of artificial intelligence. I’m Sam Ransbotham, professor of analytics at Boston College. I’ve been researching data, analytics, and AI at MIT SMR since 2014, with research articles, annual industry reports, case studies, and now 13 seasons of podcast episodes. In each episode, corporate leaders, cutting-edge researchers, and AI policy makers join us to break down what separates AI hype from AI success.

Today I’m talking with Erik Brynjolfsson, who runs the Stanford Digital Economy Lab and studies how people use technology in general and now AI specifically. He’s trying to think about how technology is affecting [the] economy and work. His book The Second Machine Age shaped a lot of this debate. And his team recently published “Canaries in the Coal Mine?”, a research paper about what’s happening to entry-level workers. Erik and I bump [into] each other a few times a year at the National Bureau of Economic Research, and he always brings up something I hadn’t considered. No pressure, Erik, but I’m expecting the same today.

Erik Brynjolfsson: [It’s] good to be here, Sam.

Sam Ransbotham: Welcome to the show. Let’s start with the Stanford Digital Economy Lab. Can you give us a quick view of what the lab does?

Erik Brynjolfsson: Sure. We study the digital economy. … I loved my time at MIT. I was there for over 25 years, but now I’m out here in Silicon Valley, the epicenter of the AI revolution. We’re focusing on how AI and other digital technologies are changing the economy. Kind of the premise of the lab and of my work, my career, is that technology is advancing very rapidly. The capabilities are amazing. At the same time, our economic understanding is not advancing nearly fast enough. Our economic institutions, skills, organizations, they aren’t keeping up. My job is to try to close that gap.

Sam Ransbotham: It’s those pesky people. The technology moves fast. Organizations and people slow us down, I guess, is the summary there.

Erik Brynjolfsson: That’s right. A lot of your work has highlighted that, too, and we’re doing what we can to keep up.

Sam Ransbotham: I was thinking [about] the Race Against the Machine book and … “Canaries” today. What connects all of it?

Erik Brynjolfsson: “Race against the machine” … was the headline. The conclusion was that we should race with machines, not against machines. And that’s what I’ve continued to emphasize — there’s an opportunity for humans and machines to work together. It’s not automatic. There’s a lot of choices that we need to make, but the technology is enabling all sorts of new possibilities. Trying to invent and discover those is a big part of what we humans need to do.

Sam Ransbotham: I like the choice part because I think it’s so easy for us to slip into [the] vernacular of saying, “AI does this, technology does that.” I know you’ve been against that.

Erik Brynjolfsson: That’s exactly right. It’s probably one of the most common questions I get: “What’s AI going to do to us? What’s AI going to make happen next?” I’m like, “Wait a minute, the premise of your question is wrong. It’s ‘What are we going to do with AI?’”

This is an incredibly powerful tool, arguably the most powerful tool that we humans have ever had. Almost by definition, that means we have more ability to change the world than we ever had before. So let’s think about how we want to use that. It turns out that our values, our choices matter more now than they did in the past because we do have this ability to make a really big dent in the universe, for better or worse.

Sam Ransbotham: Yeah, for better or worse. I think that’ll probably come up a couple of times as we talk here.

I was thinking about how much you got right in the early Race Against the Machine book. But that’s not interesting as an academic. What do you think you got wrong? What were you surprised about? What’s changed differently than you thought 16 years ago or 15 years ago?

Erik Brynjolfsson: Well, first off, since we’re talking about what we got wrong, I should highlight that this is cowritten with Andy McAfee, along with The Second Machine Age. He’s been a partner on a lot of my projects.

What did we get wrong? We were looking at some advances — I’d highlight particularly in The Second Machine Age, we started that with a ride in Google’s self-driving car. Actually, it was 2012, right after we wrote Race Against the Machine, we rode from Mountain View up to San Francisco. I’ve got to say I thought, “Wow, self-driving cars are just around the corner,” metaphorically, and I thought it would happen very quickly. It’s obviously taken longer. I do ride around in self-driving cars a fair amount here in the San Francisco Bay Area, up in the city. They’ve had them for a while, and now they’ve come down to Palo Alto as well, and you can ride them all the way on the highway. But that’s 15 years later.

It’s taken a while, and I think I was overoptimistic about that. That said, I think I underestimated how fast the technology would advance in other ways. The way you and I can talk to [large language models], whether it’s ChatGPT or Gemini or Claude, and have them do work, I think I would not have expected that to have happened that quickly. I mean basically I would have considered this [artificial general intelligence] if you had asked me in 2012 or 2015. So that’s pretty cool that you can get really good advice from them. We all know they still hallucinate and make mistakes, but so do we humans. On average, they’re really quite good. So that happened faster. We got that wrong a little bit.

Then the most disappointing part is, look, I knew that we were just talking about how human institutions change more slowly than technology. But oh my God, I didn’t expect it to be this much slower. At times it’s almost like it’s moving backward. So that’s been pretty disappointing, that our political institutions, our businesses, organizations, they aren’t keeping up. Productivity is not growing any faster than it was at the beginning of this AI revolution, maybe a smidge if you kind of squint. But we’re not really translating these capabilities into better business performance or the kinds of benefits that we all hope for.

Sam Ransbotham: I rode one of the automated cars out in Palo Alto at the last meeting where I saw you at. It very quickly got boring. Now I was fascinated for a few seconds and then bored very quickly. You touched on that would have been AGI a few years ago, or LLM stuff would have been AGI a few years ago, but it’s not anymore. How much of this is sort of a boil frog, we are getting used to these technologies, it takes even more to impress us? Is that part of it? Is it a measurement part? Is it a J-curve, which we’ll come back to in a second? All the above? None of the above?

Erik Brynjolfsson: All of the above. Part of it is, like you said, getting bored in the self-driving car; the first few minutes are kind of scary and thrilling, and then it’s exciting. And then within half an hour you’re like, “OK, I kind of get this.” These cars, they kind of drive like my grandmother, very carefully and slowly, which is, I guess, the right thing to do.

Sam Ransbotham: That’s why you left Boston maybe.

Erik Brynjolfsson: Exactly. Their accident rate is 90% lower than humans. I’m looking forward to the day that the cars are all like that. Right now they are a little slower point to point. Sometimes when I’m in a rush, I’ll actually get an Uber because I know that the driver is more willing to cut some corners, so to speak. But part of it is that we kind of take for granted things that initially were eye-popping.

I guess our brains are designed to get used to things. But the other two things you said, I think, are more important and more interesting. One is the measurement issue. Our economic metrics are just not designed to capture a lot of the benefits of AI, especially the benefits of free digital goods. So we’ve got a whole methodology for addressing that. Maybe we can talk about that later.

The other part is the J-curve. This is what we started the conversation about — there really is a gap and a difficulty in translating capabilities into real business change. It’s because there [are] so many complementary assets and investments that need to be made. Many of them have to be invented, and we don’t know what they are, these complementary co-inventions of new business processes, new skills, whole new business models and ways of working. They’re hard and we have entrepreneurs and managers and business school professors and consultants all working on figuring them out.

In earlier work I did with [University of Pennsylvania’s] Lorin Hitt and Prasanna (Sonny) Tambe and others, we found that the investments in those intangible assets are about 10 times bigger than the direct investments in computer technology. But since they’re largely intangible, largely invisible, we tend to not appreciate them as much, and we don’t measure them, and people don’t even realize that they’re there. But we’re trying to make them more visible in our research and create an appreciation that you have to make those investments as well.

Sam Ransbotham: Actually, the measurement thing is huge and difficult. I remember you and [Carnegie Mellon University’s] Avinash Collis were talking about the measurement of value from Wikipedia and from Facebook and these other technologies. All this stuff is just superhard to measure. And when we don’t measure stuff, we do a bad job, I think.

Let’s switch maybe to measurement because I think one of the cool stories about your “Canaries in the Coal Mine” study … I think your headline finding was that people ages 22 to 25 were the most exposed [and] lost about 16% of their employment. But you’re using ADP payroll data to find that. That’s a very interesting signal of information. Tell us about that study.

Erik Brynjolfsson: Like everybody, we were seeing all these headlines about AI eliminating jobs or other headlines [that] AI is creating jobs. In any given month, there [are] hundreds of thousands of people who gain or lose jobs. A half-decent journalist can easily find examples of either one and build a narrative around it, or just a man in the street they can talk to. To me at least, I don’t really know how to aggregate that. What’s the real story? So we wanted to get large-scale data.

When we looked at the current population survey, we didn’t really see much, and when we looked at the top line of ADP, which is the world’s largest payroll processor, they shared the data with us here at the Digital Economy Lab. We actually didn’t see much in the top line either. I was almost thinking, “OK, we’ll write a story about how this is all just a bunch of hype.” But then we dove in a little bit more deeply, and some of the subgroups who are having these very large effects, like you just mentioned, people who are ages 22 to 25, early-career workers, they had a noticeable fall in employment, but it really got striking when you subdivided it by how exposed they were to AI.

[OpenAI’s] Tyna Eloundou, [University of Pennsylvania’s] Daniel Rock, and others have created a taxonomy where they rank all of the tasks in the economy, 18,000 distinct tasks by how much they’re exposed to LLMs. Could an LLM help you write a memo? For sure. Could an LLM help you lift a box? Not really. If you rank every task like that and then aggregate them up to occupations, you basically have a score of whether or not this occupation is going to be affected by LLMs.

We looked at about 750 occupations. We divided them into quintiles. The most exposed quintile, a little over a hundred occupations in this youngest age group, had at first about a 12% to 13% decline. Now it’s more like a 16% to 17% employment decline relative to the other occupations. So it’s really quite striking. We didn’t see a fall in employment for the older workers, even in the exposed occupations. You really have to slice it this way.

Even when we controlled for other factors, like we controlled for interest rates, we controlled for the tech industry — you can take the entire tech industry out if you want, or control for tech overhiring — you can control for remote work, you can control for education, none of those things knocked out the core result. We continued to have this very noticeable decline.

I can say another thing, which is that, as I mentioned, the older workers were doing OK. There are a couple of other groups that were doing OK. One is the people in the not exposed occupations. So if you look at the other end of the spectrum, like home health aides, where AI is not affecting them nearly as much, they actually had growing employment, even for the young workers.

And then, the most interesting result, the one I’m most excited about, was that if you divide the way they’re using AI into mainly automating and eliminating work versus augmenting and creating new skills, new opportunities, doing new things that you never did before, the automating group had falling employment, and the augmenting group, the ones who were learning new skills, had growing employment. So it’s kind of a double win: higher productivity and growing employment. For some reason that doesn’t get picked up as much in [the] press when they write about our research, but I think it’s really important that we found that AI could be associated with higher employment for certain kinds of workers.

Sam Ransbotham: I think that ties back to how you use it, your decision-making, the choices that we all make in that. That doesn’t surprise me, because if I think about a task and, going back to this aggregation of tasks you alluded to, no job, I’m sure, runs a hundred percent down the list of tasks and a hundred percent down the list of non-automatable tasks. The composition of that task is going to leave parts of those jobs more valuable. That’s going to leave those parts more valuable.

Erik Brynjolfsson: Exactly. That’s another really important point: You can take every occupation; [each has] a bundle of tasks. In O*Net, a typical occupation has 20 to 30 tasks. A radiologist has 26 distinct tasks. [Computer scientist] Geoffrey Hinton famously said, “Reading medical images, that’s what radiologists do. That’s going to be replaced by AI.” What he didn’t factor in was there were 25 other tasks that radiologists do, and most of them were not affected by AI.

The net effect was that as reading images got cheaper and better because of AI, it actually increased the value of radiologists for doing those other tasks, and net employment grew for radiologists. I’m not saying that always happens, but there [are] definitely cases — some people call it Jevons paradox — where making something cheaper increases the demand for it, and you end up having more employment rather than less, even as it gets more efficient.

Sam Ransbotham: I think it’s the frustrating thing. I think everyone would like an answer, which is “Here’s the clean answer,” and what you’ve come across here is just a gigantic “It depends,” and it depends on an increasingly complicated set of stuff.

Erik Brynjolfsson: Well, let me try and make it a little simpler. It is somewhat more complicated than the simple story of you always eliminate jobs, or for that matter, it always creates jobs. But as an economist, a really useful tool for me — and I’ll get a little wonky for your readers, but it’ll be [a] useful time — is you can think of a demand curve. When price falls, quantity increases. Demand curves are downward sloping. Most of us intuitively know [when] you make something cheaper, people will buy more of it.

But what really matters is how steep that demand curve [is]. For some products, like apples, even if you made them 10 times cheaper, you’re not going to buy 10 times more apples. Maybe you buy a few more apples. But the demand curve’s pretty steep. So as prices go down, you end up spending less. That’s pretty intuitive. Most people think that all goods are like that.

It turns out that not all goods are like that. There are lots of goods where as the price goes down, even just a little bit, the quantity goes up a lot, like jet travel. It used to be that very few people would fly in planes across the country, but now it’s cheaper, and a lot of us fly around the country way too much. So there are goods and services like that. That’s called elastic demand or a flat demand curve.

Whenever you have situations like that, you actually have growing demand for something, and you end up having more spending. At least half the goods and services in the economy are on that side of the curve. That’s really good news. It means that as we get more productive, more efficient, some things shrink, but other things grow. I think there’s a little bit too much of a bias in the public conversation toward the side that’s shrinking. One of the things I want to do with your podcast is get more people to think about that side that’s growing.

I teach a master class [in] which we basically show people how to identify and discover these new opportunities for wealth creation, how you can use AI to do new sorts of things to create entrepreneurial ventures. The more we do that, I think the better off people are going to be. They’re going to help themselves, and they’re going to help the rest of the economy by inventing new things.

Sam Ransbotham: Apples [are] a great example. Like you say, we don’t consume them when we get more of them, but I don’t know of any company that says, “Pretty much all the IT projects that we could ever think about, we’re doing them.” You know what I mean?

Erik Brynjolfsson: Yeah, exactly. It seems like there’s almost infinite demand for solving all sorts of IT problems. When I look around the world, I don’t see a shortage of problems. Poverty, the environment, health care especially, these are all things where we could use a lot more resources, and I don’t see us running out of opportunities to address those.

Sam Ransbotham: The fantasy that we would run out of those types of problems, that’s great, but I don’t think that’s going to happen in the next couple of years though, the next couple of centuries. Let’s go back to this J-curve idea. First explain a little bit about what is the J-curve, and how should people think about that?

Erik Brynjolfsson: This is another concept I think people find very useful, and that is whenever there’s an important powerful technology, we economists call them general-purpose technologies. We used to just say GPT, but my AI friends stole that acronym. But AI is both. It is a generative pre-trained transformer, like GPT-5, but it’s also a general-purpose technology like the steam engine, electricity, the internal combustion engine, and cars. Those are the things that actually drove most economic growth, most productivity. They’re the reason we have higher living standards than our ancestors, a couple hundred years ago or even 50 years ago.

AI is the most important general-purpose technology of our era, maybe of all of history. But the thing about these general-purpose technologies is that the value really comes from the complementary assets. With electricity, it enabled light bulbs and electric motors and air conditioning and ultimately computers and a lot of other things. With AI, it’s enabling all sorts of new business processes, and it’s not just the physical technologies. It’s also these intangible assets that we talked about earlier. That’s great. The thing is that these intangible assets take time to create. They’re expensive. You have to invest in them. Whether you hire consultants or you do it yourselves, you have to reinvent how work gets done.

During that costly period, you’re spending more as you’re reinventing your business processes, but output doesn’t instantly go up. Mathematically that means more input, no increase in output, [and] productivity, at least as it is conventionally measured, goes down. So that’s the downward part of the J-curve. Then later you start harvesting that. You start using these new business processes to create more output. Now output is going up. And that’s the upward part of the J-curve.

What we’re seeing with most general-purpose technologies is that there’s an initial lull where productivity is low or even negative, and then later it takes off. Now it’s hard to see exactly where we are for AI. We’re in the middle of it. But if you look back at earlier history, if you look back at, say, electricity — I wrote [about] some of it in my Ph.D. dissertation — believe it or not, it took about 20 to 30 years of time where companies were trying to reinvent how factories were organized.

They were installing electric motors. At first they did not get much productivity. [Economist] Paul David and others documented this very carefully. For literally up to 30 years there was essentially no productivity gain in factories as they were electrifying, which is just remarkable, but, eventually, they completely redesigned how factories were organized. Instead of being clustered around a central motor like a steam engine, they had separate motors for each piece of equipment, and they distributed the work, and they laid out the arrangement based on the flow of materials.

When they did that, productivity started skyrocketing, like doubling and tripling productivity. But like I said, it took 30 years. I don’t think AI is going to take 30 years. I’m pretty sure it’s already beginning to happen. But at the same time, it takes longer than some of my technology friends, some of my AI friends expect. They think that as soon as you invent the AI, you’re done.

The reality, as you know, Sam, in your work and in my work, there’s a lot of business process redesign that has to happen. We’re in the middle of that, and I’m doing what we can to try to speed that up. My company Workhelix is very involved in showing companies how to identify the opportunities for value creation. If we can speed that up, it’ll be great, but we’re still going to have a bit of a J-curve.

Sam Ransbotham: I’m thinking back to the electricity times. I’m guessing they didn’t have the quarterly stock report pressures and the need for instant results that perhaps people deal with now.

Erik Brynjolfsson: The biggest thing, honestly, is that they didn’t really have all of our institutions. Not to pat ourselves on the back, but business schools didn’t exist. Consulting books [and] the whole science of industrial engineering and management, didn’t exist. I went back and I read some of the books. I went to Harvard Business School’s Baker Library, and they had these old books, and it was pretty primitive how they were thinking about things.

Now there’s a much bigger ecosystem for helping companies redesign their business processes. It still takes time and redesign, [and reskilling] their workforce. But my sense is managers are much more conscious of it right now, and they’re leaning in much more aggressively, and they’re sharing best practice.

At Workhelix, we have a tool that will scan through all of the opportunities and share them with the workers and look at who the superusers are who are really crushing it, and then compare them to the average worker. So it just speeds up the learning by probably 10x compared to what people would have done before they had these kinds of tools.

Sam Ransbotham: I really like that because that again ties in measurement and some other things. The annual performance review is really not helpful anymore. The idea that you could get [a] better measurement about what you’re doing and how you’re doing it helps.

Let me kind of be antagonistic a little bit. I wrote something at MIT Sloan Management Review a while back about rethinking AI objectives. You wrote something about the Turing Trap. I think the first inclination that people have is to pass the Turing test, to make the machine do what the human can do. And that leads us to incremental improvements of existing processes. It seems like measurement could accidentally feed into that as well if we’re not measuring the right things. What are you thinking about that, short term versus long term?

Erik Brynjolfsson: Well, there’s definitely a short term, long term; [those are] your objectives. I was talking to a CFO a couple months ago, and she told me, “Hey, we really need to measure the benefits of AI, the ROI.” And I said, “That’s great. You should be measuring it. Tell me about it.” And she said, “So that’s why I’m demanding that every division tell me how much head count reduction they’re getting from AI.” I was like, “Wait a minute. I’m glad you’re measuring, but isn’t that a little bit simple-minded, because, yes, there’s nothing wrong with cutting costs, but it’s really missing the bigger opportunity of doing new things, allowing your workers to create new kinds of value.”

There’s a CFO getting paid millions of dollars, and the reason they’re making big money is that they’ve got to think more creatively about value creation, not just having machines do what the humans are already doing and replacing them. That’s the classic Turing test: Can you make a machine imitate a human perfectly?

That’s a very narrow way of thinking about it. I think Alan Turing was a brilliant guy, but I think that mindset has done a lot of damage to how we use AI. I call it the Turing Trap because I think that it’s a trap to only focus on using AI to replace or imitate workers. We should also think about how AI can augment and allow us to do new things. But that requires new measures, like, how do you create new products and services? How do you have higher customer satisfaction? How do you improve quality, both the quality of the products but even the quality of the work life for the people working?

These are all things that won’t show up in your narrow replacement mindset, so this is something that also takes a longer time, like you said. Short-termism tends to focus on just cutting costs and what you’re already doing. But the ones that have the most competitive advantage, the ones that have the most lasting benefits, are the ones that think of going beyond that and doing new things, new business processes, new products and services. It doesn’t happen in weeks, but when you do achieve it, you have something that lasts for years.

Sam Ransbotham: There’s plenty of places where machines outperform us. The easiest way to tell if this is a machine or a person and pass that Turing test is that I’m not going to be able to do math very fast. Our objectives are not exactly the right standard, I guess. You mentioned that in the [beginning of] electrification people didn’t have the institutions around that. Neil Thompson [and] the MIT Initiative on the Digital Economy, [which] you were obviously involved in before you moved to Stanford, they’ve got a science article that said something like 90% of the most important new AI models are coming out of industry. We’ve had this massive switch from academia and public funding [producing] technology benefits to most of it coming out of technology companies and startups. You mentioned Palo Alto and the epicenter that’s there. All right, I’m worried. Is that the right place for that stuff?

Erik Brynjolfsson: Part of the reason it’s coming out of industry is that it’s gotten to be incredibly expensive to build these very large models. There are these scaling laws for LLMs where as you make a model bigger, if [it has] more data, more compute, more parameters, you get a predictable improvement in performance. So they went from spending millions to tens of millions, hundreds of millions, billions, now tens of billions, even hundreds of billions of dollars on training these models. No university can afford that. No nonprofit organization can afford that. So the companies are raising huge amounts of money as we’re speaking. Anthropic and OpenAI are filing to go public. Google has a nice cash engine they can pour into this as does Meta.

So that’s what’s driving that. At the same time, what I tell my academic colleagues is your comparative advantage isn’t [in] spending more money on computers. It’s in thinking creatively. There [are] some brilliant ideas that happen just from sitting in your office and thinking deep thoughts. It’s still a good strategy.

I had a conversation with Geoffrey Hinton a while back, and I asked him what kinds of hardware he uses to make his insights and discoveries. He pointed to his laptop and tapped it and said, “This is the hardware I use.” And for those kinds of fundamental breakthroughs, it’s not always a matter of having thousands of GPUs.

I think academics need to lean into their comparative advantage. That said, it’s fair to say that more and more of the frontier research is happening through a really expensive approach. In a way it’s a testament to how valuable AI is. One of the reasons industry didn’t do it before was not just that it was expensive, but it just didn’t have that big a payoff.

I’ll share a story. The first AI company I started, believe it or not, [was] in the late 1980s. That’s how old I am. I started an expert systems company with Todd Loofbourrow. Expert systems are these rule-based systems. We created a little bit of value for some banks and HR organizations, but [at] the end of the day, it was nothing like the value that companies are creating today. That brand of AI just wasn’t all that valuable, I have to be honest. It [is] only now that we have these very powerful neural networks that you really have commercial incentives that are unlike anything we had before. So in a way it’s a testament to the success of the field that industry is investing as much as they are.

Sam Ransbotham: I like that point. The U.S. government, I think, put like a billion and a half dollars or so into non-defense research spending last year, and that’s roughly what Google is putting in. You know, the [National Institutes of Health] and [National Science Foundation] budgets are small. As you point out, we are not going to be able to outspend those.

Erik Brynjolfsson: But on that, I think it’s fair. You want to have both. I mean, it’s the job of academia, it’s the job of the federal government to invest before it’s commercially viable. So much of the long-term benefits of R&D [are] not apparent, and there’s not a commercial incentive to invest. The internet, early space travel, so many fundamental breakthroughs in biology that have extended our lives, they happened in universities and government laboratories because they were not commercially viable, but then later people built on them. You can’t always tell in advance which ones are going to pay off. But if you do the fundamental research, the track record shows that eventually some percentage of them are going to really make a difference in our living standards.

Sam Ransbotham: That’s the basic applied cut. We don’t have to fund the applied stuff because the market will take care of that. And [for] the basic stuff, the invisible hand doesn’t work quite as well in those, or it works slower than everybody would like.

I think you’re going to have a bunch of academics listen to this because they’re curious what you say. What should people be looking at? What are the kinds of things that fall into that category? Maybe a critique is that a lot of the studies that I’m seeing in academia are, let’s say, consulting projects for companies. What are these basic types of ideas that people should be pursuing?

Erik Brynjolfsson: Well, there’s a lot of them in lots of different areas, and I think I don’t want to prescribe them because I think in a way part of the magic is letting people pursue their own ideas even if there’s nobody on the outside that sees the value of them.

That said, I’ll give you my own list. I’m focused on the economic side. I think that right now the biggest gap is that we don’t have a good understanding of how AI is going to transform the economy. We are creating incredibly powerful information processing entities. As an economist, you can think of a market as an information processing entity. You can think of firms and organizations as information processing entities. If you have a thousandfold or a millionfold improvement in AI, it would be a miracle if firms and markets didn’t also transform and we develop new kinds of organizations to manage work. But we haven’t figured those out yet. So [a] big part of my work is trying to understand, what does the economy look like? [What are the] firms, markets, and maybe some new kinds of entities as AI gets more powerful, as it aggregates information, as it processes information? And then how are we going to manage the concentration of wealth and power that could result?

I’m super optimistic about the potential of technology to boost productivity. I have a bet, a long bet on that with [Northwestern University’s] Robert Gordon, and I’ve written about how I expect productivity to grow. At the same time, I’m pretty worried that if we don’t manage it right, it could lead to a great more concentration of wealth, and that tends to lead to a concentration of power. And that could hurt our freedoms if we don’t figure out an economic system that balances not just wealth creation but also shared prosperity.

Sam Ransbotham: I was looking up some of your background before talking, to refresh, and I saw the phrase mindful optimist, and it kind of hit me. So what is it going to take to get us to this positive version of the future versus the negative version of the future?

Erik Brynjolfsson: Let me take a minute to define mindful optimist, because I have a very specific meaning for that. There’s a lot of people out here in Silicon Valley who are optimists, but I think too many of them are kind of blind optimists. They’re like, “Hey, Erik, don’t worry. It’s worked out in the past. It’ll work out in the future. Just chill.” There are also a lot of people who are what I call blind pessimists. They’re like, “Oh my god, we’re doomed. There’s nothing we can do.” I think both those groups make the same mistake.

It’s the one we talked about at the top of the program, which is they think of AI as something that’s going to do things to us as opposed to us having the agency to control the future. A mindful optimist is somebody who sees a future and then works toward doing it, is mindful about creating it, doesn’t just assume it’s going to automatically happen.

They’re not like kids at Christmas [who] expect presents to magically appear the next morning, but instead they’re more like kids who see a tree, and they see some boards, and they think, “Hey, we could build a tree house there.” Then they’re like, “Well, it’s not going to happen by itself. Let’s figure out how to make it.” And then they make that optimistic reality happen through their hard work and through their imagination. That’s the kind of optimism that I would like to see. I don’t see enough of it. I see way too much just passivity.

I gave the closing lecture to my class here at Stanford last Friday, and one of the things I told them was every time they hear … “AI” they should think of amplified intention, that AI is the greatest amplifier of intention ever. Whatever they want to achieve, AI can amplify it, but it’s not going to happen without that agency, without that intention.

So that’s a message I want to have for the broader world: “Let’s think about what our values are, what kind of economy we want to shape, what kind of shared prosperity we want to create.” And I mean that for everybody, not just the technologists and sending that message to them, but also to policy makers, also to citizens, to workers, to managers. Let’s think about really actively using this.

Over the next 10 years, I expect the world is going to radically change. But there’s no one inevitable future. [When] you cross the border between, say, [the] United States and Mexico or other countries, you can see that different institutions lead to very different outcomes, different levels of wealth and democracy, and that’s the same for the future. We can live in a lot of different possible futures. One of the things I’m doing here at the Digital Economy Lab is a lot of research to understand which paths are more likely to lead to those beneficial outcomes.

Sam Ransbotham: I think that’s a great way to end here. We [do not] have easy answers in either direction. The gains are real, the harm is real, and which one wins is a choice that we’re making now and something that we have control of.

Erik Brynjolfsson: More than ever. One of the interesting things is this is the time for philosophers and for humanities and for economists and people who think about what kinds of values we want to instantiate in the world. With this greater power comes greater responsibility and more potential to change the world. So we really need to think consciously about what kind of world we want to shape going forward. I don’t think there’s been enough attention to that, but the Digital Economy Lab is in part focused on that.

Sam Ransbotham: Thanks for bringing the numbers and the nuance. We’ll see you in Boston this summer.

Erik Brynjolfsson: Absolutely. Looking forward to it, Sam. Great talking to you.

Allison Ryder: Thanks, everyone, for listening today. We’re off for summer break, and we’ll be back in the fall with an exciting lineup of speakers from Instacart, GoFundMe, Dropbox, and Honeywell. Please join us then.

In the meantime, it helps our show a lot if you leave us an Apple Podcasts review or a Spotify rating. If our show has really impacted your life and work, please drop us a line. You can reach us at smr-podcast@mit.edu. We’d love to hear from you.

Balancing the daily grind of operations with the spark of innovation is one of the most persistent challenges managers face.1 Regulatory compliance, performance metrics, and the never-ending demands of customers can easily crowd out creative thinking. Work overload or mismanaged operations elevate employee stress, which directly constrains ideation capacity. This cognitive bottleneck triggers a cascade of quality incidents that require downstream intervention, further inflating workloads. The result is a self-reinforcing loop, where the lack of front-line ideas accelerates systemic stress.

Of the 150 hospital front-line workers we surveyed recently, over 80% said they wanted to engage in innovative activities. But fewer than half felt that they had that opportunity, and 70% said that they rarely or never communicate their ideas. Front-line employees often work under immense pressure caused by stressors such as regulations, staffing shortages, and unyielding time constraints. In that context, creativity can feel like a luxury — or, worse, a risk. And yet, as research has shown time and time again, the people closest to the work — front-line employees — are often the ones best equipped to improve it.2

Even as artificial intelligence increasingly shapes innovation processes, the deep contextual knowledge, empathy, and creativity of front-line employees remain indispensable to meaningful innovation. Equally critical — yet increasingly endangered—are middle managers. Although many global firms are aggressively thinning this organizational layer in anticipation of an AI-driven future, middle managers play an essential role in fostering their front-line teams’ innovation.

Our research on effective behaviors in nursing departments at several maximum-care hospitals — where new ideas can literally save lives — demonstrates how top and middle managers can help front-line workers strike a balance between innovation and day-to-day tasks.3 A critical success factor is the alignment of top and middle management around a shared ambition: the effective use of their respective roles to shape employees’ innovation identity and confidence. Below, we offer concrete recommendations and highlight common managerial pitfalls to avoid.

When There’s No Reward for Innovation

Consider the following anecdote from our interviews. When Sam, a hospital nurse, designed a new “intensive care diary” for families and patients recovering from critical illnesses, he wasn’t trying to launch a grand innovation project. He just saw that something was missing. “I thought it could help people process what they’ve been through,” he explained. “The team was open to it — some even said they’d love to try it.” Sam developed a ready-to-go concept, printed several prototypes, and even planned a short workshop to introduce it to colleagues during overlapping shifts.

Then the idea stalled. His unit’s management showed little interest, fearing that implementation of his idea might escalate into a larger project. The underlying message conveyed a disregard for his creative initiative, leading to a missed opportunity to actively shape the employee’s understanding of his creative role. Worse, it discouraged further contribution, never mind the motivation to go “above and beyond.”

For example, Sam knew that the repetitive, time-consuming checks he and his cardiology intensive care unit colleagues had to do each night to ensure that each emergency crash cart was complete and organized were largely unnecessary. And he believed that the carts could be sealed with numbered tamperproof tags so that some nightly checks could be skipped. The materials were there. The process was simple. And yet, Sam hesitated to suggest the change.

Seeing no benefit in voicing further ideas, Sam stopped doing so, leaving his potential untapped. “I still think it’s a good idea,” he said, laughing softly. “Same with a few other things — like improving the way we order medications or organize bed assignments. I just don’t bring it up anymore.” And every night, he and his nursing colleagues continued to spend (waste?) time checking crash carts.

Sam’s hesitation reflects a common tension: Employees see opportunities for improvement every day but are unsure whether acting on them is part of their role, especially if initial initiatives were not supported. The subtle hesitation — the space between noticing and acting — reveals a deeper issue: When employees don’t feel empowered to recommend improvements, or don’t feel accountable for them, creativity fades. Innovative ideas dwindle, not due to lack of interest but because of the absence of organizational structures and managerial support that validate innovation as a core capability.

Middle Managers: The Overlooked Innovation Engine

How can managers help workers strike a balance between high operational efficiency and necessary creative learning? It requires that employees develop a robust innovation identity — the understanding that generating ideas is an essential part of the job. Lacking this sense of identity, one worker explained, “It’s not in my job description to change things. My job is to do my work, go home, and everything’s fine.”

Even when employees feel psychologically safe and know they won’t be punished for speaking up, they still may not see creativity as part of their role. The question, then, becomes “What can make employees see it differently?” The answer often starts with their supervisors.

Middle managers are decisively close to front-line employees and play a large role in their sense of identity as innovative problem solvers or mere task-doers.4 Cutting management layers to “streamline” operations risks removing the very people positioned to encourage innovation. Through everyday interactions, such as encouraging inputs, backing small experiments, and recognizing contributions, middle managers can quietly build — or erode — their employees’ innovation identities while establishing idea generation as a regular and essential practice.

While having too many management layers can breed bureaucracy, removing middle management altogether can leave front-line innovation without its most important catalyst: a manager close enough to care and credible enough to champion ideas upward. AI is taking over more tasks but lacks deep, situated, tacit knowledge derived from moment-to-moment operational friction between humans. Front-line employees like Sam are oftentimes the only source of nuanced, contextualized insight necessary to identify truly nonobvious problems and generate targeted, human-centered solutions.

How to Nurture Middle-Manager Innovation Champions

Middle managers are often considered the bridge between strategy and execution, and rightly so. They translate vision into daily practice, turning abstract goals into concrete tasks. But our research shows that top managers and senior leaders play an equally instrumental yet easily overlooked role: They support middle managers’ nurturance of front-line innovation.

When senior leaders clearly connect front-line ideas to the organization’s mission — linking, for example, nurses’ suggestions directly to patient safety or operational efficiency — they shift the perception of creativity from “nice to have” to “essential.” Their role is not to micromanage but to enable innovation: providing clarity, trust, and structural support so that creative work becomes part of “real work,” not an optional add-on. One executive put it bluntly: “People always look at the leader: Do they support it? Do they care? Leadership is the decisive factor.”

Many senior leaders encourage their teams to be bold and experimental but fail to recognize that their own behavior sets the tone. If top management clings to old habits, so will everyone else. Authentic leaders demonstrate a growth mindset while simultaneously acknowledging the discomfort of operational gaps. Spotting glitches, working on them, and learning from the experience are not signs of weakness but of growth. This modeling allows middle managers to internalize a similar mindset and, in turn, further foster it.

Many organizations, however, underestimate how long it takes to build a culture in which people are not only confident to develop and communicate improvement suggestions but believe that they are expected to. One leader reflected that his organization had spent five years trying get people to understand that they wouldn’t be punished for experiments and ideas that didn’t pan out. When such a message is repeated consistently and authentically, it lays the foundation for innovation. Middle managers can acknowledge the potential of front-line ideas and be empowered to take charge, experiment, make mistakes, and continuously improve, contributing to a culture where creativity can thrive.

When Small Nudges Lead to Big Ideas: Practical Guidelines for Leaders

In balancing daily operations and innovative initiatives, timing and tactics matter. Overload your managers with new (top-down) initiatives, and creativity can collapse under the weight of “too much.” Respecting the role’s “sandwich” position means leaving enough time for mid-level managers to actively engage with their direct reports in constructive and innovative ways. One executive described the art of subtle prompting this way: “Leaders need to provide strong support to ensure that employees are allowed to participate in projects and aren’t constantly pushed back into their daily routines. It’s the supervisors who must create that framework.” Great leaders sense when to step in and when to step back, and make space for creativity without making it feel like yet another obligation.

Recommendations for Middle Managers

These actionable recommendations can help managers use their unique role to help front-line innovation thrive, by shifting from mere supervisors to coaches of problem-framing and champions of micro-experimentation. By engaging in the behaviors below, middle managers can spark a virtuous cycle that builds an appreciative culture over time and allows creative ideas to emerge as team members internalize ideation as part of their job. This shift in self-perception drives consistent idea generation, which in turn provides proof of ability and impact, further strengthening employees’ front-line innovation identities.

Allow small-scale experimentation. Use a “local-first” filter to prioritize ideas that are independent and testable at the local level. Fast-track these small, testable, and independent suggestions, such as Sam’s crash-cart tagging system, for immediate, small-scale testing. Resist the urge to immediately scale complex ideas.

Establish protected exploration. Actively carve out and protect dedicated “idea spaces” separate from daily operational demands. Use structured problem-framing discussions in team meetings: Identify problems in a first step, and then link front-line employees’ capabilities to potential solutions.

Apply process rigor. Commit to providing a concrete next step or reasoned feedback for every idea quickly, ideally within a few days. Ambiguity or slow response times can curtail new initiatives while eroding employees’ innovation identities.

Set the example of learning from failure. Ensure that every small-scale experiment that underperforms or fails is followed up with a formal, structured debrief: What did we learn? How does this inform the next iteration? It’s crucial to omit any mention of blame. Learning from experiments encourages a positive culture that accepts mistakes as necessary steps on the path to successful improvements.

Recognize and reinforce new ideas. Provide personalized, nonmonetary acknowledgment — such as a personal note or public recognition — for the act of voicing an idea, regardless of its immediate viability. The majority of employees in our study were not primarily driven by monetary rewards; they sought recognition and to be taken seriously. They valued the opportunity to contribute meaningfully to organizational improvement.

Recommendations for Senior Leaders and Top Managers

Institutional support is indispensable in shaping how middle managers cultivate front-line innovation, and top-level leaders must deliberately direct support toward empowering the organization’s middle layer.

Intentionally support and train middle managers. Provide the necessary resources — including dedicated office blocks, leadership training and coaching, and structural support — to ensure that all managers have the capacity to dedicate time to develop their team’s innovation potential. Innovation support must be an expected, institutionalized, funded function, not a secondary duty.

Institutionalize knowledge transfer and strategic coherence. Promote regular exchange sessions, such as focused meetings between department or function leaders, to share proven approaches and ideas for fostering staff innovation. By structurally supporting this exchange, your organization can establish the appreciation and formal institutionalization of front-line ideas.

Foster strategic coherence. To ensure that ideas across the organization are evaluated fairly and transparently, clearly link front-line ideas to the organization’s mission; in hospitals, for example, that might include operational efficiency and employee safety. Without this common thread and objective criteria, an employee’s idea realization — and its potential impact — may rely on chance and individual heroics.

Model discomfort. Senior leaders must openly model the discomfort and learning inherent in change, just as middle managers must for their own teams. Publicly celebrate successful experiments, but also openly discuss the lessons learned from initiatives that failed, demonstrating that risk-taking is valued. Use lessons learned to carefully teach middle managers how to walk the thin line of taking risks while avoiding unnecessary operational disruption.

The System Behind the Spark

Front-line innovation rarely happens in solo, light-bulb moments. It’s more like a relay race, with ideas passed from the front lines to middle managers, then to senior leaders and back again. Each handoff requires clarity, trust, and good timing. In complex settings, the nimble ingenuity that front-line employees bring cannot be fully replaced by AI, because such innovation depends on hands-on operations and exposure to real-life processes.

Yet that involvement creates a fundamental tension: keeping operations running while also generating ideas to improve them. Middle managers have a critical role to play in helping front-line employees strike the right balance and while gaining a sense of innovation identity and confidence. In turn, senior leaders must enable these enablers. By setting a clear purpose, granting managers autonomy, and signaling consistent support, they can shape how middle managers lead and foster innovation within their teams.

When this system works, people like Sam no longer stay silent. They speak up because they understand that contributing ideas is part of their role — an individual shift that enables organizations to move beyond managing everyday operations to shaping the extraordinary.

As multinational companies implement artificial intelligence workflows and look to adopt AI across their global operations, they are increasingly running up against country-specific regulations and policies that aim to govern AI and align its use with national priorities and local cultural norms. These regulations and policies — which fall under the umbrella of “sovereign AI” — govern where data is stored and processed, whose infrastructure is used for training and operating AI models, and how algorithmic decisions are reviewed and enforced in a given jurisdiction. Many markets are now developing their own sovereign AI frameworks to reduce dependence on the United States and China, where nearly 70% of leading AI models originated.

This creates a strategic dilemma for multinationals. Relying on global AI platforms maintains operational consistency but deepens exposure to geopolitical disruption and local market access risk. Localizing data, infrastructure, and models earns regulatory trust but incurs significant cost and complexity when a company operates across dozens of jurisdictions with differing requirements. The challenge is that policies vary significantly by country and are evolving rapidly, making a single global AI strategy untenable, and fully independent local systems impractical.

Most companies are responding defensively, treating sovereign AI as a compliance obligation managed by legal or IT teams. Our December 2025 survey of 1,928 executives across 28 countries reveals a striking gap: Sixty percent of respondents said that rising geopolitical risk makes them more likely to pursue sovereign technology solutions, yet only 15% have made AI sovereignty a CEO or board-level priority, and fewer than 13% see it as a growth driver rather than a cost.

In this article, we argue that sovereignty is better understood as a continuum of choices and that the companies best positioned to scale AI globally are those that treat those choices as a source of competitive advantage rather than a constraint to be minimized.

The Sovereign AI Landscape

The regulatory landscape governing AI has shifted significantly in recent years, from relatively narrow data residency rules to a far broader set of requirements governing models, infrastructure, and how algorithmic decisions are made and enforced. Most major markets are now developing their own sovereign AI frameworks, resulting in a patchwork of locally governed ecosystems, each with distinct rules, data standards, and expectations for responsible use.

The complexity is already visible at every layer of the technology stack. Companies operating in European Union member states must conduct formal risk assessments, maintain detailed technical documentation, and submit it to national market surveillance authorities under the EU AI Act. (The AI Act is now in active enforcement, with its most comprehensive requirements for high-risk AI systems taking effect in August 2026.) Simultaneously, they must comply with prior standards and policies, like GDPR (General Data Protection Regulation), NIS2 (Network and Information Security 2), and DORA (Digital Operational Resilience Act). Companies must also prepare to align with the sovereignty package announced by the European Commission in June 2026, which includes two legislative proposals and a strategic road map to bolster the EU’s AI sovereignty.

Meanwhile, companies looking to do business in Saudi Arabia must navigate strict data localization requirements — including obligations to store nationally sensitive data within the country — alongside cross-border transfer rules that require adequacy assessments or contractual safeguards, all within a governance framework that is still taking shape. There is no dedicated AI law, and binding obligations currently flow from data protection and cybersecurity regulations rather than AI-specific legislation. The same governments driving these requirements are also pouring billions of dollars into building the infrastructure and incentives to enable sovereign AI solutions.

Companies have started to develop strategies to navigate this fragmented landscape. We know from our consulting work that three global banks are rethinking their tech strategy in the EU: They’re limiting further migration of sensitive systems into foreign public cloud systems and instead building a shared platform in their home countries.

In the U.K., senior leaders of multinational banks are exploring a domestic alternative to Visa and Mastercard to reduce reliance on U.S.-owned payment networks. These are early signals of a structural shift in how multinationals must think about AI infrastructure. This raises an urgent question for CEOs: How do you scale AI globally when the rules governing it are local, fragmented, and still being written?

To answer this question, CEOs need to make three strategic choices: where accountability for decisions should sit, how much sovereignty their operations require, and which external partners can help them execute.

1. Make sovereignty a strategic priority.

The first and most urgent CEO decision is raising AI sovereignty to the level of a strategic concern. Our survey found that most organizations delegate decisions regarding AI sovereignty to chief data/AI officers (37%) or compliance/risk officers (28%), while only 15% of organizations have made it a CEO- or board-level priority. When sovereignty sits in IT or compliance, it results in fragmented decisions across business units, inconsistent approaches across markets, and missed opportunities to turn sovereignty into local advantage. Sovereign AI is not an IT architecture choice. It is a strategic bet involving geopolitics, capital allocation, supply chain resilience, and long-term competitiveness, and the decisions it requires can be made only at the top.

What does that look like in practice? Consider BNP Paribas, one of the largest banks in the EU. Since 2023, it has built a deepening partnership with Mistral AI, Europe’s leading sovereign AI model provider, culminating in a groupwide multiyear agreement in 2024 and a renewed three-year extension in 2025 covering software, co-development research, and on-premises deployment. The bank also backed Mistral financially, participating in both its 385 million euro ($445 million) funding round in 2023 and its $640 million Series B in 2024. The partnership is driven by the C-suite with sovereignty as a key consideration. Keeping AI on-premise, under the bank’s direct control, ensures sensitive data stays within European regulatory jurisdiction. Such decisions — say, which AI ecosystem to depend on, which regulatory frameworks to operate within, how to manage geopolitical exposure — carry implications for capital allocation, competitive positioning and long-term resilience. They belong on the C-suite and board agenda.

2. Treat sovereignty as a continuum.

As such CEO-level choices demonstrate, sovereign AI does not require full independence or the wholesale replacement of existing systems. CEOs must recognize that sovereignty is a continuum. Depending on the industry, national context, and specific use case, aspects of the company’s technology stack may need varying degrees of adjustment to meet sovereignty concerns. Only about one-third of all executives we surveyed believe that their AI workloads require any such sovereign adjustment. Those actions can involve anything from implementing targeted safeguards around data residency or legal oversight to replacing entire elements of the technology stack, such as models or infrastructure, to satisfy sovereignty requirements.

The degree to which sovereignty matters to the design of a company’s AI system rests on three considerations.

The first is industry risk. Defense, health care, energy, and financial services are more sensitive industries because AI systems have implications for national security, citizen safety, and economic stability. Governments are likely to have greater AI sovereignty concerns for companies operating in those domains. In contrast, retail, tourism, and consumer services can often rely on global platforms when strong local safeguards are in place.

The second is national context. Countries pursue different strategies shaped by geopolitics and economic priorities. China has built a full-stack China-for-China model. Singapore emphasizes interoperability and cross-border trust. The U.K. and much of Europe favor hybrid approaches that combine global platforms with selective local control. Companies operating across these markets must design for variation.

The third and most decisive consideration is the use case. Even within an industry, AI that is used in high-stakes decision-making, such as credit decisions, medical diagnostics, or energy grid optimization, carries far greater risk than AI used in contexts like customer service, marketing personalization, or internal productivity. High-stakes use cases call for greater scrutiny around data governance, model transparency, and regulatory exposure.

All three considerations are relevant to AstraZeneca, a global pharmaceutical company. Pharmaceuticals is a sensitive industry, so AI use carries high stakes by default. But AstraZeneca calibrates its sovereign controls deliberately by national context and use case. In China, adverse drug reaction data must be reported to the National Medical Products Administration and remain within the country under strict governance. To address this requirement, AstraZeneca is using Alibaba Cloud’s sovereign infrastructure to deploy local AI models, such as the Qwen large language model, for pharmacovigilance in a private, locally controlled environment.

Outside China, AstraZeneca made a different calculation. For R&D and clinical development — work that is sensitive but not subject to the same regulations as pharmacovigilance — it uses the Amazon Web Services (AWS) public cloud to run large-scale AI and machine learning, including multi-agent systems that enable scientists to query complex biomedical data and generate insights in seconds. The same company, facing different sovereign AI pressures in different contexts, arrived at two different infrastructure decisions.

3. Build hybrid sovereign ecosystems.

Few organizations can, or should, build the full AI stack independently. Our survey found that 55% of organizations plan to use a mix of global and local AI providers, reflecting a shift from dependency toward flexibility.

CEOs must adopt a tailored mix of global and local AI providers, depending on use cases, the regulatory landscape, and risk tolerance.

For organizations seeking global scale, partnering with hyperscalers such as AWS, Google, Microsoft, and Oracle is an attractive option. In software and platforms, Oracle’s EU Sovereign Cloud and Microsoft’s Delos Cloud partnerships are enabling European companies to run sensitive workloads fully under EU law.

When earning local trust is a priority, partnering with country-endorsed national champions may be the right strategic move. Indosat Ooredoo Hutchison — Indonesia’s trusted national telco, jointly owned by Qatari and Hong Kong-based investors — is building Indonesia’s first sovereign AI cloud with international partners Accenture and Nvidia while ensuring that national data stays onshore to support local startups and government clients.

Companies that want to avoid full dependence on hyperscalers can access specialized capacity from AI-native players like Nebius, CoreWeave, and Lambda. These providers build infrastructure specifically for AI workloads, allowing companies to access high-performance compute more efficiently, often with regionally deployed capacity that meets local regulatory or data requirements. Helical, a Europe-based biotech company, trains its large-scale biological AI models on Nebius’s infrastructure. Nebius is headquartered in Amsterdam and operates data centers across Europe, giving it the compute performance it needs while keeping data within regional boundaries.

When the deepest level of sovereign control is required, and where competitive advantage comes from shared capability rather than proprietary infrastructure, federated consortia offer something other models cannot. The OpenBind Consortium, for example, received 8 million pounds ($10.8 million) from the U.K. government’s Sovereign AI Unit to bring together a group of partners that includes the University of Oxford, Diamond Light Source, and MedChemica to build a nationally governed data set for AI-driven drug discovery that is 20 times larger than earlier efforts. The real advantage, however, lies in the sovereign governance model: It allows companies to collaborate, share risk, and scale AI with confidence — something fragmented, ad hoc data sets cannot deliver.

The Strategic Imperative

Most companies focus sovereignty on the data and cloud layers of their tech stack. Our survey found that while 60% apply sovereignty controls to data and 46% to infrastructure, only 22% extend them to AI models themselves, leaving a critical layer exposed. As AI systems become more autonomous and agentic, the decisions and actions that matter most will increasingly happen at the model and agent layers.

This doesn’t mean that every company needs full-stack control. Companies need to apply sovereignty to the layers that matter most for their highest-stakes use cases.

In a world of persistent geopolitical uncertainty, sovereign AI has become a competitive capability.

Companies that treat it as a strategic design choice rather than a compliance obligation will shape how AI evolves in their markets. Those that don’t risk discovering too late that the intelligence driving their most critical decisions is no longer under their control.

Digital platforms and generative AI have lowered the barriers to accessing global talent, capital, and knowledge for companies everywhere while making it possible to reach customers across languages and cultures. Research my colleagues and I have conducted suggests that such tools make it easier for entrepreneurs to serve global markets and for investors to evaluate startups from afar.

Access to those technologies should result in a leveling of the global playing field that allows new ventures to thrive anywhere. Promising early-stage startups are emerging in areas like Jakarta, Nairobi, Kyiv, and São Paulo. But when it comes to scaling, the old pattern remains: Companies that scale and become category leaders are disproportionately concentrated in a handful of traditional hubs, such as Silicon Valley, while early-stage companies outside of them struggle to scale into larger businesses. Technology, it appears, is not enough to overcome the barriers to scaling.

When Technology Reduces Some Gaps — but Creates Others

Even as digital technologies reduce structural differences across locations, some companies respond to the new opportunities they create in ways that undermine the ability to scale. When entering new markets or adopting new technologies becomes as simple as a click, businesses can fall into one of two traps: either chasing every available opportunity or defaulting to what is closest and most convenient. Both responses can systematically disadvantage companies outside major hubs.

The first trap stems from the urge to go global before the company is ready. Today, companies can attract users from around the world with a single post on a digital product platform. As a result, many organizations — especially those in smaller markets under pressure to show global traction — try to pursue multiple global markets at once. But in doing so, they often overlook a critical resource: early users whose feedback they could more easily interpret because they share a common background or geography. My research shows that business leaders can more easily recognize the demand signals of local users and, as a result, learn more effectively about their company’s nascent product and refine it before expanding further.

Generative AI is making market expansion more complex, particularly for companies based outside English-speaking hubs. Research I conducted with colleagues shows that while GenAI helps high-quality ventures in these contexts stand out by improving the pitches of expert entrepreneurs more than those of non-experts, it disproportionately enhances pitch quality in English-speaking environments. As long as investors and customers rely on pitches as an input, ventures in non-English-speaking hubs may continue to face disadvantages when entering hub markets. This challenge is particularly acute for non-hub firms because they rely more heavily on text-based signals to reach global audiences.

Yet overcorrecting to avoid this trap can lead to the second trap: defaulting to what is close and familiar. This is particularly evident in startups’ technology choices. The global boom in tech entrepreneurship has dramatically expanded the set of available tools, many of which have the potential to accelerate adopters’ growth. At the same time, tool vendors are increasingly relying on GenAI to craft persuasive product descriptions, meaning that a well-written pitch is no longer a reliable signal of a tool’s value. This makes it harder for potential buyers to distinguish tools that will accelerate growth from those that are merely well packaged. When managers face too many persuasive options, they often fall back on simple rules — like choosing to adopt tools that were locally developed or are already familiar to them. Indeed, in research I coauthored, judges evaluating many polished pitches in a startup accelerator competition favored ventures from their own regions — even though the judges were no better at assessing local ventures’ quality — and passed over 1 in 20 promising startups as a result. The same bias shapes how companies choose their technologies.

Defaulting to what is local as a heuristic can particularly penalize companies in remote markets. These companies often encounter fewer locally developed or locally targeted tools — partly because tool providers themselves face pressure to cater to hub-based customers. Ongoing research by my colleagues and me shows that, as a result, genuinely useful technologies often go unseen, are misunderstood, or are deprioritized by the companies that could benefit most from them.

In this way, AI can unintentionally reinforce geographic disparities rather than eliminate them unless companies bring something that technology alone cannot provide: strategic focus.

Strategy as the Missing Equalizer

Prioritizing what is nearby can narrow the opportunity set for companies in remote locations because frontier innovations are often concentrated in hub markets. At the same time, pursuing every distant opportunity — which is often encouraged via external investor pressure — can diffuse scarce resources and weaken execution.

The solution is strategic clarity: a clear articulation of how a company intends to combine local and distant opportunities to create a differentiated position in the market. This begins with a single question: What is your competitive advantage? Are you serving a market that competitors have largely ignored? In this case, your advantage may lie in access — bringing a technology or use case to customers who have been overlooked. Or are you competing in an established market by offering a superior solution? Here, the advantage may be quality — delivering better performance than existing alternatives.

This distinction has direct implications for subsequent technology and market choices that can either widen or narrow the global scaling gap. When a company’s strategy centers on serving an underserved local market that it knows exceptionally well, the key may be to adapt an existing technology to the needs of that market.

For example, Grab and GoJek, founded in Malaysia and Indonesia, respectively, adapted the ride-sharing model pioneered by Uber to conditions that a hub-based competitor wouldn’t have easily been able to read or replicate: cash payments in largely unbanked markets, motorbike taxis suited to dense urban traffic, and an eventual expansion into food delivery and financial services that matched how people in the region lived and spent. Lack of local knowledge became a barrier to entry for global competitors.

When the advantage is quality, the strategic imperative is different: Identify distant markets where demand for a superior solution is strongest, and draw on differentiated local assets that hub-based competitors cannot easily access — such as exceptional software engineers or exclusive access to local university research labs — to deliver it. 

Grammarly, which has its roots in Ukraine, illustrates this logic well. Its founders used the country’s deep pool of local developers to build a technically superior writing-assistance tool — harnessing a local talent advantage that hub-based competitors couldn’t easily replicate. Its value proposition meant that its target market didn’t have to be local; they could go after global professional and academic users concentrated in English-speaking markets. Spotify similarly leveraged its exceptional local internet infrastructure in Sweden to create a product that targeted and ultimately transformed the global music industry.

In this way, strategic clarity transforms the search for technologies and markets from a reactive scramble to a deliberate capability that is available to companies everywhere, not just those situated in hubs.

The practical question is how to build this strategic clarity. Start by asking yourself four essential questions:

1. What is our core value proposition? How does your offering improve customers’ lives, processes, or outcomes? What makes it difficult for competitors to replicate it?

2. What is the target market that benefits most from this value proposition? Are you addressing an underserved segment — perhaps close to home? Or competing for customers in an established market, potentially abroad? Strategic focus begins with clarity about who gains the most from what you offer.

3. How will we deploy technology to reinforce our competitive advantage? If you are targeting an underserved market, harness your unique local knowledge to adapt globally available technologies to needs others have overlooked. If you are targeting established markets abroad, draw on differentiated local assets that hub-based competitors cannot easily access or replicate.

4. Where should we begin? Early adopters should resemble your eventual target customers while also providing feedback you can clearly interpret. For example, if the local market, where you can get clearer feedback, resembles the target market, start there; if not, start with the target market. The right testing ground generates learning signals — not just early traction.

Digital technologies, including AI, are not automatic solutions to geographic disadvantage. Without strategic clarity, they can just as easily widen the gap as close it. Grab, GoJek, Grammarly, and Spotify were able to achieve scale because they understood and exploited what made them different. With strategic clarity, location becomes a choice rather than a constraint, and technology becomes what it was always meant to be: not a substitute for strategy but a powerful amplifier of it.

To understand their customers and markets, a growing number of customer-oriented companies are using generative AI tools, alongside the language and reasoning capabilities of popular large language models (LLMs), to access and analyze their own internal content. These hybrid knowledge approaches, which typically employ a technique called retrieval-augmented generation (RAG), allow the integration of a company’s own customer insights with the general knowledge base on which an LLM was trained. Companies taking this approach to what was previously called “knowledge management” reap several benefits, such as enabling employees to access and summarize content in natural language. That capability is particularly important in large organizations — where employees searching for insights often have no idea where those insights originated or how they might be found.

Organizations gather and attend to insights about what their customers want, how they want to be sold to, and what products and services they have interest in. Customer insights typically originate from market research departments or external market research agencies, but they can also be found in sales interactions, customer letters and emails, website behaviors, social media interactions, customer service tickets, purchase patterns, focus groups, and other channels. It can add up to a lot of structured and unstructured information, so tools to help summarize, categorize, store, and access it certainly make sense.

However, companies that focus exclusively on storing and accessing knowledge are making the same error that many organizations made in the earlier generation of knowledge management: That focus is too narrow. Instead, companies should also address knowledge flows — how customer and market insights are created, analyzed, stored, and accessed. New technologies — using generative AI to at least some degree — can assist with all of those steps.

While earlier tools for knowledge management (such as Lotus Notes and Microsoft SharePoint) allowed broad access to customer and market insight content, many organizations did not experience a revolution in insights access and usage. The technology provided more access, but cultural challenges — difficulty in organizing and retrieving the knowledge, indifference to the content, organizational silos and overlaps, lack of collaboration with external agencies — often prevailed. Today, organizations still face those same cultural obstacles when tackling knowledge management.

To learn how generative AI can and can’t help leaders beat knowledge management obstacles, we spoke with customer and market insight specialists or leaders at eight different consumer-oriented companies. Some of these leaders had responsibility for creating and overseeing the generative AI system; others had broader responsibility for market research or insights-oriented technology. We also spoke with several vendors of GenAI-based technology for customer and market insights, and one market research agency that makes extensive use of GenAI-based qualitative research technology.

How Generative AI Tools Help

Companies can combine GenAI tools with customer and market insight content on their own (though it may be difficult), use external software vendors to simplify the work, or use a combination of both approaches. Procter & Gamble, for example, uses vendor-supplied software for knowledge storage and access but has created its own system for GenAI-based analysis and categorization of the content. As Kirti Singh, P&G’s chief analytics, insights and media officer, noted, this way they get “sharp, pointed answers from GenAI” rather than just links to documents.

Among the tool vendors, focus differs. Some primarily focus on the storage of and access to insights, but the tools’ value goes beyond providing simple automated virtual filing cabinets. Their functions include automated curation of documents, integration of diverse content types, on-demand analysis done in response to queries, and synthesized answers to prompts. This type of tool, however, assumes that data analysis has already been done and that insights are waiting to be found. Analysis of quantitative data is done by analytical software, and the capacity to analyze qualitative data is limited.

Other vendors concentrate on the analysis of qualitative customer data and documents for customer insights. Still other vendors emphasize rapid testing of consumer responses to advertising. Today, there is overlap among these categories, and most vendors are attempting to address the broad process of identifying or creating customer insights, curating and categorizing them, and making them available for later access. We expect that at some point, broad customer insights platforms will emerge, employing generative AI and other capabilities to address the entire process.

If the vendor’s primary function is insight storage and access, the approach typically involves adding the customer’s custom and proprietary content on top of large language models. The content stored can include structured data (quantitative market research results, spreadsheets, or customer satisfaction ratings, for example) and unstructured data (such as transcripts of interviews, social media comments, or focus group results) from both internal and external sources.

In most cases, organizations pursuing this route centralize as much customer and market content as possible in one system. Curation is required to reduce content overlaps, eliminate obsolete/irrelevant documents, and generally maintain quality. As one customer insights leader told us, “AI is only as useful as the data it learns from.” To make sense of insights, the GenAI tools typically must tackle categorization, summarization, and content tagging. Tagging the content makes it more likely to be retrieved later. Some companies use manual tagging, while some systems employ GenAI-based tagging using a predefined taxonomy.

Novartis offers an example of a company successfully revamping insight storage and access using GenAI tools. Working with an external vendor, the company developed a customer and market insights system, called Sherlock, for its consumer business. After users pose questions, the system gives answers by pointing to a specific line of text or a time stamp in a video. Sherlock also incorporates expert-curated microsites, known as Knowledge Zones, on particular topics, such as packaging. Users who add content to the system must adhere to strict governance guidelines about document formats and quality. Novartis’s research vendors can upload project deliverables directly into Sherlock.

The system helps Novartis avoid spending on redundant insights services across its business and helps employees find relevant insights quickly, without overgeneralization. (For example, it could flag results that were based on patient data from Europe only, using a feature called WatchOut.) The results have added up: Novartis saved more than $29 million in primary market research costs in just one year. Such use of GenAI to enhance insight storage and access can facilitate the democratization of information by helping employees find both knowledge and knowledgeable people.

Qualitative Data Analysis: A Special Problem

For a long time, the ability to do qualitative data analysis — a messy business — has not been included in off-the-shelf analytical tools. Specialized software for this purpose has mostly been used in academic qualitative research; historically, most market researchers have conducted semi-manual analyses, using spreadsheets. GenAI tools offer an alternative to this extremely time-consuming qualitative analysis work.

A recent academic article argued that GenAI is unsuited to qualitative analysis, but our analysis suggests that this is true only of generic AI chatbots. More specialized tools have other capabilities that make them suited to effective qualitative data analysis.

Tracy Tuten, who leads qualitative research at market research agency Illuminas (now part of Radius Insights), became an early adopter of a vendor’s generative AI software in order to mine customer insights. Tuten, who has taught market research at several universities, refers to this approach as “conversational qualitative data analysis.”

Via GenAI-based software, Tuten uses natural language prompts to analyze qualitative data from interviews and focus groups. The system lets her upload audio and video files for automatic transcription, generate summaries, surface themes, and compare them across audience segments. A large-scale qualitative project such as a global study with 30-plus interviews might have taken six weeks to analyze in the past but can now be synthesized in a day, Tuten said. The tool also lets her surface secondary insights that she might have missed in the unstructured data. Tuten often uses the software collaboratively with clients in workshops, enabling faster, more participatory insight discovery.

Given that many qualitative researchers previously relied on spreadsheets and manual cut-and-paste coding to analyze data, this AI-based approach represents a major advance in efficiency and rigor. However, uncritical use of generative AI or other forms of AI may have significant shortcomings. Conversational qualitative data analysis does not replace the researchers; it only augments their performance.

PepsiCo makes extensive use of software for creating customer and market knowledge, including both structured and unstructured data. The company has particularly focused on determining how customers respond to specific advertising campaign and brand messages. But that isn’t the only application the company has employed. In an interview with us and in The Consumer Insights Revolution, a book describing a transformation of customer insights at PepsiCo, Stephan Gans, senior vice president and chief customer insights and analytics officer, described the company’s “platform” for marketing research, called Ask Ada. It includes:

• The ability to test new creative content on real or synthetic customers.
• A data repository on the results from advertising, influencer, and other types of campaigns.
• Social listening capabilities.
• Predictive modeling.
• Knowledge management of customer insight lessons, present and past, including meta-learning.
• A conversational interface to all Ask Ada content.

Gans also credited Ask Ada with reducing PepsiCo’s dependence on external agencies and consultants.

Why AI Isn’t Enough: Four Challenging Factors

Despite these new capabilities, a GenAI tool cannot replace a marketing leader for strategy work. As Gans noted, “Raising the bar on marketing and innovation effectiveness to fuel commercial excellence will become increasingly automated. Leading the understanding of consumer demand is much more strategic and still requires humans.”

As we discovered in our research, several important issues inhibit the ability of AI technology to transform customer and market insights. These issues predate knowledge management and generative AI and will cause problems if not addressed. Let’s examine four of these factors, along with examples of organizations that have encountered and overcome them.

1. Geographical and business units lack common approaches.

One global consumer goods company where we conducted interviews had acquired a GenAI-based customer and market knowledge tool from a vendor but thus far had made little progress improving access to global knowledge. The problem was that the company does business in over 100 countries, and country-based units have a large degree of autonomy. There’s no companywide consensus on names for brands, categories, and distribution approaches. “We have pockets of knowledge, but they are very incoherent,” the head of knowledge management for insights told us. “Different people are investing in different things. We have contradicting numbers and outputs — they are all contextualized differently.” No senior executive has attempted to create greater commonality of information and knowledge formats across units; it would be viewed as counter to the company culture. As a result, the generative AI tool is used by only a few geographical units, and the company’s marketers and product developers are unable to learn from each other. The company is attempting to develop a strategy to create a more centralized, global approach.

At PepsiCo, the story is different. When Gans was first named chief consumer insights and analytics officer in 2017, the company had a diverse set of approaches to customer and market insights. But Gans wanted to create “one nation” of market research so that the company’s marketers could learn from one another and share relevant insights. With strong support from the CMO, Gans created the Global Insights Council, which comprised 15 insights leaders representing all regions and central/global capabilities. Today, customer insights are tightly integrated into PepsiCo’s innovation work.

2. Customer and market insights aren’t part of strategy and culture.

Even the best technologies won’t succeed if the organization’s decision makers aren’t ardent consumers of that type of knowledge. Without those passionate consumers of information, a company’s market research efforts will fall flat, our research showed. So culture and change management work will often be required of leaders.

Passion for data runs high at P&G, which is known for its long-term focus on being customer- and market-driven. Indeed, P&G recently celebrated its 100th year of market research; in 1924, the then-CEO asked a researcher to determine why customers were buying Ivory soap. P&G’s Singh told us, “At the heart of everything we do is the consumer. … Our strategy is to provide a superior product experience to our consumers. We employ experimental science, human and behavioral science, data science, and technology platform knowledge to understand our consumers.”

3. Agency relationships introduce data ownership complexity.

Many companies use external advertising and marketing agencies for consumer research. The client/agency relationship may lead to uncertainties and dysfunction involving analysis strategies, interpretations of analyses, and ongoing ownership of the data and results. If agencies end up owning all or most customer and market insights, a company’s employees will be unable to meet customer needs without external help. PepsiCo’s Gans argued strongly that the client company has to own all research results and insights created by agencies on the client’s behalf. He added that it’s not a good idea to own the data but then outsource the learning from it, because employees should apply lessons learned from market research in future campaigns.

However, for consumer-oriented companies that continue to work with agencies, some vendor software can facilitate collaboration between clients and agencies. Both parties can view, edit, and query customer research and produce a variety of outputs.

4. Analytics professionals may be seen as low-status “order takers.”

If that is the case in your organization, that reputation needs to change. At one of the consumer products companies where we conducted interviews, the insights and analytics function always had a library-like focus. Previously, internal customers who were interested in insights had to consult with a researcher or “librarian.” With the advent of a vendor-supplied GenAI tool, the function has been democratized.

However, users of the company’s system still treat it as a library; they don’t contribute much to the stock of insights. Enabling internal customers to serve themselves hasn’t substantially increased demand for the content and analysis. Users also don’t always supply high-quality prompts; they might ask, “What do we know about back to school?” not realizing that the company has decades of market research on the topic. As at Novartis, experts created micro-sites of curated content within the platform, to address particular information-access issues for certain areas. Still, budgets and head counts in the insights and analytics function have been cut in recent years. Similar functions outside the U.S. don’t want to pay for the GenAI-enabled tool, so they take different approaches to customer and market knowledge.

PepsiCo previously had something of an “order-taking” mentality for market research, and researchers were rarely asked to collaborate with the internal customers or help to shape the requests for insight. Research team members had little respect for their roles, and the function was asked to cut its budget several times. When Gans arrived in the leadership role for the function, he and the CMO concluded that PepsiCo was spending hundreds of millions of dollars per year on consumer insights and that it made little sense to do that unless the company were to become more customer-centric. He made a series of changes — including the implementation of a new AI and insights software system and the Ask Ada platform — that eventually made the customer insights and analytics organization well respected and well funded.

Overall, any AI tool should not be considered a replacement for what the organization is already doing well. As P&G’s Singh told us, “We brought together our tradition of being focused on understanding the customer with the latest AI solutions — not replacing, for example, customer home visits, but augmenting them with AI.”

Our study also suggests that the AI-enabled software for managing customer and market insights is evolving rapidly. Leaders are often interested in different features and functions that fit their company’s specific situation. But leaders should be aware that today’s software has limits. In poorly integrated global organizations, humans have created different names for customers, brands, and marketing approaches across geographies; even AI can’t pull together a unified set of customer and market insights in that case. Companies need humans to integrate and standardize that data to analyze it and act on it effectively.

Finally, if a company’s workforce isn’t actually interested in gathering and acting on customer and market insights, no software is likely to change that situation. These are problems that need to be addressed by humans, not AI.

What does artificial intelligence do to creativity? Are generative AI tools making us more creative, or less? Given that creativity is often the engine behind the most successful ideas and ventures, and that 83% of senior executives rank innovation among their top three priorities, understanding how using AI affects human creativity is critical for businesses.1 On the one hand, generative AI can act as a valuable brainstorming partner, enabling inventors and designers to rapidly prototype ideas and concepts. On the other hand, it risks inadvertently constraining creativity by narrowing the search space too early and encouraging users to anchor on AI-generated suggestions that seem “good enough.”

Across four recent studies, our research reveals that the truth lies beyond this simple binary. We have found that although AI can enhance individual creativity, it reduces collective creativity. To explain why this occurs, we should first clarify what we mean by creativity.

From Individual Creativity to Societal Innovation

Scholars typically define creativity as the intersection of novelty and usefulness.2 Novelty is the degree to which an idea or artifact is original or rare, and usefulness is the degree to which it is valuable or effective in achieving its purpose. An idea that is novel but useless or useful but unoriginal is not creative.

While these dimensions capture the creativity of a single idea, the dimension that best captures the creativity of a group of ideas is its diversity. Any collection of ideas may contain a few that are novel to some but obvious to others, or novel yet not useful. But a highly diverse set is more likely to contain a few highly original outliers that are both genuinely original and potentially valuable. This breadth provides teams with more raw material to recombine, compare, and refine over time, yielding products that better match the full range of customer preferences. After all, ideas that initially appear to be impractical can turn out to be breakthroughs once they have been refined or recombined: for example, the “failed” adhesive that became the Post-it Note or the abandoned video game whose internal communication tool became Slack. In other words, creativity requires more than just quantity and quality of output. It also requires diversity of output, where different ideas can spark new lines of inquiry, new speculation, and new seeds that breed new innovations.

By casting a wider net, organizations can guard against premature convergence on safe, conventional options, increasing the odds of surprising, high-impact breakthroughs. This, however, is where our research reveals an interesting paradox. Individually, AI often enhances creativity, particularly by enabling less experienced or less inherently creative individuals to generate more novel and useful ideas. But collectively, AI often “compresses” the idea space. Because many people anchor on similar AI-generated suggestions, outputs converge. A typical output produced with AI assistance is more creative, but the variance of the full set of outputs decreases. In short, even if an AI-inspired idea looks good, it may turn out to be similar to everyone else’s AI-inspired ideas.

For managers, the implication is profound. The challenge is to harness AI’s productivity and quality benefits while preserving the diversity of ideas that fuels long-term innovation.

Impact of AI on Idea Diversity

To understand how AI affects creative diversity, we analyzed four recent studies we worked on that span different creative domains: short-story writing, circular-economy solutions, humor caption contests, and collaborative storytelling. Despite the varied contexts, a consistent pattern emerged: AI assistance improved individual output quality while narrowing collective diversity.

In the first of these studies, two of us (Anil and Oliver) examined how access to AI influences the creative process and the diversity of collective output.3 In this experiment, participants were asked to write short, eight-sentence stories. Some people wrote entirely on their own, while others were given up to five three-sentence story seeds generated by an AI model. Independent evaluators rated the individual creativity of each participant’s story. We also used AI-based text analysis to measure the degree of semantic similarity among the stories, comparing those written with versus without AI assistance.

The results revealed the core tension. We found that AI assistance improved story novelty, especially for writers with a lower baseline level of creativity (as measured beforehand with an existing paradigm, the Divergent Association Task). Yet at the collective level, diversity declined. Stories from the AI-assisted groups converged on more similar beats or structures, showing less variance than those written without AI. (See the story-writing graphics.) This suggests a social dilemma: While individuals gain from AI, especially those who struggle most with creative tasks, widespread reliance risks narrowing the collective pool of ideas, leaving us with higher average quality but fewer distinctive outliers.

In a second study one of us worked on (Léonard, with four collaborators), this pattern held in a very different domain.4 We asked participants to propose circular-economy solutions to address sustainability challenges, such as repurposing waste materials. A human-only crowd produced a broad range of ideas, from conventional recycling proposals to unique, unconventional ones such as innovative bricks made from foundry dust and waste plastic with a Lego-like interlocking design to reduce construction-related air pollution. In contrast, a single human working with AI often surpassed the crowd in independent evaluators’ ratings of overall quality, strategic viability, and financial and environmental value. But the human crowd scored higher on novelty, and the unusual ideas that might spark breakthroughs emerged mostly from the human-only group. Once again, AI raised the floor of performance but narrowed the variance in outputs. (See the circular-economy graphic.)

A third study pinpointed where in the creative process this narrowing occurs.5 One of us (Kartik, with a co-researcher) conducted a randomized experiment modeled on the New Yorker cartoon caption contest to test whether AI reduced the diversity of creative outputs during the idea-generation or idea-evaluation stage. We tested four collaboration designs: human-only, AI use for idea generation alone, human idea generation with AI use for idea selection alone, and AI support during both phases. The study found that AI boosted both the quantity and average quality of ideas, with the greatest gains when it supported both the generation and evaluation stages. Importantly, the diversity effects diverged depending on when AI was used: AI in idea generation consistently reduced diversity, whereas AI in idea selection preserved variety at levels comparable to those of human-only work. (See the humor caption contest graphic). This finding suggests a potential way forward: The stage at which AI enters the workflow may matter as much as whether it is used at all.

A fourth study that one of us worked on (Kartik and a co-researcher) tested this insight directly.6 It examined how different human-AI collaboration models affect both the quality and the diversity of story writing, as well as the impact of those collaboration models on self-reported writer satisfaction. Four designs were tested: human-only, human-led ideation with AI drafting, AI-led creation with human approval, and continuous human-AI collaboration throughout (the “copilot” scenario). The results confirmed the pattern observed in the previous study: Ceding creative control to AI produced the most homogeneous outputs; having humans and AI work together as copilots mitigated the effect to some extent; and keeping humans in charge of early creative tasks preserved significantly more diversity — approaching the variety seen in fully human work. (See the Hosanagar and Ahn story-writing graphic.) This indicated a clear design principle: whether diversity survives depends on where humans are introduced into the workflow.

The graphs above visualize these findings across all four studies. To measure the diversity effect, we used AI techniques to calculate the similarity between the various outputs within each group and then averaged those scores. Think of it as a clustering metric: A higher similarity score means that ideas bunched together; a lower score means that they were spread out across a wider creative space. In each panel, the horizontal axis represents the average similarity score, and the distribution curve shows the frequency of that score among participants. When AI is involved, the curves consistently shift to the right — toward higher similarity — signaling that the outputs become more alike.

How to Use AI in Creative Workflows (Without Sacrificing Diversity)

The evidence across studies points to one conclusion: How you use AI in creative work matters as much as whether you use it at all. Leaders who seek the efficiency gains of AI while preserving or enhancing originality must intentionally design their workflows. Here are some practical strategies you can use.

1. Keep humans in the driver’s seat for ideation. The fourth study described earlier, which tested different modes of human-AI collaboration in story writing, offers direct guidance here. Participants who retained responsibility for ideation produced stories that were rated higher by independent evaluators in terms of interestingness and overall quality, and they reported greater satisfaction. The diversity effects were equally important: Ceding creative control to AI produced the most homogeneous outputs, while keeping humans in charge of early creative tasks resulted in significantly more diversity, approaching the variety seen in fully human work. Even the copilot model, which involved AI throughout, narrowed the diversity of output compared with human-led ideation.

This has immediate practical implications for managers: Let humans take the lead in any creative and innovative workflow to capture more unique ideas. Start by having a team sketch ideas or draft early outlines before integrating AI into the process. This sequencing preserves variety while still capturing efficiency — and ensures that AI complements, rather than substitutes for, the uniquely human capacity to make messy, surprising leaps.

Video game producer Ubisoft’s in-house AI tool, Ghostwriter, offers a concrete illustration of this sequencing in action. Designed to assist scriptwriters working on large open-world games, Ghostwriter takes on one of the most repetitive narrative tasks: generating first drafts of short lines of dialogue spoken by nonplayer characters. Crucially, the tool does not replace the writer’s role in shaping character or story; scriptwriters define the character and context first and then select and edit from among the AI’s generated variations. Human judgment remains in the driver’s seat throughout the process. The result is a workflow that frees writers to invest their creative energy where it matters most while AI absorbs the volume work downstream.

One other insight emerged consistently across the four studies: If what you are after is the greatest diversity of ideas possible, then humans are hard to beat. But diversity alone is not enough. In the circular-economy study, for instance, a single human iteratively working with AI produced solutions that scored higher in overall quality because targeted prompting enables rapid refinement toward practical value, even though a human crowd produces more diverse and novel ideas. Leaders should choose wisely when deciding how to employ human-only groups in their organization’s workflow.

2. Diversify AI inputs. Homogenization often stems from everyone using the same AI tool in the same way. Managers can push back against this by deliberately introducing variety: Rotate prompts, experiment with role-playing instructions (such as “Argue against this idea”), run parallel AI models, or integrate novel data sources.

Research supports the use of those tactics to increase idea diversity. For instance, chain-of-thought prompting, which involves asking the model to reason step by step before generating outputs, produces substantially greater dispersion in idea sets than plain-vanilla prompts and, in some cases, approaches the variance achieved by human groups.7

Complementing this, the circular-economy study demonstrated that when humans iteratively instruct a large language model to generate solutions distinct from previous iterations, they significantly enhance novelty without sacrificing value. That field study found that this human-guided differentiation approach, which explicitly prompted the model to “tackle a different problem than the previous ones and propose a different solution” after each output, produced solutions with novelty ratings comparable to those of human crowds while maintaining superior strategic viability and overall quality.

3. Deploy multi-agent and multimodel approaches. Just as diverse human teams surface a broader range of perspectives than homogeneous ones, diversifying AI “voices” can counteract convergence. Incorporating agentic workflows, where one AI generates ideas while another critiques them or multiple specialized agents tackle different aspects of a problem, can significantly broaden the search space.

Organizations can design systems where different models or agents address the same challenge from distinct perspectives. At the research frontier, multi-agent AI systems are beginning to assist with scientific discovery itself, generating hypotheses, critiquing them, and refining them in self-improving cycles. Some early use cases are already producing experimentally validated results.8 Colgate-Palmolive offers a practical illustration of this architecture in action. Rather than routing innovation work through a single AI interface, the company uses different AI systems in conjunction with one another: One mines consumer data to surface unmet needs, a proprietary AI generates product concepts, and a third uses “digital consumer twins” to simulate consumer reactions — with humans guiding each handoff.

The key insight is architectural: Rather than channeling all creative work through a single AI interface, organizations should build workflows that create productive tension across multiple AI perspectives. This approach takes advantage of AI’s efficiency while maintaining the divergent thinking that drives breakthrough innovation.

4. Build guardrails and mindful friction to protect human comparative advantage. The temptation will be to let AI act as a copilot everywhere. But if employees outsource their core creative tasks, they risk losing the very skills that make them distinctive. We recommend introducing some friction to AI use — small design choices that prevent people from becoming passive consumers of AI output. For instance, teams could be required to submit human-generated options before consulting AI, or justify why an AI-suggested idea should be selected.

There’s more at stake than just another good idea that might benefit the organization: These practices keep people’s creative muscles active while still harnessing AI’s efficiencies. Without such guardrails, efficiency gains will quickly become commoditized, leaving little basis for competitive differentiation and depriving the workforce of its ability to drive new ideas forward in an age when everyone will have access to AI. Research backs this up. A 2025 study found that students with unrestricted AI access performed significantly worse once that access was removed — but carefully designed guardrails eliminated this penalty.9 Similarly, in a different study, consultants who blindly adopted AI recommendations underperformed compared with those who maintained critical oversight.10

Ultimately, what will define organizations’ competitive edge in the years to come is their ability to cultivate a diverse set of creative ideas through human ingenuity, complemented by an efficient, research-backed workflow that uses AI’s capabilities at the right time to achieve superior quality and feasibility.

AI can support creativity, but only if humans engage actively and early in shaping the process. The organizations that stand out will not be those that use AI the most but those that use it most intentionally, designing AI use in ways that allow human originality and machine efficiency to amplify rather than cancel each other out.

When leaders talk about generative AI tools, one promise comes up repeatedly: These tools will save time.

Fewer emails. Fewer meetings. Less administrative drag.

That expectation shapes how many organizations decide whether GenAI is “working” — and why they’re often disappointed when people’s calendars don’t suddenly open up.

But for our organization, time savings turned out to be the wrong place to look.

What we saw inside a large public higher-education institution, the Community College of Philadelphia, wasn’t less work, but work that changed form. When generative AI entered our organization’s everyday workflows in 2026, coordination didn’t vanish. It shifted away from meetings and toward writing, away from clarification and toward clearer first passes, away from back-and-forth deliberation and toward faster closure on decisions.

To understand what really changed, we looked at how three of the professional roles within one administrative unit — executive leaders, operational leaders, and student-facing professionals — worked during the same six-week period across four different years. What emerged wasn’t a story about automation replacing people. It was a story about how work gets shaped, completed, and passed along.

That distinction matters. Organizations that judge AI only by hours saved risk missing the real gains and feeling underwhelmed by AI, even when it’s quietly doing what it’s supposed to do.

Three Groups’ GenAI Gains

GenAI tools showed up in day-to-day work at our college in 2026. To understand the impact GenAI had on the three groups of professionals, we examined the same six-week window each year (February 1 through March 15), comparing work in 2026 with patterns from the previous three years.

We didn’t ask, “How fast did people work?”

We asked, “What kind of work were they producing?”

Because staffing levels and work hours remained essentially the same across all four years, any differences we observed reflected changes in how work was done, not changes in capacity.

Throughout this article, we use “baseline” to refer to the same six-week period in 2023-2025 — the three years when work was performed under comparable, but pre-AI, conditions. Here’s a breakdown of the results.

Executive Leadership: More Decisiveness

For executive leadership, generative AI usage brought a clear shift toward more decision-focused communication.

Outbound email volume increased in 2026 compared with the previous year. At the same time, the share of messages that were decision- or execution-grade rose sharply, from roughly 60% at baseline to about 80% in 2026.

This was not an anomaly or simply noise. It was evidence of more direction, clarity, and closure. Emails regarding decisions were sent once rather than negotiated repeatedly.

The productivity gain didn’t come from people writing emails more quickly. It came from finishing the thinking before hitting “send,” which reduced the need for downstream clarification and prevented issues from bouncing back up the chain.

Operational Leadership: Faster Work

Operational leadership’s pattern was different from that of executive leaders. Instead of increasing, overall email volume remained relatively stable across the four years. What changed was the quality of that communication.

The share of decision- and execution-grade messages increased substantially, from about 65% at baseline to roughly 85% in 2026. That translated to less drafting, redrafting, and reclarifying of decisions. As a result, productivity gains appeared as organizational speed rather than a simple reduction in email volume: Time was freed up for direct engagement with faculty, staff members, and students.

Given that staffing levels and work hours were unchanged, these gains reflected faster turnaround and lower effort per decision (as opposed to a shift in communication channels).

Student-Facing Professionals: Resolution Efficiency

This group had experienced a pronounced spike in email volume in 2024, but communication had declined sharply by 2026. The share of decision- and execution-grade messages increased modestly in 2026, from an already high baseline of around 80% to about 85%.

This lighter communication pattern did not signal disengagement but resolution efficiency. Clearer guidance upstream, combined with a procedural shift that routed certain interactions through a centralized portal rather than email, reduced the number of clarification cycles required to resolve student queries.

Generative AI played a supporting role by reducing the time needed to draft and redraft responses, which enabled staff members to answer student questions with fewer steps. As a result, the staff now had more time to interact face-to-face with students when the need arose.

What We Gained

Across all three roles, meetings did not go away but many escalations did. Standing meetings remained. External meetings continued. One-on-ones didn’t vanish.

But issues that once triggered quick “let’s talk this through” meetings were increasingly resolved in writing. Decisions were made and communicated with enough context to stand on their own. Clarification moved out of synchronous time and into first-pass clarity.

AI didn’t eliminate meetings. It reduced unnecessary escalations — a far more meaningful gain.

What about economic benefits? The productivity gains we observed did not come from reducing head count or extending work hours. Staffing levels remained stable. Yet more decision-grade work was completed, and more time was available for direct engagement with students.

For leaders, the implication is not immediate cost cutting thanks to GenAI tools but avoided friction. When decisions display clarity and enough context to stand on their own, faculty and staff members spend less time seeking clarification, revisiting prior guidance, or navigating uncertainty. At our college, that reclaimed time is redirected toward student support, instruction, and problem-solving rather than internal coordination.

Over time, this type of shift will have economic consequences, and not just at postsecondary institutions. Clearer coordination allows organizations to absorb more work without adding layers, roles, or meetings. It reduces the hidden costs of delay — repeated emails, follow-up meetings, and stalled actions — that quietly consume employee capacity and often feed burnout.

In this sense, the economic value of generative AI may show up less as line-item savings and more as structural resilience. In our case, this means having the ability to keep organizational focus on student success while slowing the growth of administrative overhead.

Takeaways for Leaders

Although our analysis draws on a higher-education setting, the coordination patterns observed — decision escalation, clarification cycles, and role-specific workflows — are common to many people-intensive organizations.

Three lessons stand out:

1. Don’t judge generative AI only by time saved. Look at how work changes shape.
2. Expect communication to evolve, not disappear. Clarity has organizational value.
3. Design for specific job roles, not averages. The same tool produces different gains depending on how work is organized.

Across roles, generative AI did not create a single productivity effect. It amplified what mattered most in each role: decisiveness for executives, speed for operational leaders, and resolution efficiency for student-facing professionals.

For our organization, AI did not produce empty calendars or fewer emails. We gained better first drafts, faster closure, and more time to deal with people directly. Those are valuable gains.

Avoid chasing the wrong success metrics: Consider your organizational dynamics and where workflow gains are most needed.

In 1951, philosopher Martin Heidegger told a small audience, “The most thought-provoking thing in our thought-provoking time is that we are still not thinking.” Few understood him then. Seventy-five years later, the observation has become unavoidable because AI has forced every leader to confront a question about the nature of intelligence and thinking itself. If thinking is nothing but what machines can do, only faster, we have no case: We outsource it to machines. But if thinking is something else — an embodied, attentive activity through which reality reveals itself — then leadership in the age of AI is the task of cultivating a generative capacity no machine can replicate.

Consider the doctor who treats screens instead of patients, the teacher constrained by standardized testing, or the World Cup referee whose real-time decisions are repeatedly overturned by a video assistant referee. Everywhere, situation-sensitive judgment is being replaced by what Hartmut Rosa calls execution logic: prestructured parameters that turn decision makers into mere executors.1 As spheres of discretion disappear, the creativity of human agency drains away. Beneath these surface symptoms sits the deeper question now beginning to surface in boardrooms: What is irreplaceable about us, and which intelligence will be the foundation of durable advantage once everything codifiable has been automated?

Every leader I work with — in business, government, international institutions, and nongovernmental organizations — reports the same thing. The machine is spinning faster than they can process and think. The acceleration extends far beyond AI: the inbox, the KPIs, escalating disruptions, tools meant to save time that consume more. Overwhelm has become a shared planetary experience. It is also an early warning signal that something essential is being eroded, precisely when we most need it.

This erosion has a name, a diagnosis, and a response. The name is intelligence monoculture: the assumption that AI is the only intelligence worth investing in. The diagnosis is that monocultures, sooner or later, collapse. Our response should be to create a second infrastructure, running in parallel to the agentic AI-enabled IT stack: a deep-sensing leadership infrastructure that cultivates the collective capacities to co-sense and cocreate at the level of the whole system. With it, AI becomes survivable and useful. Without it, the first infrastructure depletes the very soil it is rooted in — heading toward erosion and, eventually, collapse.

This is the blind spot. Leaders have a strong grasp of what they do (the actions they take, the strategies they execute) and how they do it (the processes, the systems, the tools). What remains hidden is the inner place from which they operate: the source of attention, intention, and creativity that no machine can replicate.

The age of AI forces us to clarify our assumptions about intelligence. Can thinking be reduced to computation and pattern recognition? Or is human thinking qualitatively different? And underlying this looms a deeper question: Who are we as human beings? Are we mere extensions of increasingly powerful algorithms — or genuine sources of awareness, intention, and agency?

Three Intelligences for the Age of AI

Intelligence is not one thing. At a minimum, three forms must be differentiated and integrated.

AI in the form of large language models (LLMs) is a pattern-prediction machine, matching and meshing existing human knowledge at a superhuman speed. Trained on existing data, AI deals extraordinarily well with dynamic complexity. It’s powerful — but structurally backward-looking, even when it appears to look forward.

Organic intelligence (OI) is the intelligence of structurally coupled living systems in ecologies of relationships. It senses multiple perspectives and orients to see with rather than just look at.2 This is where empathic listening lives. OI handles social complexity — the texture of multiple worldviews, cultures, and interests.

Source intelligence (SI) is the intelligence of the whole social field — the social soil from which all perspectives emerge. It is sourced not only from what is but from what is emerging. SI also stands for soil intelligence: the intelligence of the social mycelium running through that soil that connects what looks separate aboveground. Examples are entrepreneurs and leaders who sense and create a future that does not yet exist.

SI is grounded in what Eva Pomeroy and I have called fourth-person knowing: the source from which collective action arises.3 It handles emerging complexity: Where the solution is unknown, the problem keeps changing, and it is unclear who needs to be at the table.

The three intelligences are highly interwoven and nested, with source intelligence at the core and organic and artificial intelligences in the surrounding spheres. An intelligence monoculture — almost entirely dominated by AI — would look like an empty shell. There would still be some hardware. But the living, breathing inner core would be gone, turning the shell into a superhardened iron cage for those trapped inside.

The standard fear about AI runs one way: Machines are becoming more like humans. The real danger, though, may run in the opposite direction. We are becoming more like machines — not physically, but epistemically: We see thinking as computation, learning as data processing, creativity as recombination, decision-making as optimization, and the human self as algorithm. That epistemic conversion is what makes LLMs so seductive: They do not need to actually understand. They only need us to have already redefined understanding as what they do.

The Cave and the Sun

At the heart of the leadership challenge in the age of AI lies the question of where human attention, creativity, and agency originate. The late Bill O’Brien, a former CEO of Hanover Insurance, named it in a single sentence: The success of an intervention depends on the interior condition of the intervener.

In our work with teams across sectors, we have identified four structures of attention that organize how we listen, think, and act:4

1.0: Downloading. I listen to what I already know. Attention originates from inside the system; the interior condition is enclosed and reactive (ego-centric).

2.0: Factual listening. I lean into new facts with curiosity. Attention originates from the boundary of the system; the interior condition is transactional (object-centric).

3.0: Empathic listening. I see the world through the perspective of another. Attention originates from the field of relationships (relation-centric).

4.0: Generative listening. I listen to what is emerging from the edges, leaning into its best future potential. Attention originates from the surrounding sphere of potential; the interior condition becomes permeable to what wants to emerge (eco- or cosmo-centric).

The blind spot operates differently at each level. The arc from 1.0 to 4.0 is a shift in the structure of attention. What Plato names allegorically, leadership in the age of AI must name operationally. Prisoners chained in a cave see only shadows cast by a fire behind them; at levels 1.0 and 2.0, they take the shadows for reality. Much of today’s management lives among shadows — AI-generated projections, KPIs, dashboards, pattern matches mistaken for understanding. At Level 3.0, we turn around: We see the fire that casts the shadows. At this level, systems begin to see themselves. At 4.0, we step outside the cave into sunlight, into the realm of the source, which illuminates all things but cannot be seen by looking directly at it.

AI produces ever-more-convincing shadows. It simulates all four levels with astonishing mastery — text patterns that sound ego-centric, object-centric, empathic, even field-aware. But the simulation comes from patterns without interior condition — no witnessing awareness, no deep thinking. No one is there.

Perhaps the ultimate gift of AI is this: It holds up a mirror that forces us to see ourselves and ask, “Who are we? And who do we want to become?”

Four Levels of Collective Action and Strategic Innovation

Resilient organizations operate and innovate across four levels of collective action. Each level involves a distinct structure of attention and, in the age of AI, a distinct set of core leadership skills for the respective human-AI interface.

Level 1.0: Pattern-Executing — Automating. The first level is pattern executing and replicating: This operates with the logic of downloading, as in the levels of attention above. Agentic AI is an unprecedented driver of this level. The human-AI mode is delegation: AI or machine intelligence takes over well-defined cognitive tasks. Think about a fully automated production line. The core leadership skills here center around judgment, or how to recognize plausible but false AI outputs and results. The focus on automation can liberate human attention for higher-level work. This is where most investment flows today — and where Rosa’s execution logic operates in its purest form.

Level 2.0: Pattern-Adapting — Augmenting. The second level is pattern adapting and adjusting to the context of the environment. Human attention engages in object-centric ways — noticing disconfirming data, exceptions, and anomalies — but intention and agency remain within existing frames. The human-AI mode is navigating — Kasparov’s centaur: human strategist on top, AI as the powerful body underneath, with the human steering. This is what Nobel laureates Daron Acemoglu and Simon Johnson call “machine usefulness”: AI complements rather than replaces human labor.5 An MIT Media Lab study on cognitive debt found that LLM-assisted writers’ showed neural connectivity up to 55% lower than that of those who wrote without AI — and that sequence matters: Those who worked brain-first and then engaged AI showed significantly stronger metacognitive engagement than those who started with AI.6 At Level 2, the core leadership skills involve intention setting, sensemaking, and good judgment.

Level 3.0: Pattern-Shaping — Co-Sensing. The third level is pattern sensing and pattern shaping. Conversations shift from debate to reflective dialogue, that is, to thinking together. The move from sensing to shaping defines this level. Here, all three intelligences interact. OI tunes into the multiple perspectives at play. SI leans into emergence. AI surfaces patterns across large-scale data that no individual could perceive — and, used well, holds up a mirror in reflective dialogue that helps humans become more aware of their own assumptions and agency. The human-AI mode is shaped by a partnership with machines, revolving around orchestration and mirroring. This mode requires holding spaces for multiple intelligences to interplay, which in turn requires the core leadership skills of holding space for co-sensing, discernment, intention setting, and co-shaping to happen.

Level 4.0: Pattern-Originating — Deep Sensing and Cocreating. The fourth level is pattern-originating: deep sensing and cocreating. Here, SI moves to the core. Sensing what is shifts to sensing what emerges — the highest future potential. Reflective conversation shifts into generative dialogue: collective creativity and flow. The human-AI mode is holding the space: Origination emerges from human attention that becomes permeable to the field (eco- or cosmo-centric). AI moves from the center to the periphery, if it appears at all (a transcript, a reflective surface to return to later), and is not part of the originating act. The core leadership skills at this level center around holding space for deep sensing, moral discernment,7 shared intention, and cocreating.

In other words, the core leadership skills of the lower levels are included and recontextualized in the higher levels of collective action. One of the most critical leadership capacities today is the metacapacity to balance all four of these levels appropriately. Without that rebalancing, the gravity of AI pulls everything toward Level 1.0 and 2.0 monocultures.

From Machine to Living System

Industrial-era companies were designed like machines: standardized, process-driven, hierarchical, and replaceable. AI-era organizations, as my colleague Lili Xu has observed, increasingly resemble living ecosystems: dynamically collaborative, decentralized, adaptive, and responsive in real time. The most powerful companies of the future may not be the largest but the ones that learn and sense into emerging opportunities the fastest.

As AI dramatically reduces the cost of replicating expertise, what was once the source of competitive advantage — proprietary methods, scale, 10 years of training — collapses. What is truly irreplaceable about a company in the age of AI? Not algorithms; those are commodifying. The real source is the capacity to build organizations where technological intelligence and human field intelligence can evolve together.

The hidden infrastructure for this resilience, says Xu, is the people who sense tensions before they become crises, who hold trust across stakeholder groups, who perceive what customers cannot articulate. These forms of intelligence rarely appear in KPIs, yet they are often the source of an organization’s deepest competitive durability. This constitutes the paradox of the AI era: The more that intelligence becomes abundant, the more the relational and field-based intelligence becomes scarce — and therefore valuable.

What organizations now need to do is invest in deep-sensing infrastructure with the same seriousness they invest in AI. This is the other half of the infrastructure that is missing today in most organizations and economies.

For leadership teams ready to assess where they currently stand, the first question to ask is “How much leadership attention is currently going to the first and second levels of action, and how much to the third and fourth?”

Four mini diagnostics can help to clarify that picture:

• How much time in meetings is spent downloading and debating (levels 1.0 and 2.0) versus engaging in reflective and generative dialogue (levels 3.0 and 4.0)?
• Where is the center of gravity of how your organization currently operates: pattern-executing, pattern-adapting, pattern-shaping, or pattern-originating?
• To meet the demands of our age, which of those levels needs strengthening and more leadership attention now?
• What support structures — tools, practices, places — have you created that help your teams and organization to develop deep sensing and innovation capacities around levels 3-4?

Beyond the Blind Spot

Inside the cave, we take shadows for reality. AI-generated projections are mistaken for understanding. What is missing is the Level 3.0 capacity to turn around and see the structure that generates the projections — and the Level 4.0 capacity to step outside into sunlight, to originate new patterns from source.

AI produces ever-more-convincing shadows — depth, empathy, even wisdom — simulated from patterns without an interior condition: without the awareness that notices its own awareness. The current cave that we are operating in is our blind spot. Turning around and stepping outside requires what no AI can do for us: the cultivation of an interior condition from which we can see more deeply, more clearly, and more collectively.

Max Weber warned of modernity’s iron cage a century ago. Today, the cage has a new name: the 1.0-2.0 machine, supercharged by a trillion-dollar industry, the logic of inevitability, and the daily downloading that floods our calendars and shapes our attention. Each of us faces a choice: Get absorbed into the machine, or turn around and step outside. Choose what story of the future you want to be part of — and give AI the role it deserves: tool, partner, mirror or master. That move, if performed collectively, requires a new minimal enabling infrastructure: deep-sensing spaces that enable organizations to upgrade their operating systems and their capacities to levels 3.0 and 4.0.

Every day, leaders have two critical allocations to make: the allocation of attention, and the allocation of budget. What percentage of each is going into automation? Into navigation? Into orchestration? Into deep sensing and pattern origination? If your ratio is vastly out of whack, you already know what the next move should be.

The cave is comfortable. The shadows are mesmerizing. The logic of inevitability whispers that there is no alternative.

There is.

Leaders at literally every Fortune 500 company will tell you that they are governing their AI — every single one of them. Now ask those same leaders who’s responsible for shutting down an AI model that’s causing harm. Most people can’t answer that question.

That silence is the most important story in enterprise technology right now, and it’s rarely, if ever, addressed.

During the past several years, a governance industry has quietly grown up around artificial intelligence. Companies have built registries to catalog their AI models. They have implemented classification systems to label their data. They have stood up dashboards to monitor model behavior, and risk councils to review new deployments. They have written policies, hired compliance officers, and presented slides to their boards. The infrastructure of governance has proliferated.

What’s missing is the governor.

Who Can Shut Down an AI Model

I run AI and data governance at Adobe. My job is to build exactly the kind of program that every company now claims to have. And what I have learned from the inside, while getting my hands dirty, is that the hard part is never the technology. The hard part is a question that sounds almost insultingly simple: When an AI model does something it shouldn’t, who has the authority to stop it?

The question is not who gets notified. It’s not who writes the incident report. It’s who has the authority, the organizational standing, and, frankly, the job security, to walk into a meeting and say, “We are shutting this down.”

In most companies, that person either doesn’t exist or is a paper tiger set apart from development teams organizationally. Roles like chief AI ethics officers and groups like data governance councils and responsible AI teams are completely necessary but merely advisory. They can flag. They can recommend. They can escalate. What they generally cannot do is hit the stop button. The actual decision authority sits somewhere else. Inside most Fortune 500 companies, it generally sits with someone whose primary job is shipping products and hitting revenue targets and who has every incentive to treat the governance flag as an afterthought rather than a mandate.

This is not a criticism of individuals; it is a description of a structural problem that the governance industry has largely chosen to ignore because the governance industry is selling tools, not accountability.

These tools are genuinely useful and vital to regulatory compliance and effective governance. A model registry tells you what AI systems exist and what they’re doing. A risk classification framework tells you which ones deserve the most scrutiny. A data lineage system tells you where the inputs came from and whether they were clean. All of this is real and important infrastructure. But it is infrastructure for visibility, not infrastructure for action.

You can have perfect visibility into a problem and no mechanism for solving it. You have to see something to do something, but you also have to pick up the hose to put out a fire.

Think of it this way: A fire alarm is not a fire department. You can wire every room in your house, dutifully change the battery of every smoke detector annually, and route every alert to a beautiful dashboard — but still have your house burn down because nobody picked up the hose. Yes, it’s critical to know that the house is burning — otherwise, you wouldn’t know that you need a hose. But you need somebody to tell the fire department to take action.

At Adobe, we addressed this by creating a federated governance model with named owners for every AI system and a centralized steering committee, with escalation authority, reporting into the trust and security organization, not the product team. That’s the key design choice we made: to give governance a reporting line independent of the teams shipping AI products. This is essential so that the person who can say “no” to an AI decision doesn’t report to the person who benefits from saying “yes” to shipping products.

Why Urgency Is Required

The stakes here are rising fast. The European Union’s AI Act is now in force, and it does not ask companies to demonstrate that they have dashboards. It asks them to demonstrate that they have meaningful governance. It requires documented decision-making, clear lines of accountability, and the ability to show, after the fact, who made a consequential choice about an AI system and why. When regulators come asking those questions, a policy document and a risk registry aren’t going to be sufficient answers. Regulators want a name.

The urgent need for governance is compounded by the speed of AI deployment. Most large enterprises are now running hundreds of AI systems across their organizations, in places leaders may not even be aware of. You will find AI tools being used in customer service, hiring, content moderation, pricing, fraud detection, and anyplace well-intentioned employees are just trying to make their lives easier. Many of these systems were deployed quickly, under pressure, with governance treated as something to be sorted by future-them. The time for future-them has arrived.

The answer is not to slow down AI adoption. It is to take the organizational design question as seriously as the technical one. Every AI governance program should be able to answer three questions: Who has the authority to stop a model? Do they know it’s their job? And do they have the standing to exercise that authority when it conflicts with someone else’s road map?

If your company cannot answer those questions, you do not have a governance program. You have paperwork.

The companies that will navigate the next five years of AI regulation and public scrutiny are not necessarily the ones with the most sophisticated tooling. They are the enterprises that did the harder, less glamorous work of building a human accountability structure to sit underneath the technology.

These organizations are the ones that appointed an AI governor, gave them real authority, and made clear that the job was not to make AI deployment easier but to make it defensible. These organizations have a federated team deputized to identify, remediate, and escalate, and they know that escalation requires a destination. That means a governance function with a direct line to senior leadership, independent of the product teams shipping AI, with explicit authority to stop an AI deployment.

Every company says it governs its AI. The real question separating governance from theater is simpler than any framework. Ask yourself: Who in my organization can say “no” and have the authority to mean it?

What would you do with artificial intelligence if you were confident that it would transform your industry? What actions would you take if you felt that you were at an inflection point in that transformation? Would you try to be an early proponent of AI-first in your industry, or a fast follower?

Those are some of the questions faced by the leaders of Franklin Templeton — officially Franklin Resources Inc. — a large investment and asset management firm with about $1.7 trillion in assets under management that was founded in 1947.

Over its 79-year history, Franklin Templeton has grown through strategic acquisitions that have enhanced its capabilities and global reach and expanded its competencies across asset classes, geographies, and investment philosophies.

Today, however, AI is an important driver of future growth and profitability. Consultants, academics, and industry associations agree that the technology is already powering research, compliance, and client relationships in the investment field and that it will transform them further in the future.

Jenny Johnson, Franklin Templeton’s CEO and a third-generation leader of the firm, has long combined investment leadership with deep technology fluency, having managed technology organizations earlier in her career. Before AI became a board-level mandate, she had already been focusing on AI for years. She personally experiments with generative AI, building AI agents and using techniques like vibe coding (using generative AI prompts to write code) to create computer programs.

But even Johnson has been amazed by the rapid advancement of AI in the industry. “This is faster than even I thought it was coming,” she said in a November 2025 video interview. “Every big financial institution spends a lot of money on reconciliation between systems and reconciling data. AI can help with that.” She noted that AI could also review company research reports and sell-side reports, analyzing, for instance, how tariffs would affect U.S. pharmaceutical companies versus those in Europe. “I don’t think everyone will have the same models,” she said. “Training the model is all going to be about your own data.” The future, she said, will be having the company’s entire talent force using AI as a tool.

AI Capabilities Today at Franklin Templeton

Franklin Templeton is moving rapidly toward that future, with a huge variety of internal AI capabilities and transformative platforms at both production and pilot status. The company has product teams that work with business units such as distribution, operations, and investments. Each product team operates under an AI-first model that combines product management, engineering, and data science into one unit. There is a common AI platform team and a research team. There is also an adoption and solutions team that drives employee implementation of AI and helps align business benefits with the products.

Deep Ratna Srivastav, the company’s chief AI officer, is responsible for AI product management, engineering, research, and adoption. He was involved in the conceptualization and launch of Franklin Templeton’s Goals Optimization Engine, one of the company’s early portfolio selection and optimization offerings. He told us that the engine integrates with global fintech ecosystems — including those with recordkeepers, managed account providers, custodians, and digital wealth platforms — to deliver personalized investment strategies aligned to investors’ financial objectives. It currently generates recommendations for over 40,000 investors, primarily focusing on retirement goals. The application has been embraced by several of the company’s strategic partners and is part of the firm’s forward-looking AI road map. The next phase, Srivastav said, will apply reinforcement learning to advance portfolio optimization.

Franklin Templeton offers its sales and distribution team its Intelligence Hub, which brings together AI and digital capabilities designed to enhance insights, facilitate territory management, and strengthen client engagement in meetings with financial advisers. The hub centralizes previously fragmented data sources, research, and over 15 workflow tools into a single interface, reducing manual search time and accelerating access to important content for sales meetings. AI-powered workflows automate list generation, meeting preparation, and dynamic prioritization. A Franklin Templeton salesperson can get a recommendation on which independent financial advisers to highlight, what to focus on in a client conversation, how best to get visibility with the adviser, and the most appropriate clients to meet with based on geographical proximity.

Following a yearlong pilot, Intelligence Hub was made broadly available to the company’s sales professionals in early 2026. Srivastav said it has delivered measurable efficiency improvements, including reduced daily preparation time before client meetings. It has also led to a significant increase in value-added client interactions.

The company has also applied AI to end-to-end processes in the middle and back offices of the organization. There are AI-enabled platforms in production for automated reconciliation of trades and for creating scalable communications with custodians, counterparties, and core trade operations.

Investment analysis is also increasingly supported by AI. The goal is not to automate investment advice but to support it with better information, faster iteration, and insights that humans alone couldn’t arrive at. “Copilot, not autopilot” is the overall objective.

To that end, a system called MosaiQ combines portfolio construction, manager research, and analysis into a single platform. An AI assistant named Pixel guides users through MosaiQ using natural language to explain complex investing concepts and, increasingly, to perform end-to-end tasks on users’ behalf. There is a new portfolio manager “copilot” assistant in place that can provide early warnings of market shocks, identify behavioral biases in training, and provide insights on portfolio creation. Franklin Templeton has also built an agentic investment analyst called Gromit that can independently analyze nuanced topics (for example, the impact of higher oil prices on U.S. labor trends), fact-check humans, and offer contrarian viewpoints by analyzing a breadth of proprietary and third-party data sources. Those systems are primarily powered by generative AI.

Looking Forward

To position the company for evolving client demands, Franklin Templeton’s $103 billion multi-asset group, Franklin Templeton Investment Solutions, tasked Max Gokhman, formerly its deputy chief investment officer, to lead the new AI & Digital Asset Solutions team. It will focus on three areas: further developing AI-driven investment capabilities, launching strategies incorporating digital assets and tokenized products, and advising clients on the effective use of these technologies in their own portfolios and organizations. Gokhman’s experience as an AI company founder, digital asset investor, institutional asset allocator, multi-asset portfolio manager, and chief investment officer made him uniquely suited to lead this effort.

“I’ve seen our industry change multiple times over my career, but never at a pace this rapid,” Gokhman said. “Tenacious focus and a willingness to pivot are requisite for any asset manager that wants to be relevant five years from now.”

Chief AI officer Srivastav and his colleagues are working across many other end-to-end processes. One involves voice intelligence for the U.S. retail business to transform customer engagement. “Portfolio commentary” AI, which will deliver timely insights to strengthen the client experience, is in the planning stage. Utilizing the multi-agent orchestration portfolio management copilot for the investment team is another step in the end-to-end redesign of the investment process. Marketing is streamlining its content creation process, enabling it to produce more personalized, timely, and high-quality content. Other corporate functions — including legal, compliance, HR, and finance — will be similarly reengineered with AI.

Neither Srivastav nor CEO Johnson is terribly concerned about whether Franklin Templeton’s employees will go along with the AI transformation. While the opportunities for AI education have been only somewhat popular, there has nonetheless been rapid adoption of virtually every AI tool made available to employees, Srivastav said. In many cases, these tools have been visible to clients and partners, which is helpful in persuading employees to use them. Srivastav noted that noncompliance with the company’s extensive AI governance policy and procedures has not been a concern thus far.

The leadership team of Franklin Templeton isn’t sure whether its AI capabilities will result in a “big bank” transformation or whether they’ll power a slower evolution toward increased efficiency and effectiveness. They do know, however, that they want to be ready in advance of customer and market demand and that they need to be among the industry’s leaders.

“Professionalism” encompasses the broad set of shared beliefs and expectations about how people within an industry or workplace should interact with one another: Think communication style, punctuality, or meeting etiquette. But opinions differ: Cameras on? Cameras off? Do meetings start precisely on the hour? Is arriving a few minutes late acceptable or offensive?

Our conversations about professionalism tend to proceed like a garden that has been allowed to grow without controlling for weeds or pests and is then subject to endless debate over whether the result is “good” or “bad.” But that has never been the right conversation, because context matters: Are your organization’s professional norms good or bad for your particular workplace?

While some norms are common to many workplaces — such as following through on commitments, treating colleagues with respect, and communicating appropriately — professionalism has no single definition. It varies across regions, cultures, sectors, and industries. But as a set of norms for differentiating wanted (“professional”) from unwanted (“unprofessional”) behaviors, professionalism is inherently about excluding some for the benefit of the whole. When defined well and fairly, professional standards can effectively guard against harmful behavior while creating a shared sense of identity among people from a range of backgrounds, compounding their individual efforts into collective impact. But, defined poorly, professionalism can divide and distract teams, systematize active discrimination, and discount — or even incentivize — detrimental behavior.

As an organizational consultant, a leadership adviser, and an analyst of workplace systems, I’ve learned that the key to establishing a professionalism that works is to actively define norms and standards for your particular organization. Far too many leaders ignore their own agency to shape what professionalism means, defaulting to “how we’ve always done it” rather than questioning which norms would, in fact, serve their people best. As a result, workplace professionalism is often a mixed bag: norms that signal competence and skills alongside outdated norms that can unintentionally disadvantage some team members. For example, norms that discourage discussion of caretaking at work can exclude caretakers and parents; expectations of “normal” appearance and body language can hinder neurodivergent or LGBTQ+ people’s self-expression; and dress codes defining “acceptable” hairstyles can stigmatize people with natural, Afro-textured hair.

Contextually Defined Norms

Every leader has the responsibility to create a version of professionalism designed for their unique workplace context. By incentivizing helpful behaviors that bring the best out of every person and disincentivizing harmful behaviors that impede performance, leaders can design a bespoke code of professionalism that serves people rather than functioning as an obstacle. Here’s how to lead a collaborative process of rethinking their workplace’s approach to professionalism, regardless of geographic region, sector, or industry.

1. Define success for your unique context. Take a step back to see the bigger picture. Ask your workers and key partners to share with you what they believe success looks like for your workplace. More products sold? Satisfied customers? Highly engaged workers? Trusting relationships with key community leaders? A succession plan for solid leadership over the next decade? Defining the outcomes that matter most to your organization grounds everything you do in a “why” that goes deeper than “because a leader said so.”

2. Identify deal-breaker behaviors. Imagine an employee who is highly effective at delivering results — but the way they do it is egregious enough that it compromises their own, or possibly their entire team’s, success.

Clear deal-breakers are physical violence, harassment or intimidation, verbal abuse, or discrimination — even on the part of your top performer. Defining more subtle offenses is trickier. What if their workstation is messy? Not ideal, but perhaps excusable. What if their lack of personal hygiene causes their colleagues to avoid them? More troubling. What if they cause important clients to feel disrespected or belittled after meetings? That might be a deal-breaker.

But deal-breaker behaviors aren’t universal and may vary across cultures or industries. The practice of identifying your organization’s particular deal-breakers is powerful precisely because it can reveal cultural norms or shared beliefs so deeply held that they’re practically invisible. Discuss this as a group to identify where your key partners might agree or disagree about what behaviors constitute deal-breakers.

3. Identify the minimal expectations required for success. This is the most uncomfortable step. If professionalism is up to us to define, we might want to define it aspirationally, as the highest expectations we can set to be the best version of ourselves. Always saying please and thank-you, always following every cultural norm to the letter, embodying perfection in all workplace interactions — that’s the ideal. But no person is perfect in any setting, to say nothing of the workplace. As a pragmatic tool, professionalism is best used to define the minimum standards of behavior that we expect from our colleagues, one step above our deal-breakers.

For example, it may not be feasible to expect our colleagues to wear a uniform, but we might define success in our workplace as having a strong sense of shared group identity and attention to detail. Those criteria may be reflected in a dress code that sets the expectations that clothing will not have visible dirt or stains but will include an accessory with the company logo.

Ideally, everyone in the workplace would be gracious and warm in every interaction, but human nature makes that infeasible. However, we can define success in our workplace as requiring effective communication and good teamwork. A respectful conduct policy might set the expectation that the way we communicate will make our colleagues feel safe and respected, and that if we miss the mark, we will swiftly make amends.

Similarly, it may not be feasible to expect our colleagues to always have their video on during virtual calls. But we might define success during important discussions as requiring deep human connection — and so our leadership team might set the expectation that webcams will be on during retreats, culture-building events, and teamwide discussions.

4. Understand the gap between expectations and reality. Ask your key partners what behaviors are really rewarded or punished in practice. You may find that aspirational norms have unintended consequences. Leaders may, for example, officially encourage workers to respond to emails within 24 hours — but in practice, managers may penalize workers who don’t respond quickly, even outside of traditional working hours. Leaders may communicate that deliverables and results matter more than busywork — but in practice, they may still extend promotions to workers who seem to always be working rather than to their more efficient colleagues, simply because the busier workers seem “more committed.”

Each of these gaps has a real cost, not just to people but in terms of your ability to align your actions with how you defined success in Step 1. If these gaps represent behavioral shortcomings of your starting point of “passive professionalism,” closing them will help you establish a far more functional and beneficial definition of professionalism, tailor-made for your context and directly linked to your organization’s success.

5. Incentivize what you want, and discourage what you don’t. Professional norms are not rigid policy but a means to an end. Your particular definition of professionalism can help ensure that everyone in your workplace is rowing in the same direction, is protected from abusive and harmful behaviors, and can expect the same standard of mutual respect throughout the workplace.

If old norms are no longer contributing to success, or new norms are needed to reach success — or both — it’s not enough to simply declare a policy change in an email or during a team meeting. Leadership has to align their behaviors — particularly their informal rewards and rebukes — with the professional norms they’ve defined. To support a norm of punctuality, for example, managers can praise and acknowledge those who best embody that norm while confronting any deal-breaking behavior. (For example, an employee who routinely joins meetings halfway through should be addressed directly to correct the behavior.)

Be on the lookout for any existing behaviors that contradict the norms you’re trying to build. For example, the new norm of punctuality might clash with an unspoken norm that seniority grants flexibility, with certain employees held to a far looser standard than others. To truly ensure that timeliness becomes prioritized across the workplace, you may need to clarify that senior leaders must now show up on time as well, with no exceptions, even if they have been excused for not doing so in the past. Focusing on changing one high-impact behavior or practice at a time, and clarifying what is and is not expected, can make this shift feel more tangible.

Professionalism will always be a potential source of debate as times change and work evolves. Critiques of professionalism — that it may not meaningfully align with success, that it may be biased in its application, or that it may result in harm — reflect the real possibility that the norms you have today may not be the norms that your organization needs. Especially during contentious times, be open to revisiting what you consider professional behavior and asking yourself whether your norms are most effectively serving their purpose: empowering your people. When in doubt, return to these steps to design a strategically aligned set of professional norms that enables everyone to bring their best.

After several years of AI experiments and pilot initiatives, a crucial question remains open for most companies: How much of a return — and what kinds of returns — are we getting from all of this AI investment? To many executives, AI ROI still often feels more like art than science: elusive, imprecise, and industry-dependent.

Surveys and benchmarks paint a confusing picture about current returns. Much of the guidance also remains focused on measuring inputs — encouraging organizations to invest, experiment, and build capabilities (“You should invest in …”) — rather than on outputs and how to assess impact (“Here’s how to measure results”). Today, few companies apply the same financial discipline to artificial intelligence as they would to a new factory or piece of machinery.

Our interviews with more than 30 CEOs and senior leaders across various industries confirm that measuring AI ROI is anything but standard practice: Two companies making nearly identical investments may define success in entirely different ways. Yet companies that fail to identify an explicit approach to AI ROI — or that simply roll out generic AI tools and hope for productivity gains — rarely realize credible, lasting returns.

ROI measurement differs by the type of AI technology being used. Analytical AI projects, which are typically based on established machine learning techniques like prediction and optimization, often produce more directly attributable financial returns but tend to be applied to targeted, well-defined use cases. Generative AI, in contrast, is broadly applicable, given its ability to perform a range of knowledge work tasks previously done by humans. A GenAI tool often creates improvements in speed, quality, or volume of work, requiring deliberate translation into financial impact. And some companies combine both analytical and generative AI solutions in a customized manner.

AI ROI also depends heavily on industry context. In the consumer goods sector, companies streamline their supply chains by using analytical AI, enhancing demand responsiveness. A B2B marketing agency using generative AI may focus instead on creative throughput and ideation, proposal win rates, or lead conversions — a different definition of “return.”

Three Pathways to Tangible AI ROI

Based on our interviews with executives, we identified three practical approaches to measure and manage AI ROI. These approaches reflect a range of AI maturity levels among companies, and varying strategic intents.

By comparing your organization’s current approach against this framework, you can identify where you are and what it will take to move forward. The overarching goal for leaders: to ensure the translation of AI activity into verifiable business results.

1. Function-focused approach

Who it serves: Companies trying to build credible proof points before scaling.

With this approach, you select one or a small number of business functions, such as customer service, marketing, production, or HR, as the starting point for focused AI tool deployment. In each function’s case, you build or acquire tailored AI solutions and equip people with rigorous, function-specific performance metrics. This means tracking outcomes such as shorter response times, fewer errors, improved quality, or reduced unit costs. For leaders, the logic is “If we can demonstrate credible ROI here, we can justify broader deployment elsewhere.”

Function-focused AI initiatives often deliver some of the most tangible ROI, especially when paired with deliberate workflow redesign. In customer service, organizations that deploy GenAI-driven agents and decision-support tools have reduced handling times and call volumes — often automating a high percentage of routine customer requests — and translated those gains into lower service costs and improved customer satisfaction.

For instance, Unilever redesigned its early-stage recruitment process around AI-based candidate assessment, reducing HR’s reliance on external recruiters while shortening time to hire and lowering recruitment costs. In other companies, finance units have experienced similar dynamics, where AI-based forecasting, pricing, or fraud detection systems embedded into core decision workflows have improved accuracy, reduced losses, and delivered measurable cost benefits.

The function-focused approach to AI ROI is particularly effective for building organizational confidence in AI investments. The plus side: By limiting scope and maintaining clear ownership, organizations can create credible proof points that are easier to measure, explain, and defend. The negative side: Because specific needs and contextual factors shape function-specific ROI, different success stories might be difficult to compare or aggregate as AI adoption expands.

Your next move: If you’ve already done several function-specific AI initiatives, it’s time to begin laying the groundwork for the next stage: coordination. As function-level proof points accumulate, leaders can gradually move toward a shared AI ROI playbook with consistent definitions, financial logic, and data instrumentation standards. Start by standardizing metrics that can be transferred across functions and aligning financial assumptions across key use cases. As one CEO said, “We’re iterating toward a more structured model, linking AI impact to planning, budgeting, and playbook development; it’s an ongoing loop of learning.”

2. Coordinated approach

Who it serves: Companies trying to make AI ROI comparable across functions or units.

With this approach, you’re managing a growing set of AI initiatives across the organization. Concurrently with function-specific deployments, or even earlier, you’re also rolling out some general-purpose AI tools and shared AI capabilities that touch multiple teams and workflows. The defining challenge here is coordination: maintaining broad visibility into AI activity while selectively focusing on the metrics that have the most significant business and economic impact. Ideally, this approach facilitates shared learning, reuse of proven metrics and assumptions, and faster replication of successful AI use cases.

Organizations taking a coordinated approach often use shared AI platforms and capabilities to manage initiatives spanning multiple teams. At JPMorgan Chase, an internal AI platform called LLM Suite has been deployed to more than 200,000 employees across legal, research, client services, operations, and other functions. This gives people broad access to generative and analytical AI tools while requiring coordination mechanisms to ensure consistent value creation. At Amazon, the evolution of internal AI systems resembles an AI flywheel, whereby innovations — such as recommendations or robotics — that begin in isolated teams spread and are reused across the organization through shared machine-learning platforms and practices.

In both cases, value comes from coordinating how results are interpreted, compared, and scaled across the organization. At this stage, generative AI tools are often used both inside and across business functions, heightening the need for coordination. Analytical AI tools deliver results that are often easier to compare, via clearer links to operational and financial outcomes.

The logic and business motivation for coordination are straightforward: “We’ve invested in many AI initiatives, and we need a way to stay on top of them all.”

However, especially in larger organizations, coordination without clear standards can result in a patchwork of ROI methods, making it difficult to align priorities, compare outcomes, and decide what to scale.

Your next move: During this phase, it’s important to continue prioritizing and standardizing. Identify where deeper ROI instrumentation is warranted, and apply consistent financial logic across the full set of AI initiatives, regardless of whether they involve broad tools and capabilities or targeted deployments. Standardizing how results are translated into financial terms enables meaningful comparison and scaling across initiatives. As one CEO put it, emphasizing the need for a common baseline, “If an AI initiative claims to replace the work of four employees, I want to know who they are; otherwise, it’s not real savings.”

3. Enterprise portfolio approach

Who it serves: Companies that are ready to govern AI ROI at scale.

This stage represents the highest level of ROI maturity and is where you’re applying rigorous financial logic across the entire portfolio of AI initiatives. An AI initiative is treated like any other significant investment: It is governed through forums similar to those for capital projects and is evaluated with business cases, financial models, and portfolio metrics such as net present value and internal rate of return. This approach emphasizes funding projects that create measurable value as quickly as possible.

At Morgan Stanley, for example, AI initiatives are assessed through a structured evaluation framework that tests each use case against real-world criteria before deployment. This approach enables disciplined enterprise-level oversight and scaling of AI tools. In comparison, one equipment manufacturing company we studied applied strict financial discipline: Both analytical and generative AI initiatives were allowed to run for a limited trial period and were routinely terminated if they failed to demonstrate measurable value within six months. This ensured rigor but risked premature rejection of promising efforts.

A professional services firm pursued another option: It separated two kinds of AI initiatives — those that built mandatory foundations for generative AI adoption, where ROI was not enforced upfront; and targeted, tailored applications, where clear financial returns were required. Effective enterprise AI ROI management depends on deliberate and company-specific choices about timing, risk tolerance, and evaluation rigor.

At the portfolio level, both analytical and generative AI are evaluated as part of the investment mix, but often under different expectations. Analytical AI work fits naturally into traditional financial models, whereas generative AI initiatives may require staged evaluation and adapted governance. For example, milestone-based funding or phased ROI thresholds may be needed to save worthwhile initiatives from premature rejection when those projects have indirect benefits, delayed adoption, or value creation driven through learning.

The enterprise portfolio approach to AI ROI offers clear benefits. You can compare AI initiatives side by side, compare them with other technology investments, track portfolio-level value creation, and make more confident decisions. As AI initiatives begin to reshape the operating model, however, initiative-level ROI comparisons become less informative; leaders should then rely more heavily on enterprise-level performance indicators to assess systemwide impact.

Your next move: If you choose to take an enterprise portfolio approach, it’s important to decide how strict you want to be. Fully enforced ROI can kill breakthrough AI bets too early if you overlook the value of new capabilities, learnings, and spillover benefits. The goal is to balance financial discipline with strategic patience: Apply lighter ROI tracking to early-stage AI experiments and introduce more rigorous scrutiny as projects scale. Consider creating a separate unit or governance track for more radical AI initiatives. As one executive told us, “You don’t need to measure everything from day one; start with clear KPIs for each area, then layer in more rigor as solutions mature.”

Getting AI ROI Right: Three Takeaways

Many organizations will move through all three approaches to AI ROI over time. Here are three parting takeaways from the executives we interviewed:

• Prioritize high-value, scalable AI use cases. ROI becomes most visible and meaningful when AI is applied to high-volume, high-leverage work. Whether through enterprisewide deployment or targeted use cases, even small productivity gains in large-scale activities can compound into significant value.
• Lead decisively. AI ROI doesn’t materialize by accident. The benefits come only when you provide direction, follow through, and rethink how work gets done.
• Remind yourself that your company and AI technology will keep evolving. To navigate ongoing changes, avoid both overengineering and under-measuring.

As your organization accumulates AI maturity, use the three approaches to track your progress and see your ROI grow.

Large companies seeking access to new technologies — as well as the high returns promised by early investments in successful startups — have been establishing corporate venture capital (CVC) units for many years. But returns on those investments can be erratic, and new technologies can be difficult for the parent company to take advantage of. Why do many companies struggle to derive adequate benefits from their CVC efforts? We think that at the heart of the issue is a persistent confusion over objectives that ultimately makes CVCs difficult to sustain.

Dueling objectives have long been a problem: According to a 2015 survey of CVC investors, 79% aimed to support the parent company’s strategic aims, while 76% of respondents from the same sample claimed to prioritize financial returns.1 A 2021 study found that most CVCs still rely on ad hoc structures and governance processes that confuse parent companies and result in weak executive support and frequent shutdowns.2 Our research indicates that many CVCs continue to pursue both strategic and financial benefits, only to discover that these two goals are very difficult to mix in practice. There are no easy solutions to this problem, but our data and interviews have led us to some specific recommendations.

Our primary argument is that, once the parent company and the CVC unit agree on what they seek to gain from investments, that decision needs to drive everything else the CVC does: investment guidelines, team composition, the decision-making process, and the extent of its integration with its parent. Failure to align CVC objectives with parent expectations and then with organizational implementation is likely to be fatal.3

The Spectrum of Investment Models

Strategic-priority CVCs benefit the parent company by investing in startups that provide insight into and access to new technologies, products, services, and business ideas that the parent can take advantage of. Realizing these benefits requires close integration with the parent company’s business divisions. Financial-priority CVCs invest in startups primarily to generate a monetary return. Hybrid CVCs try to give equal weight to strategic benefits and financial returns. (See “CVC Investment Models.”) While financial returns are easily calculated by comparing sums invested to the current market value of a portfolio, evaluating strategic returns is much more difficult, especially when CVCs mix strategic and financial goals.

As of January 2025, our sample of 59 CVCs had adopted those investment models in relatively similar numbers. We classified 21 (36%) as financial-priority leaning, 20 (34%) as strategic-priority leaning, and 18 (30%) as hybrid. We included only CVCs that had made CB Insights’ annual top 10 list in terms of active investments between 2017 and 2024 and were still active in 2025. On average, these CVCs were 19 years old with a recent estimated fund size or investment budget of $749 million.

It’s important to note that most CVCs fall along a spectrum, not at the extremes (that is, wholly devoted to one or the other objective). We do not recommend a strategy on the extremes or squarely in the middle. These positions are difficult to sustain, either because they fail to provide any strategic value or financial returns or because they are mediocre at both. Instead, we suggest that CVCs prioritize strategic benefits or financial returns but aim to gain some benefits in the lower-priority category. In the majority of cases, it makes the most sense for CVCs to focus on strategic investments that yield some financial benefits, since this investment strategy is most likely to identify viable startups that can benefit the parent company. (See “The CVC Spectrum.”) All CVCs, theoretically, have a lower cost of capital than independent VCs, to the extent that they receive money from their parent companies and don’t have to compete for outside investors. Since most CVCs have some financial criteria, their main differentiation occurs in how high those financial bars are and to what extent CVCs access their parents for help with investment decisions.

Pros and Cons of Strategic, Financial, and Hybrid Approaches

Some practitioners argue that a CVC should emphasize strategic objectives as an investment in the parent company’s future. Les Vadasz, the founder of Intel Capital, one of the oldest, largest, and most successful CVCs, strongly holds this view, saying, “If you don’t have a strategic reason to invest, then I don’t think the CVC has a reason to be in business.” Vadasz expected his investments to help build demand for Intel’s semiconductor products and to provide some insight into future trends.4 He looked for a relatively quick impact on demand for the microprocessor business — for example, by investing in software companies whose applications ran on the Intel x86 chip architecture.

Siemens’s experience illustrates how difficult it is to maintain a strategic focus if that means passing up potentially good financial investments. Frank Andrasco, a veteran of Siemens Ventures and its successor, Next47, and now a senior investment director at Aramco Ventures, agreed that strategic benefits should be the main focus of a CVC. However, he found a pure strategic portfolio to be difficult to sustain. Siemens Venture Capital had been strategically oriented, but, because it had declined to invest in many deals that it later realized would have offered good financial returns, Siemens’s top management made its successor CVC unit, Next47, financially oriented. Andrasco moved on, frustrated with this decision. CVCs “are always going to be beaten to the best deals. … They are competing with Andreessen Horowitz and Sequoia. Why are they going to be better than those guys?” he told us.

Missing out on good financial investments is only part of the frustration for strategically oriented CVCs. Bailing out failing startups is also not sustainable, as Vadasz explained. “We invested money for strategic reasons,” he said. “Now, a little caveat here: You have to invest with financial discipline because companies that don’t succeed do not help you.” Also, strategic CVCs tend to become less strategic over time, according to Andrasco. “You can’t get in on the best deals because your strategic constraints create adverse selection,” he said. “So the only solution is to remove the strategic constraints.”

Another school of thought — the one to which Siemens pivoted — is that CVCs should focus on making money because startup investing offers the potential of extraordinary financial returns. Tim Chiang, a veteran of GE Ventures and Xerox Ventures, strongly holds this view, arguing that financial-priority CVCs can make quicker decisions than strategic-priority units because they don’t need to coordinate due diligence and priorities with a parent company.

In our sample, most CVCs owned by financial services firms based in Asia (such as Mitsubishi UFJ, SBI Securities, Daiwa, Mitsui Sumitomo Insurance, Fosun Capital, and CreditEase) were in the financial-priority category. They often treated a new venture fund as one of several investment vehicles for their clients. However, we’re seeing many other CVCs in this space as well, owned by Google, Panasonic, Baidu, Legend/Lenovo, and Next47 on the tech/industrial side, and SR One (GSK), Roche Venture Fund, Novartis Venture Fund, and Novo Ventures on the pharma/biotech side.

Some financial-priority CVCs in our sample occasionally capitalized on expertise in their parent companies for due diligence, startup mentoring, and business development, suggesting a different strategy than pure financial motivation. We call this model financial-priority/hybrid. Financial returns remain the primary goal, and there are no strategic investment criteria, but there is some help from the parent, such as making investment decisions or providing startup mentoring. (See “CVC Investment Models.”)

Many prominent global corporations try to give equal weight to strategic and financial objectives and create hybrid CVCs, but these give rise to the most difficult implementation challenges. There is no overarching goal to guide decision-making, and such initiatives can fall short on both strategic and financial expectations.

“The challenge is that hybrid CVCs are trying to do something that is ... inherently serving two masters. And they don’t know which one will try to kill them,” Chiang told us. Investment teams also struggle to combine different goals: “It’s hard to force an embedded VC group to change colors on the spot,” he said.

Even CVCs that successfully balance financial and strategic objectives may be shut down when the parent company runs into trouble or shifts direction. GE Ventures illustrates this point. The unit’s founder, Sue Siegel, told us that she felt compelled from the outset to balance financial and strategic criteria. “With no financial discipline, you don’t have anything. … It’s all about the healthy exit,” she said. The GE Ventures portfolio did well, but in 2024, General Electric’s board closed the CVC and divided the conglomerate into three separate companies.

Execution Challenges for CVCs

While each of the investment approaches described above comes with particular execution challenges, the most noteworthy that we saw in our research involved deciding how to measure and realize strategic benefits, maintain financial discipline, and recruit and compensate a top-notch investment team. We’ll review each in turn.

1. Measuring and realizing strategic benefits. Getting an accurate and consistent picture of strategic benefits afforded by their investments seems to be a huge hurdle for strategic and hybrid CVCs. Managers we interviewed used both qualitative and quantitative metrics. Intel Capital analyzed investment success based on the money, time, and effort the company put in, and any strategic benefits and financial returns achieved.5 Vadasz focused on two types of strategic benefits while also trying not to lose money. One benefit was access to startups that had technology Intel wanted to use, such as advanced chip production equipment. The other, as described earlier, was relationships that would increase demand for Intel’s core microprocessor products.

GE Ventures tracked the number and type of partnerships that a portfolio company had with a GE business unit, such as for distribution or commercial product development. It recorded how much money GE Ventures put into the investments and how many employees were involved in supporting partnerships. GE Ventures and GE executives reviewed the portfolio at quarterly meetings.

Based on his experience at Siemens, Andrasco developed a model at Aramco Ventures to estimate what potential value a startup investment might create for the parent company. This model also gave the CVC a basis to compare actual strategic returns — losses avoided or revenues and profits gained.

Realizing strategic benefits requires that a CVC be tightly connected to the parent company. Intel Capital did this through a matrix structure when Vadasz managed the CVC. At that time, 15 to 20 people (of about 100 total employees) were attached to one of Intel’s functional and geographic divisions but worked primarily for him. These employees attended Intel Capital staff meetings, helped with due diligence, and worked closely with portfolio companies to develop their businesses. They were assigned to work with the CVC unit for a minimum of two years and often did so for longer.

Andrasco also relies on a matrix at Aramco Ventures, with about 15 of the 40 CVC employees based in the Saudi Aramco home office doing business development and recruitment for startups. Andrasco considers this structure to be “lightweight strategic,” which he defined as being open to the “possibility of the company and the startup working together … although it may not actually happen.”

As a hybrid, GE Ventures operated more like an independent VC, but with its parent company represented on the investment committee. Siegel invited GE executives to join the committee when the CVC was considering a startup in their business area. The GE executive got one vote but did not have veto power. The three GE business units that engaged most closely with startups assigned their employees to work with them while paying their salaries. In other GE business units, the CTO or chief strategy officer sat on the investment committee for a particular review. If the investment went forward, that executive became responsible for assigning people to serve as “shepherds” and develop a partnership between the GE business unit and the startup.

Our interviews suggest that for strategic-priority CVCs, a realistic target for close relationships or partnerships with the parent company might be one-fourth to one-third of the portfolio investments. But building and maintaining these relationships requires both the CVC and the parent company to make serious commitments in terms of people and time. Acquisitions were another way to realize strategic benefits, but the CVC managers we interviewed saw M&A as a separate corporate or divisional activity.

2. Maintaining financial discipline. Financial criteria are straightforward to implement. The CVC needs to pay attention to cash burn rates and possibly set a threshold floor for “exit value” — the minimal level of desired return should the startup be sold or go public. Determining the exit value requires estimating what comparable startups have sold for or noting what their IPO values have been, or who potential acquirers might be. Establishing value requires the investment team to estimate how far from a commercial product or service a startup actually is, what the competition looks like, and who the likely customers and acquirers might be.

Financial discipline also means spreading out your bets. For example, GE Ventures adopted what Siegel called a layered investment strategy. In the first layer, early-stage investments (Series A and some seed funding) were limited to 20% of the portfolio, given that they might take 10 to 15 years to pay off, while 80% of investments were later-stage — more likely to have an earlier payoff but less likely to have a supersized return. The second-layer investments were in strategic domains, such as health care, advanced manufacturing, or energy startups. The third layer of the strategy was to target syndicate members that might become investment partners. GE Ventures wanted to invest with the top 25% of VCs, such as Sequoia and Kleiner Perkins, based on their returns over the past 15 to 20 years.

Another aspect of financial discipline is to understand what leads to a healthy portfolio. Andrasco looks for a 12% annual appreciation in the value of Aramco Venture’s investments. Similar to Vadasz and Siegel, he has established a modest financial floor because of his experience that “CVCs that lose money don’t stay in business.” Andrasco also insists that CVCs should not negotiate special deals for their portfolio companies and create situations where the parent is the startup’s least-profitable customer.

3. Recruiting and compensating the investment team. These challenges are intertwined, because choices on how to compensate the investment team affect recruitment. We found that CVCs generally struggled to compete with independent VCs on this front. Independent VCs raise outside funds and charge a management fee (usually 2%). They compensate partners with a share of any equity gains (usually 20%), called carry or carried interest. In the U.S., tax authorities treat this type of income as long-term capital gains and impose taxes at a lower rate than for ordinary income. As a result, carry often leads to huge paydays. In contrast, most CVCs compensate managers and teams at a level similar to that for new business development. One alternative is for a CVC to offer large bonuses, sometimes called “phantom” or “shadow” carry, that are indirectly tied to investment returns. Another option is to create a separate CVC fund and compensate with carry, like an independent VC.

Next47 and Siemens Ventures, as well as Intel Capital and GE Ventures, did not compensate with carry, because senior management and board directors would not permit it. In contrast, Aramco Ventures gives out bonuses that incorporate financial returns based on phantom carry and estimates of strategic value achieved.

Intel Capital looked for people from Intel business units who were interested in a temporary assignment in business development compensated via bonuses. GE Ventures looked for talented early-career VCs who had not yet made partner in independent firms. Siegel offered the equivalent of a general partnership, heavy on cash and with long-term GE stock options.

The Bottom Line

We started this research believing that most CVCs should prioritize strategic returns because parent companies have a responsibility to invest in the future and startups can help them do that. We still think that strategic investments are the most valuable bets, especially since CVC financial returns are likely to be small for multi-billion-dollar parent companies. Nonetheless, if a parent company believes that it can make more money from venture capital than from other investments, and it wants to directly influence those investments, then a financial-priority CVC makes sense. In that case, financial CVCs should at least try to take advantage of their parent companies’ domain expertise, because this is their main advantage over independent VCs.

CVCs also need to realize that the objectives and situations of their parents will change over time, which in turn, will impact their missions and evaluations. During our research, for example, Time Warner, General Electric, and Xerox all closed their CVC units, even though we’d been told that the portfolios were performing well. Several companies (including NTT, Samsung, and Siemens) also launched multiple CVCs and funds to achieve different objectives. In early 2023, Microsoft’s M12 venture fund, which started out prioritizing financial returns, announced it was adapting its investment approach to incorporate more strategic considerations.6

There will no doubt always be some tension and change in priorities for CVCs that don’t have clear objectives and performance metrics. Perhaps the biggest challenge for CVCs is to build close relationships with their parent companies, for either strategic or financial investments, while still maintaining enough independence to avoid potentially bad investments.