If one were making a list of the companies that have most exploited the potential of data, analytics, and artificial intelligence in the 21st century, Capital One would probably be at or near the top. Spun out of the credit card business of Signet Bank in 1994, Capital One was founded on an “information-based strategy.” It was a pioneer in the use of data and analytics (and one of Tom’s poster children for his 2006 article about competing on analytics) for its targeted marketing, marketing experiments, variable credit card interest rates, and analytical risk management. As of September 2026, it is the eighth-largest bank in the U.S. by assets and focuses not only on credit cards but also consumer and commercial banking.

It’s not surprising that Capital One is very active in AI as well. It is ranked second among global banks in AI capabilities, in Evident AI Index of 2025, following only the much larger JPMorgan Chase. Both of us (Tom here and Randy here) have written about AI at Capital One in the past, but, given how quickly AI is changing, we thought its activities were worth updating for our last AI in Action column for MIT Sloan Management Review.

To learn about the bank’s current focus, we spoke with Prem Natarajan, executive vice president, chief enterprise AI officer, and chief scientist; and Amy Lenander, executive vice president and chief data officer. Natarajan has a strong technical background, having worked in a prominent AI role at Amazon and as a professor and senior vice dean of the School of Engineering at the University of Southern California. Lenander has a 23-year history in different business roles at Capital One, including heading the bank’s U.K. business.

Tom was in the midst of doing some research on the economic value of AI when he interviewed Natarajan, so that was one of the first issues he explored. Natarajan, a long-term AI enthusiast and expert, had an interesting take on the issue: “You can’t achieve transformation if you’re focused on quarterly ROI.” He believes that with a powerful, fast-changing technology like AI, it’s both too early and short-sighted to measure short-term ROI. Capital One’s leaders believe that AI could reshape the banking and finance industry, and the company is dedicated to thriving through the changes to come. “The ability to manage and get access to financial services is crucial to having a satisfying life, and AI will be a massive enabler of that,” Natarajan said. “The ultimate ROI is a happy customer.”

Core AI Use Cases at Capital One

Natarajan mentioned several areas in which AI is already having a substantial impact at Capital One, including software development, fraud reduction, and customer service. Providing high-quality customer service has long been a focus of the bank, and AI is being used to further improve it. Generative AI is being used to surface and deliver the right answers to human agents faster. That provides stress relief for the human agent, and the caller experiences a faster transaction with more accurate answers to their questions.

Another AI use case in customer service is call summarization, which employs agentic AI to recall the most important elements of conversations with customers and use reasoning capabilities to interpret complex customer issues, particularly on fraud-related calls. Agentic AI allows a complex summarization task to be broken into subtasks for each agent to perform. Natarajan said that only when the bank added a reasoning agent to the summarization did the performance of the process become satisfactory. Several hundred agents are now using call summarization and have realized substantial productivity gains, with the summaries added to the customer record. The early adopters have seen a “massive improvement in customer experience,” Natarajan said.

Any well-managed bank also needs to excel at fraud prevention and risk management, and Capital One is using AI in multiple ways to improve its processes. One of them is in auditing risk and compliance; the bank has built an intelligent assistant for risk associates, using open-weight models, advanced analytics, and analytical AI. Risk assessments can be done more quickly with better quality control, script review assistance, root cause analysis, and data-driven assurance at scale. Use of the intelligent assistant in audit risk and compliance has reduced quality errors by 85%, Natarajan said. It has also reduced time spent on manual audit tasks, saving an estimated 5,000 hours of audit time across 30 auditors in only a few months. One audit task, for example, has been reduced from five hours to three minutes to perform.

Natarajan emphasized that the bulk of Capital One’s AI initiatives are oriented to enterprise and line-of-business capabilities rather than individual productivity. “Whatever we introduce,” he noted, “we put our backs into it. It’s not ‘Here’s a tool — go use it.’” The bank invests heavily in training for any new use case and in making sure people have the capabilities needed to use new tools competently. As Natarajan, who worked closely with the intelligence community in a previous job with Raytheon BB&N, put it, “Every revolution in technology has only happened when tradecraft has evolved alongside it.”

The Data That Makes It Possible

Natarajan and Lenander both asserted that data is critical to competitive advantage with AI. Natarajan commented that “proprietary data brings competitive advantage with AI,” and Lenander noted that “building a unified data ecosystem creates a competitive moat.”

Lenander views the current focus on AI as an extension of previous capabilities. “We have a long history of automated decisioning for credit, marketing, and other important domains,” she said. “Early on, the method was regression, then machine learning, then more complex machine learning, and now generative AI. The data we used were initially structured, and now it’s a mix of structured and unstructured.”

For example, she said, the data used to train customer service assistant models is from call transcripts, among other sources. Those transcripts are also analyzed by generative AI to determine what themes the bank is seeing in customer interactions, such as how they’re responding to the rollout of new products. With the combination of that data and AI, the bank can see trends in near real time. The bank also analyzes clickstream data from the public-facing website, and from the mobile app and website customers use to make transactions. With that data, analysts can learn how customers spend their time and attention and can optimize the company’s marketing funnel, sharing the most relevant information with customers to resolve their banking issues.

Lenander pointed out that even though AI can take very unstructured data and make use of it, its outputs are better when the data is organized. Just as Capital One organized its data to train machine learning models, it is now investing resources in curating its unstructured data. One key tool that has been helpful is knowledge graphs, which characterize and map the relationships between data entities such as people, concepts, and objects, to identify, for example, which customers use which banking products. Curating more data underneath the knowledge graph makes it more valuable.

The bank gives responsibility for the curation of unstructured data to the people who are closest to it. In fact, it uses the same framework with unstructured data as structured data by managing it as “data products” that are packaged and ready for data consumers to use, in whatever format those data consumers need. With data products, ownership stays close to the data itself: The customer service team, for example, owns the data product for customer pathways through the call center system. This approach to data products is similar to Capital One’s general approach to data management, with federated data ownership and central data capabilities and standards. Lenander oversees a central team that ensures that all of the data products fit together, and she helps federated groups build their data products and apply a consistent, high bar for quality and usability.

Capital One has also taken a step into the market for data management software. Databolt was developed by the bank’s internal software team to tokenize data for security purposes. Databolt enables organizations to secure sensitive data across large language models and AI workflows without exposing personally identifiable information and is now marketed to external customers through Capital One Software. It is the second product for that business, following Slingshot, a tool that helps companies predict and manage cloud spending.

Our discussions with Natarajan and Lenander reassured us that this company that we’ve studied for decades is still leading the field of data, analytics, and AI-driven enterprises. It’s a trait that at this point is deeply embedded in Capital One’s DNA.

Leaders tend to talk about culture in abstract terms: values, norms, trust, belonging, and the like. But employees watch for the moments that show how culture actually manifests. During meetings, who gets interrupted? Employees notice whose sentences are cut short, redirected, or quietly overtaken — and whose ideas are allowed full expression.

Interruptions are rarely considered in culture surveys or engagement dashboards; they’re often instead dismissed as the outcome of a particular conversational style or personality clash, or the by-product of fast-paced work. Yet interruptions are one of the clearest cues employees use to assess whose voices carry authority, whose judgment is trusted, and whose presence requires constant negotiation.

In our research on workplace microaggressions that was recently published in the Journal of Business Ethics, we found that subtle acts of exclusion can erode employees’ sense of psychological ownership and shape how they respond to the organization. Across two vignette experiments and a two-wave survey, a pattern became clear: Small acts of exclusion had real behavioral consequences.

How do those dynamics play out in everyday interaction? To answer this question, we ran a focused substudy of in-depth interviews with leaders who had experienced interruption at work and observed it among others. We also analyzed meeting recordings and field observations, using an interactional lens to code turn-taking, interruption, and idea uptake.

Across both strands of the research, we found that employees didn’t talk about culture in the abstract language leaders used but in terms of moments — and that moments of interruption were among the most revealing. Culture is evident in how interaction unfolds moment by moment, especially when the stakes are high and attention is scarce. People judge organizations less by what leaders claim than by whether people’s everyday experience signals responsiveness, dignity, and care.

Interruptions as Cultural Indicators

Interruptions are not inherently problematic. In some contexts, they can signal energy, urgency, or collaboration. What matters is how leaders handle them. Who interrupts, who gets interrupted, and what happens next? When someone is interrupted, three things happen at once: Their contribution is cut short; the direction of discussion shifts; and a judgment is made, implicitly, about whose input merits completion.

Across the settings we examined, interruptions were spread unevenly. Participants described them clustering around particular people, viewpoints, and moments, especially when the speaker was still forming a point or introducing a dissenting view. One leader told us, “It wasn’t that anyone was openly dismissive. It was that my ideas rarely got to finish breathing.”

Over time, those patterns shaped participation. Many people we interviewed described adapting their speech and behavior after repeated interruptions, by shortening their contributions, adding disclaimers, or waiting for an explicit invitation to speak. It’s through these repeated micro-adjustments and moments of self-correction, not through policy, that culture becomes internalized. This adaptation effect was one of the most consistent findings in our follow-on study.

Because interruptions are common and often ambiguous, leaders tend to minimize them, attributing them to communication style, time pressure, or poor meeting etiquette. But for those on the receiving end, repeated interruption is rarely experienced as neutral. Employees quickly learn who is likely to be cut off, when, and by whom. That predictability turns interruptions into cultural signals that reveal whose authority is fragile and whose is buffered by status.

In organizations that publicly emphasized inclusion and psychological safety, such moments were especially consequential. Employees often contrasted stated values with what interruptions revealed in practice. As one participant put it, “You don’t need a policy to tell you where the power is. You hear it in who gets cut off.”

We found that interruptions often served as an early indicator employees used to decide whether broader cultural commitments, such as openness, candor, or collaboration, were genuine. In several settings, participants referred to interruption patterns as “the real culture” regardless of what leaders espoused.

The Myth Of ‘Jumping Back In’

Executives often assume that interruptions are harmless because anyone can simply reenter the conversation. In practice, that assumption ignores how power shapes participation. Speaking again after an interruption requires confidence that one’s contribution will be welcomed rather than penalized. That confidence is not equally distributed. In our data, some participants were far more likely to reclaim the floor after being interrupted than others, who were more likely to withdraw, wait, or defer entirely.

In one tech firm that prized robust debate, leaders interpreted interruptions as signs of energy, but employees experienced them differently. Interruptions clustered around dissenting views and were rarely moderated. One employee told us, “It wasn’t that disagreement wasn’t allowed. It was that disagreement had to sound a certain way, from a certain person.”

What leaders described as dynamism, employees experienced as constraint. This gap between leadership interpretation and employee experience was one of the most persistent disconnects in the study.

One reason interruptions persist is that leaders rarely experience them symmetrically. Status secures attention: More senior people are interrupted less, deferred to more, and insulated from conversational risk. As a result, meetings can feel open and participatory from the top while feeling constrained from below.

In our interviews, several executives were surprised when presented with interactional findings. Many believed that they had been modeling inclusive behavior, but culture is not transmitted through intention. When shown the pattern of interruptions, they saw the gap between intention and impact. One executive said, “I thought I was helping by keeping things moving. I didn’t see who was paying for the pace.”

We also found that the same interruption does not land the same way for everyone. Seniority, role, identity, and organizational standing all shape how these moments are experienced and interpreted. Across our interviews and observations, employees described interruptions as especially consequential when they came before a point had fully formed, or when they redirected discussion away from uncertainty, risk, or challenges. In those moments, interruption did more than cut someone off. It filtered what counted as usable knowledge in the room.

This matters not only because it is unfair but also because it is costly. Across our observational data, interruptions often truncated contributions that raised concerns about risk, ethical issues, or longer-term implications. In other words, the voices most likely to be cut off were the ones broadening the conversation in useful ways. Diverse representation alone does not address this. It may widen the bench of voices in the room, but interaction determines which voices prevail.

How to Address Interruptions

Once leaders recognize interruptions as cultural signals, their instinct is often to correct the interrupter directly. But calling out interruptions in real time can embarrass the interrupter, harden hierarchies, and produce temporary politeness rather than real behavior change.

The more effective starting point is careful observation. Leaders need to understand the architecture of conversation: who anticipates interruption, who pushes through it, and who reads it as a warning. Most leaders assume that they already know. They usually do not.

One practical step is to ask a trusted junior colleague to map a leadership-team meeting: who gets interrupted, who finishes their point, whose ideas are taken up. Leaders are often startled not by how often interruptions occur but by how predictably they cluster.

Once leaders have that information, they can take the following actions to facilitate cultural change.

Hold the floor for the speaker. Protecting the interrupted speaker is more effective than correcting the interrupter. The most effective responses reported in our study were short, neutral phrases: “Let’s hear them finish,” “Hold that thought,” or “I want to stay with this point a moment longer.” This encouragement shows employees that their incomplete ideas deserve space, that authority does not depend on polish, and that contribution is measured by insight rather than dominance.

Over time, team habits will begin to change. People who once anticipated interruption will start offering earlier-stage thinking. The meeting will become less a contest of reflexes and more a site of shared judgment.

Slow the meeting down. Many leaders try to solve interruption issues by encouraging equal participation, which often means speeding the meeting up so everyone gets a turn. But our research suggests that speed amplifies inequality. The faster the exchange, the more authority defaults to people who already feel entitled to the floor.

The counterintuitive move is to slow the room down deliberately, allowing for a brief pause after someone starts speaking, a moment of silence between agenda items, or a short delay before a response. This can feel inefficient to meeting leaders, but employees experience it as permission to think aloud, finish a sentence, or voice a concern in full before it hardens into disengagement.

We found that teams that slowed their meetings maintained efficiency and improved effectiveness. Risk surfaced earlier. Disagreement sounded less like defiance and more like contribution. The culture shifted when leaders deliberately widened the aperture of who could shape the discussion.

Redistribute facilitation, not airtime. The intention of equalizing airtime is good, but the effect can be more performative than practical. People learn to contribute visibly rather than meaningfully.

A better lever is facilitation. In our study, teams that rotated meeting facilitation showed some of the sharpest shifts in interruption patterns. Shared facilitation disrupted entrenched conversational habits. Authority no longer pooled so predictably around one leader or one type of speaker.

Making meetings perfectly fair is an impossible goal. Instead, leaders should keep discussions sufficiently flexible so that no single conversational habit dominates.

Make culture observable. Culture work often collapses under aspiration. Leaders write down values, run workshops, and publish commitments, but employees don’t experience culture as an abstract promise. They trust what they can observe. Concrete signals tell them whether the organization is genuinely responsive and whether their dignity is being protected in practice. Interruptions are observable, as are reclaimed sentences, deliberate pauses, rotating facilitators, and ideas that are allowed time and space for completion.

The most meaningful cultural shifts in our study occurred once leaders started making interaction visible in those ways and treating conversational dynamics as the practical mechanics of culture. They paid attention to patterns they had previously dismissed as interpersonal noise and changed the dynamic.

Interruptions communicate more to employees about culture than leaders may realize. The task is not to abolish interruption entirely but to address it thoughtfully. How leaders respond in those moments shows employees who is trusted to think out loud, who can explore uncertainty without penalty, and who must present a fully formed argument to be heard at all.

Leaders serious about creating a culture where people bring their best ideas can use interruptions as a practical lever for change. Employees will notice and trust the shift. Culture, for them, has been speaking all along.

HR technology is no longer peripheral to how organizations operate. A fast-growing assortment of platforms is touching the employee life cycle and capturing workforce data. At the October 2025 HR Technology Conference, more than 400 vendors showcased their wares, including new AI-enabled offerings. But despite this wealth of options, industry and consulting reports frequently attest that these proclaimed advancements in systems and decision-making remain largely unrealized; notably, SHRM found that fewer than half of both HR professionals and employees regard their company’s HR technology as effective.

Each platform promises an alluring array of benefits, the most compelling of which is the ability to improve evidence-based decisions by establishing a single source of truth: current, centrally accessible workforce data. On this foundation sits the additional promise of advanced visibility through dashboard and reporting capabilities that provide instant access to information about the workforce’s skills, capabilities, and outcomes. Yet another promise is the ability to embed and standardize best-practice frameworks across units, functions, and geographies.

The logic is straightforward: When talent decisions are supported by structured processes, shared data, and embedded analytics, leaders can shift from making fragmented judgments to intentional, data-informed decisions. These platforms serve as the lenses through which you are able to see and understand your own systems in order to pursue your own priorities.

But your interests differ from those of your vendors, who prioritize scalability and repeatability. Commercial platforms are fit for sale; your organization’s unique strategy requires systems that are fit for purpose. The criteria, workflows, and reporting abilities embedded in vendor platforms are designed to work for many clients — including your competitors.

The Accumulation Trap

The proliferation of HR technology in the market is too often mirrored within organizations. Unlike in the days of enterprise resource planning, businesses today maintain an ever-expanding portfolio of specialized tools across the talent life cycle. (Both of us have collaborated with organizations that have maintained anywhere from 80 to hundreds of different systems.) Decision makers are often surprised to find that the issue is not a lack of technology but an over-accumulation of platforms.

Few organizations’ HR technology stacks emerge from intentional architectural design. Instead, systems evolve through the steady addition of point solutions — individual tools acquired to solve localized or immediate problems and gaps. Each addition seems beneficial in isolation; it may, for example, fix a reporting limitation, add a missing hiring capability, enable employee surveys, or support a new learning initiative. But the assumption that more technology equals progress is fundamentally flawed.

Over time, these point solutions accumulate, as talent processes and decision-making authority are distributed across teams, functions, and geographies. Individual teams select different tools without having a comprehensive view of the broader technology ecosystem — undermining coherence, given that data often needs to be extracted, translated, reconciled, and transferred across systems. As AI reduces coding barriers, it is easier than ever for an organization to augment or build its own tools to replace limiting vendor solutions. Local vibe-coded point solutions are initially inexpensive but may lack stability, security, or be difficult to integrate.

Without intentional design tailored to the enterprisewide infrastructure, adding more platforms to your talent infrastructure can breed misalignment, with two significant organizational consequences: increased waste, and divergence from your strategic priorities.

How HR Tech Can Create Waste

The accumulation of HR technology systems often results in functional, financial, and decision waste for organizations. A fundamental and nearly universal business problem is data fragmentation. When data privacy and security are managed within each tool, as is often the case, competing models can create risks, such as inappropriately blocking data needed for essential functions in other tools — or inadvertently releasing data from one tool that is protected in another. Entire roles are now emerging for employees to bridge these gaps as organizations rely on “human glue” to manually extract, translate, reconcile, and transfer data across multiple systems.

Such fragmentation is a governance issue, not a technical one; when various systems collect, store, and interpret data in conflicting or incompatible formats, fragmented pipelines and algorithms that are using incomplete or inconsistent inputs can produce imprecise outputs. Lags in data availability delay decision-making as stakeholders wait for needed information, which could be out of date before it even arrives. Worse, if different systems provide different answers, competing interpretations and misaligned perspectives can proliferate. When different people have a different set of data about the same problem, it’s difficult to come to a shared understanding of what’s important or even what’s actually happening, which limits the organization’s ability to develop coherent plans.

Confusion and blurred accountability arise when there is no single party responsible for the quality of the entire tech stack. Duplication — when multiple tools perform the same or similar functions under separate contracts, renewals, and support models — can erode coherence and multiply costs. As multiple systems perform overlapping functions, different tools operate as the system of record for different processes. This competition is created and sustained by vendors and their internal champions, who attest to the value of that specific system above others and promote it as the authoritative platform.

Lastly, systems can grow outdated and diverge from current decision-making processes. Many vendor platforms prioritize static logic despite “talent” being inherently dynamic. The workflow still runs and the data is generated, but continuing to use logic or criteria that are outdated can actively sabotage progress. Imagine wearing shoes you’ve outgrown for the rest of your life just because they fit you at one time. Decision-making underpinned by “that’s the way we’ve always done things around here” is a red flag for your infrastructure. Strategy-as-experience appeals to precedent, but strategy is executed through evolving practices and ways of operating. Prioritizing organizational legacy contributes to decisions based on outdated assumptions.

Mapping Your Intent: Strategy First

When adopting vendor technology for talent management, organizations that don’t first consider whether a system supports desired outcomes can unintentionally cede command of their own strategy to vendors.1 A platform’s embedded decision logics, criteria, and processes reflect the vendor’s commercial interests, not the user’s. Vendor offerings are fit for sale, not fit for purpose, with initial and ongoing customizations requiring repeated investments to address recurring limitations. By accepting system defaults, you can lose ownership of the infrastructure that shapes your talent decisions — a primary mechanism of strategy execution that shapes your talent decisions.

‘Satisficing’ Undermines Strategy

In balancing many competing priorities — expense, speed, capacity, strategy — organizations need to make trade-offs. Satisficing is a trade-off that occurs when someone accepts “close enough” default categories, thresholds, workflows, and definitions. All organizations face the practical realities of maintaining systems, minimizing customization costs, and integrating platforms. If decision makers aren’t clear themselves on what strategy execution requires, a solution that is low-cost or easy to integrate seems like a sensible choice — and the trade-offs aren’t immediately obvious.

Satisficing occurs when organizations restructure their workflows to fit with the systems, adopt standardized talent definitions, and use available capabilities and data fields as proxies rather than designing more nuanced, context-specific criteria and processes. Settling for “close enough” happens too often, especially when organizations accept the embedded aspects of vendor-designed systems. Enacting talent decisions based on externally designed assumptions will limit outcomes. Perhaps the platform assumes that talent is an attribute inherent in all individuals, whereas your organization believes that talent can be acquired and expanded through further development opportunities and that talent is transferable across contexts when leaders value a team-based approach.

A “close enough” tool may meet expense and speed requirements but fundamentally hamper the pursuit of strategic objectives. This is particularly salient because talent management and data-driven decision-making must take into account an organization’s unique strategy and its specific workforce needs and preferences in order to make effective talent decisions. Simplifying a hiring profile to fit platform fields, amending a performance rating scale to ensure data sharing across systems, or selecting only one critical role for succession planning because the system won’t allow users to select more: Each decision seems reasonable in isolation because it allows work to continue. But what began as a small and isolated compromise compounds as each satisficed area becomes embedded in everyday practice. A patchwork of suboptimized and unscalable processes can skew what is prioritized in talent decisions.

Advancing Strategy-First Talent Systems

Articulating the outcome you want to achieve — whether it’s greater agility, more innovation, or better retention — is essential for effective business impact, as is the decision-making infrastructure embedded in your HR technology stack. When the links among criteria, processes, and outcomes are not articulated or customized for your particular organization and objectives, the parallel link between talent decisions and strategy execution becomes weak or nonexistent. The system may produce operational output in the form of data and reports without facilitating meaningful strategic impact.

With a strategy-first approach, talent systems, criteria, and processes are intentionally designed to reflect your organization’s strategic priorities. Being specific about the reasoning — the “why” — underpinning talent decisions begins with clarity: What is the issue you seek to address or rectify? What is the opportunity you seek to capture? Unstated assumptions and beliefs are often embedded in this reasoning — many times implicitly, such that you may not even realize you’re making them. Assumptions seem as natural as the sun rising in the morning but may not be fit for your purpose and should be explicitly surfaced and tested.

Criteria are the “what” — what you look for, measure, and prioritize. Processes are the “how” — how you assess, identify, develop, and structure your workforce. These design choices should be anchored by your organization’s unique strategy. In this context, hiring profiles match strategic priorities, performance reviews evaluate what matters, and development plans target capabilities required for strategy execution.

The Strategy-First Check-In

Based on our extensive work with organizations, we’ve learned that examining an organization’s talent management decision infrastructure is essential to ensuring that it advances the company’s strategy. To assist in this investigation, we offer the Strategy-First Check-In, a practical diagnostic you can use to help determine whether your HR tech stack supports strategic precision. These four steps aim to surface precisely where your systems may diverge from your priorities.

The Blank-Page Exercise

Defining what an ideal system would look like in the absence of “close enough” constraints helps clarify what are you tolerating that you would never actively choose. Take a holistic view of your existing HR technology stack to identify which elements add value and which have merely become accepted ways of doing things.

Consider what you would include when designing a fit-for-purpose talent system from scratch. Reassess the links between talent criteria, processes, and outcomes to map out an ideal technology ecosystem that actually serves your strategy. What strategic outcomes and business impact do you seek to achieve? Which processes are essential or add value? Which salient strategic criteria and definitions of talent should be used in the newly designed system?

The Ecosystem Inventory

The next step is to inventory the active systems and tools running in your HR technology ecosystem. In large or complex organizations, that alone can be a significant amount of work. Building a list of the user-facing and back-end tools that are currently in place to run all of the various operations across HR, never mind determining the owner accountable for each system, can be daunting. Often, forming partnerships with IT and finance is required to uncover all of the separate elements at play within just the HR space.

The Accumulation Audit

The biggest benefit comes from understanding how each tool’s embedded criteria, processes, and decision logic shape talent decisions. Consider the following aspects of your HR technology stack.

System purpose and history: Establish why each system exists and how it became part of the ecosystem.

• What systems do we have?
• What does each system do?
• How long has the system been in operation? What is its history in the organization?

Ownership and users: Identify who is responsible for and uses each system.

• Who or what is responsible for the system?
• Who or what uses the system?
• Who depends on each system?

Data and decision logic: Understand how each system collects, stores, and deploys data and shapes talent decisions directly and indirectly.

• What data does each system ingest? What’s the input?
• What does each system feed? What’s the output?
• What constraints does each system create?

Ecosystem fit and value: Assess the fit and value each system contributes to your unique strategy.

• What value does it contribute? To whom or what?
• What would or might happen if this system were removed?
• Which systems are duplications or in conflict?
• What system capabilities are available but not currently used?
• How much support does each system require?
• How much does each system cost?

The Strategy Test

Gathering answers to the questions outlined above will enable you to answer deeper questions about how your talent tech stack enables or limits your goals and priorities and whether the system fits your current strategic needs. Of course, doing this effectively requires clarity on your organization’s existing talent strategy; for many organizations, getting a clear, coherent and consistent talent strategy may be the first, most critical step in this process.

Strategic alignment: Determine how each system reflects your strategy.

• What is the core of your talent strategy?
• How is your talent strategy reflected in and supported by each tool?
• How does each system link to business outcomes?
• Where does the system enable or constrain strategic priorities?

Embedded criteria, processes, and decision logics: Specify assumptions embedded within each system.

• What talent criteria, processes, and decision logics are embedded within each tool? Do they differ across platforms?
• Which talent criteria and processes are vendor defaults? Which were internally designed?
• Are criteria and processes linked to business outcomes and current strategy? How, or how should they be?
• Which talent criteria and processes require customizations and repeated investment?

System interdependencies and decision logic: Examine the coherence of decision logic across systems.

• How does data move between systems?
• Which system is set as the authoritative source of logics and data? Why?
• What are each platform’s embedded decision logics? Are they consistent?
• How are outputs interpreted and used?

System value and human work: Assess the strategic value of each system and identify human work.

• What value does the system generate relative to its costs?
• What critical processes are still run manually and why?
• Where must “human glue” be used to reconcile data, interpret outputs, or compensate for system limitations?
• Which workarounds are of strategic value?

This investigation will help you determine whether your systems are shaped by business priorities or by software limitations, and where you have compromised ideals for convenience.

Design and Drive Your Talent Systems

The Strategy-First Check-In provides a structured way to assess the extent to which your talent technology stack and associated decision-making infrastructure are advancing or constraining your unique strategy. This comprehensive analysis lays the foundation for designing fit-for-purpose systems.

The purpose of the exercise is not simply to cancel your vendor contracts. Organizations benefit from designing and driving their talent stack intentionally, and a thoughtful transition requires an examination of interconnected chains of systems and workflows to identify risks to data and operational health. That involves evaluating the power and agency vendors exercise over your talent criteria, processes, and decisions to understand who — your organization, a vendor, or a group of vendors — is currently driving your talent decisions. Next, you should examine the implications of using vendor-designed definitions, workflows, and default settings, and understand whether their influence is intentional or has emerged through gradual slippage over time. From there, you can determine where leaders and functions have continued to be the authors of talent systems.

When you design your talent systems, it enables your organizational strategy to serve as the anchor for talent conversations. And when your talent decisions are treated as investments rather than expenses that must be controlled, it strengthens your ability to deliver strategic outcomes. Rather than restructuring your talent decisions to accommodate vendor-designed tools, you are customizing your HR systems to enable better organizational decision-making. When doing so, keep the following in mind:

• Prioritize where to act first by identifying elements to depreciate, eliminate, adjust, or replace.
• Intentionally define organizational talent criteria and processes that account for many contexts rather than a single, static context. And regularly revisit those definitions — talent, strategy, and organizational contexts are dynamic.
• Ask about the value of talent-in-action by identifying the contexts in which the specific talents you’ve defined are of strategic value.
• Maintain systems not because of legacy usage but for the quality, compliance, or strategic relevance that they currently enable.
• Remain aware of the opportunities and dangers of using generative AI, to avoid metastasizing errors on scale.

As vendors continually add capabilities, new features might obviate the need for specialty tools. Eliminating such tools and letting the updated platform perform their functions reduces strain on the IT function: Having fewer data connections to manage is more cost-efficient and frequently creates higher-quality output as multiple functions or processes work together coherently.

Examining new functionalities will also help you decide whether to standardize on a single HR platform. Many organizations find that they’re using only portions of multiple platforms, often for legacy reasons, which adds operational complexity and data confusion to their HR IT system. HR leaders should identify where tasks or workflows could be simplified, streamlined, or automated.

Another common design strategy is to identify “uni-tasker” platforms that perform single functions. They are often the best candidates for replacement, if few or no other processes depend on them. Where a function can’t be absorbed by other tools already in your ecosystem, in some cases it may be a reasonable candidate for a homegrown or vibe-coded addition, where an AI agent may contribute. Using AI to automate routine tasks or to remove steps altogether can return both human and IT capacity back to the organization. The caution here is that you must pay careful attention to data privacy and stability.

Designing and driving your talent systems means that your organizational strategy serves as the anchor for talent conversations. Ultimately, the promises of HR technology stem from intentional design, not technology accumulation. When you invest the time reviewing how your technology stack advances or diverges from your strategic priorities, you’re better able to reclaim ownership as the designer and the driver of talent systems and ensure that your organization focuses on the wins that matter.

Employee stress is a major social issue both within and outside of our workplaces. As of 2022, stress had increased medical expenses and decreased productivity at an estimated annual cost of up to $300 billion in high-income countries. In the workplace, a proliferation of digital tools and applications, such as Zoom and generative AI, can cause technostress, or mental exhaustion resulting from technology use.1

While few businesses assess the high costs of technostress, even fewer are aware of its more positive potential. When introduced and managed deliberately, tech tools can promote employees’ work satisfaction and engagement by offering new opportunities for personal and professional development, and a resulting sense of accomplishment. On the other hand, constant adaptation to new digital tools dampens employee morale, which can be further exacerbated by the introduction of surveillance and tracking tools that may increase productivity but also provoke negative employee responses. Beyond raising employee stress levels, technology can also increase job dissatisfaction, burnout, and turnover rates, further compounding the pressure on the employees who remain.

Our laboratory experiments and large-scale surveys of thousands of employees have confirmed that many aspects of technology can hinder productivity, such as its inherent complexity, distracting notifications, and capacity to blur the boundaries between people’s work and personal lives. In addition, anxiety can increase and progress can slow when employees perceive tech tools as mechanisms for managers to replace them or monitor their performance.2 But while tech challenges may be interpreted as highly stressful, they can also produce endorphins and motivate achievement, greatly improving employee well-being and productivity, as well as loyalty to an employer.

Managers’ Role in Changing Worker Perceptions

In workplace contexts, it is the kind of stress and employees’ reaction to it, not the quantity of stress, that matters.3 Irrespective of the type of technology that creates stress for employees, the impact depends on whether workers perceive it as an inconvenient obstacle or a welcome challenge. A reframing that presents new technologies as exciting opportunities to grow, work smarter, and collaborate better can elevate employee productivity, morale, and retention.

That reframing starts with distinguishing challenges from hindrances and avoiding ambivalence.

Distinguishing Challenges From Hindrances

Most managers overlook the productive potential of tech stress, focusing instead on reducing hindrances, which we define in our research as stressors interpreted as obstacles to work-related activities. Hindrances offer employees no challenge or sense of motivation, fulfillment, or achievement. They are purely negative, inconvenient stressors that do not provide constructive opportunities for growth. The following are common examples:4

• Technology that is unreliable, unavailable, overly complex, or malfunctions frequently, impairing work activities.
• Frequent interruptions and notifications, which can fracture concentration.
• Surveillance or tracking tools that invade employees’ privacy to monitor and assess their performance.
• Technology that frequently updates or changes, which requires employees to continually adjust work processes and creates uncertainty and frustration.

Challenges, on the other hand, are positive stressors that are interpreted as true growth opportunities. They can encourage motivation and boost feelings of achievement and fulfillment. Here are some examples:5

• Technology that enables a fast pace, the ability to work on multiple tasks or projects simultaneously, and the completion of work.
• Technology that requires employees to use a broad set of skills, enabling them to learn and develop new abilities.
• Technologies, such as enterprise systems, that offer front-line employees access and authority previously reserved for supervisors, so they can solve complex problems.
• Organizational practices that encourage and reward experimenting with and using new tech tools.
• Organizational practices that involve employees in making IT decisions rather than subject them to top-down decisions.

Beyond the challenges and obstacles discussed below are new concerns about generative AI, which were not within the scope of our research. The managerial response to the perceived stress and threat of AI will differ from the approaches we suggest below. Most tech stressors concern how work is accomplished, but because GenAI may call into question the very necessity of human employees, the remedies described here are insufficient. Leaders who leave these fears unaddressed will find them surfacing as disengagement, absenteeism, and reduced productivity.6

Avoiding Ambivalence

Our data shows that hindrances and challenges do not cancel each other out. The two inherently conflict and are difficult to reconcile. Yet the advice commonly given to managers is to control hindrance stressors while simultaneously promoting more challenge stressors.7

Our recent data sets have shown that the simultaneous presence of negative and positive tech stressors leads to a highly problematic side effect: ambivalence, or the state of having mixed or contradictory attitudes and feelings. The difficulty of reconciling a plurality of conflicting stressors sets in motion a subconscious process that employees perceive as taxing, effortful, and exhausting.

Ambivalence is almost always experienced negatively and is strongly linked to frustration and distress. It can manifest in internal conflict, anxiety, and even angst that drives over-stressed workers to seek ways to cope — possibly, in extreme cases, by quitting their jobs.8 When employees experience this ambivalent tension between embracing and feeling overwhelmed by technology, job satisfaction drops and attrition and turnover increase.

Guidance for Balancing Tech Stressors

To challenge their workforces rather than impede productivity, leaders must address the tech stressors they impose on employees to mitigate ambivalent and negative reactions.

Reducing hindrances before implementing challenges is novel advice that has not been previously given to managers.9 We propose a proactive, three-part approach and offer additional strategies to create a stimulating work environment that offers positive challenges while managing stressful hindrances and mitigating employee ambivalence.10

Step 1: Reduce Technology Hindrances

First, leaders need to understand the current level of technical challenges and hindrances their employees face. This can be done through surveys about technology-related work stress.

Approaches to addressing hindrances can be divided into three categories:11

• Targeting personal factors involves mindfulness and technology training, strengthening employees’ control over how, when, and on what they work, and reducing pressure to remain available after working hours.
• Targeting organizational factors entails creating a culture that accepts and embraces mistakes, provides mentoring relationships, and offers access to supportive resources.
• Targeting situational factors involves engaging in job redesign to reduce unpredictable work demands and ensuring that the tech tools provided do not introduce unexpected obstacles.

Many hindrances are obvious and simple to eliminate. Technology that is unreliable and malfunctions frequently, for example, should be upgraded or simply replaced.

Step 2: Increase Positive Challenges

Only after hindrance stressors, such as invasive distractions and constant connectivity, have been reduced should leaders focus on positively challenging employees with technology to boost their job satisfaction and performance.

For example, leaders can modify roles to require a broader, more complex set of IT functionalities and technological skills, and give employees extra time to learn and develop new abilities. They can also deliberately provide tech challenges that allow employees to develop special expertise in mainstream office applications, such as SAP or Microsoft Excel, or update the IT infrastructure to include applications and devices that can help employees improve their work methods and become more innovative. Recognizing employees for adopting new technologies and involving them in technology change and implementation can also have a positive effect.

Another action leaders can take is nurturing a culture that encourages employees to proactively explore and experiment with new technologies. Informal meetings, such as brown bag lunches, are a great way for teams to share knowledge about new applications in a relaxed setting, as are videos that employees create themselves and share on enterprise channels. Giving employees the opportunity to request changes to the features of these applications or suggest different tools entirely can help them feel more involved in the process of choosing workplace applications. This can cultivate a sense of ownership and empowerment.

Step 3: Help Employees Overcome Ambivalence

If ambivalence lingers, leaders can help employees in two main ways. The first is to offer ample IT support and guidance. Few leaders have the resources available to employ more help desk personnel, but specialized software, such as NinjaOne, provides fully integrated tools to manage the entire help desk support process, including functions for IT ticket management, problem resolution, collaboration, and documentation. Such tools can help address current IT issues while employees build an internal knowledge base and, by helping them resolve hindrances faster, achieve greater efficiency and reduce their sense of ambivalence, even when resources are limited.

Another way to help reduce ambivalence is by investing more in IT training and coaching initiatives. An especially cost-effective approach is the use of self-paced online learning platforms, which are designed to make the learning experience fun while users develop skills in using complex workplace applications. Leaders can subscribe to these platforms or simply encourage employees to access them. Microsoft Learn, for example, features hundreds of training modules covering topics ranging from basic applications to advanced prompt engineering. Likewise, the SAP platform offers hundreds of what it calls Learning Journeys, and Salesforce provides free access to its training tool, Trailhead, which features hundreds of pathways for learning how to solve practical problems through step-by-step guides.

The most effective approach to addressing ambivalence, however, is to establish a mentorship program that pairs employees with expert users who can provide system-related advice and on-the-job coaching. Compared with virtual or video training formats, mentorship is more informal, friendly, and specifically focused on individuals’ particular needs.12 Mentoring can make employees feel more comfortable about seeking advice, especially when they are frustrated by the dual role of technology as a negative hindrance and a positive challenge at once.

For employees to grow through challenges rather than be hampered with technical inconveniences, leaders must carefully consider the tech stressors new tech can impose. Optimally — and contrary to conventional wisdom — leaders should reduce unproductive hindrances before increasing the number, frequency, and complexity of constructive challenges. Effective leaders pay close attention to the role that ambivalence plays in their employees’ stress levels and, by extension, their well-being, performance, and quiet-quitting or resignation rates.

The cost of getting this wrong is measured not just on an individual level. It also shows up as licenses paid for never-used tools, enterprisewide transformation programs that stall, and strategies that fail not because of technology but due to limited workforce capacity. Eliminating hindrances and offering new challenges can increase employees’ abilities, engagement, and drive.

Imagine that, as an operations manager at a global retailer, you have invested heavily in data science. Your data science team has built sophisticated models that produce demand predictions for the company’s products with impressive accuracy. Yet, when the CEO, seeking to increase profit, asks you how much inventory should be held and where, the answers are not clear. The models predict demand but do not tell management what to do with those predictions. After investing millions of dollars in data science solutions, the company realizes it has many insights about demand but few decisions to act upon.

This is not a unique scenario. Studies suggest that 60% to 85% of data science projects end in failure.1 The reason is not usually bad algorithms or incompetent data scientists. More often, there is a disconnect between the models that the data scientists produce and the information that business managers actually need. Often, the data scientist does not speak the language of business, and management does not speak the language of the data scientist.2 Data scientists are trained to focus on prediction and accuracy metrics, whereas managers’ focus is on making choices that improve business metrics such as cost, market share, or profit. As a result of this communication gap, models often look elegant on paper but sit unused in practice. As one pair of analytics experts put it, “The number one reason data science projects fail is because businesspeople and data scientists, individually and collectively, don’t understand the real business problem at hand.”3 The focus on the actual business decision is missing.

This is where decision science comes in. Unlike data science, which focuses on modeling the data, decision science models the decisions that business managers must make. This aligns the data model analysis with the business goals. By ensuring that the models are directly related to the business problem, your team can avoid relying only on “convenient data” and tie the model metrics directly to business outcomes.

For business leaders, the message is clear: If your analytics investments are not delivering ROI, stop modeling just your data and start modeling your decisions. Here is what that looks like in practice.

Decision Science Versus Data Science

Decision science focuses on improving decision-making using data: As the name suggests, the emphasis is on decisions rather than on data. In a recent study, we examined academic programs and job ads in data science and decision science, with the goal of determining how their skills differ and/or complement each other.4 We examined required coursework in both data science and decision science master’s degree programs and looked at the skills listed in job ads for each role.

We found that data science places a greater emphasis on skills related to coding, prediction, and machine learning. In contrast, decision science focuses more on directly modeling decisions using data, logic, and mathematics. This skill is known as decision modeling. Decision science, therefore, tends to focus more on problem framing, problem-solving, and business domain knowledge. Decision science relies on data science for predictions and on other data-based relationships as inputs to decision models that ultimately inform and drive decision-making. Your organization needs both skill sets.

The bread and butter of data science is prediction and machine learning. As a previous study of ours confirmed, data scientists are trained with a focus on modeling data, not on decision modeling. As Bipin Chadha, senior vice president of data science at CSAA Insurance Group, said in a presentation at Machine Learning Week 2024, “Many data scientists mistakenly think predictions are decisions.” Indeed, this is the crux of the disconnect between data scientists and management.

This data science disconnect — focusing on the data rather than on the decision to be made — can result in teams solving irrelevant problems and/or using the wrong metrics. This often leads to management rejecting solutions.5

Models do not create value until managers use them. And managers use them only when the analysis reflects the decisions they face. Shifting the focus from data models to decision models compels analysts to engage directly with leadership, sparking conversations that uncover root causes, align priorities, and build trust in the recommendations.

How Decision Models Can Help

When a data science team models an organizational decision, not just the data, it forces communication between the analytics team and management, mitigating a big hurdle to project success. Here is how decision models make a difference.

Decision Models Strengthen Problem Framing

Managers expect the data science team to generate ROI by solving problems and providing the intelligence needed to make better decisions. However, managers often do not have the time to truly understand the root cause of the problem. Managers tend to experience the symptoms of some deeper problem.

A good decision scientist is like a good medical doctor: They must ask a lot of questions about symptoms to navigate to the real underlying problem. Decision modeling forces this discussion and reinforces that the most critical step is identifying the true problem that can then be modeled. The problem must be properly framed before it can be solved.

For example, consider the case of forecasting demand for 50 products over the next four quarters. Your data scientists build a set of 50 highly accurate machine learning forecasting models that include internal and external features. The data science team reports to management that, based on the model validation holdout samples, the mean absolute percentage error (MAPE) on the models ranges from 6% to 9%. The data scientists have done the job that they were trained to do.

As a manager, you may or may not know what MAPE is, but you know that solving the decision problem you face requires accurate demand forecasts. But the decisions you need to make are more complex than that: how much of each product to produce, and how much to hold in inventory to maximize the contribution to profit and fixed costs. Management’s metric is clear: dollars.

The real decision problem, then, is not one of forecast accuracy, though the decision problem is dependent on forecast accuracy. It is something like the following: Maximize contribution to profit and fixed costs while following our business rules.

To model contribution to profit and fixed costs, you will need estimates of the per-unit contribution for each of your products, the opportunity costs of unmet demand, and the cost of holding inventory. Business rules will have to include limited production and storage capacity, material availability, and labor availability. And you will need the demand estimates from the forecasting models because they will inform you about how much you can expect to sell. Importantly, the costs of producing too much and too little often differ significantly, and focusing purely on model accuracy glosses over this reality by treating each error equally.

To model this decision, the decision scientist’s approach would, perhaps, be to use an optimization model in combination with scenarios or simulation to factor in the uncertainty in the model inputs. The result would be a recommendation (how much of each product to produce and hold in inventory each quarter) or, better yet, a set of alternative recommendations that allows you to decide which of several possible plans to implement based on a return-versus-risk trade-off.

Decision Models Protect Against the ‘Convenient Data’ Trap

A common shortcut in data science modeling is to use the data that is easiest to access rather than first clarifying the decision problem and then asking, “What data would help us solve this problem?” This shortcut leads to elegant models that do not actually model the business decision that will drive business value.

Take, for example, the challenge of customer churn — a common issue for many companies in a variety of industries. Telecommunications companies frequently build models to predict which customers are most likely to leave, drawing on readily available data such as usage history and demographics. These models can generate accurate predictions, but they fall short of answering the managerial question “Which customers are worth retaining, and what actions will convince them to stay?”

Answering that decision problem requires a different lens and different data: Which customers are likely to purchase additional products/services? Which customers are most profitable? Which customers are most valuable to competitors? Which retention offers change customer behavior? These insights are rarely found in the “convenient” data sets already on hand.

By modeling the decision first and then identifying the right data, managers can avoid the convenient data trap. Doing so ensures that analytics not only predict outcomes but also guide actions that deliver measurable business value.

Decision Models Foster Metrics Alignment

Data science metrics (such as goodness of fit and maximum likelihood) may be used to construct inputs to the decision model, but those metrics are going to be of little interest to management. Decision modeling ensures that the analyst uses metrics that matter to managers.

For example, consider the challenge many financial institutions face when deciding whether to approve or decline a loan request. A typical analytical approach uses data on past loans to build a model that estimates the probability that a new loan holder will default. From there, a threshold is set to create a rule — for example, “If the probability of default is higher than 0.30, then decline the loan request.”

To evaluate how well a model performs, analysts often use a tool called a confusion matrix. In simple terms, the confusion matrix shows where the model gets its predictions right and where it gets them wrong. For example, it identifies cases where the model predicts that loans will be repaid but the loans later default, or predicts that loans will default but they are ultimately repaid. Analysts then weigh these different kinds of errors to fine-tune the model.

The problem is that those measures, while important for statistical accuracy, are not the same metrics that matter to management. Executives do not just care about whether the model is right or wrong in general; they care about the financial consequences of those mistakes.

A decision modeling approach starts with the financial goal, such as maximizing profit within acceptable risk. The predictive model supplies inputs, but the decision model translates those inputs into dollars, showing the financial consequences of different threshold choices. Most importantly, it captures the asymmetry of outcomes: The cost of approving a bad loan is different from the cost of rejecting a good loan.

Four Steps to Building a Decision Focus

If your organization’s analytics investments are not delivering the anticipated ROI, here are some practical ways to bring a decision focus to your data science team and projects.

1. Start projects with a decision brief, not a data request. Amazon’s “working backwards” process requires teams to begin a new product or initiative by writing a short mock press release and an FAQ document before building a product. That discipline forces the team to define the intended customer, the problem being solved, the value that will be created, and the tests of success, before the company commits resources. Analytics projects need a similar discipline.

Rather than beginning with a data request or model specification, managers should require a decision brief that works backward from the managerial choice to be improved: who will make the decision, what alternatives are available, what constraints and uncertainties matter, and what business outcomes will define success. In both cases, the value of the upfront document is not bureaucracy; it is clarity. It prevents teams from building something technically impressive that does not solve the problem that matters.

Here is a real-world example from Ben Martin, vice president of data, analytics, and AI at apparel company HanesBrands, who has a history of leading effective analytics and data science teams. Martin has developed a process that he uses for all data science projects: No project goes forward without acceptable responses to the following questions.

• What is the business problem to be solved, and who cares?
• What decisions need to be made to solve the problem?
• What is the anticipated business value in solving this problem?
• What are the analytical techniques and data that are needed to help make these decisions?
• Is the project feasible, with a high likelihood of success?

Note how the process starts with the business problem, the decisions to be made, and an assessment of the value and impact of solving the problem. Martin’s guiding principle is “Decision analytics drive ROI; everything else is just fun and interesting facts.”

2. Create an enterprise digital twin as a lab for your data science team. Chadha, an engineer by training, started out on CSAA Insurance’s strategy team and later moved to lead its data science efforts. He recognized the disconnect between data scientists and management and the importance of connecting the data scientists to the business. According to Chadha, “You have to learn the business to be focused on decisions. A lot of people say that but don’t do anything about it.”

So Chadha created an enterprise digital twin — a simulation of the company — that effectively became a learning lab for his data scientists to get up to speed on the business. This not only helped his data scientists learn the business and become more decision-focused but also forged a laboratory that allows the data scientists to test the potential impact of results by applying the business rules in the digital twin.

3. Do not accept predictions without prescriptions. Managers should expect analytics teams to translate predictions into recommended actions. For example, UPS’s On-Road Integrated Optimization and Navigation (ORION) system is a route-optimization platform that uses data, advanced analytics, and operations research to recommend more efficient delivery routes for UPS drivers. ORION illustrates the difference between prediction and prescription. The value of the system is not simply that UPS can predict traffic, delivery times, or package flows more accurately. Jack Levis, who led the development of ORION at UPS, estimated that 10% to 25% of ORION’s success can be credited to better predictive models.6 The real value comes from translating those inputs into optimized route recommendations that drivers can use. In other words, the analytics did not stop at “What is likely to happen?” It answered “What should we do?”

4. Build and lead the analytics team for decision impact. When hiring talent, stop chasing the mythical analytics unicorn. Instead, build teams that combine data science and decision science capabilities. Also, make business translation a leadership responsibility. Data scientists bring coding, data engineering, machine learning, prediction, and statistical modeling skills. Decision scientists bring problem framing, optimization, simulation, decision analysis, and business acumen.

The key is not simply to assemble technical talent but to ensure that the team can connect data, models, business rules, organizational constraints, and managerial choices. Disney’s decision science and integration group provides one example of this model: This centralized internal consulting group uses optimization, simulation, machine learning, data engineering, and AI to address business challenges in pricing, revenue management, and operations.

Leadership matters as well. In many successful analytics organizations, the leader is not necessarily the best technical modeler but rather a talented translator who connects technical work to the decisions that managers actually face. Chadha has degrees in mechanical engineering. Martin studied textile engineering and technology management. Levis was a psychology major. Their examples reinforce an important point: Analytics teams create value when they are built and led for decision impact, not just technical output.

The takeaway for leaders: Data science creates value only when it is connected to the decisions that managers must make. Predictions and insights are valuable inputs, but the end point is a better organizational choice. As a leader, your role is to ensure that analytics teams do not stop at modeling the data. They must model the decision, translate technical performance into business impact, and recommend actions that managers can take. The goal is not simply to become more data-driven but more decision-driven.

For the fifth year in a row, MIT Sloan Management Review and Boston Consulting Group (BCG) have assembled an international panel of AI experts comprising academics and practitioners to help us understand how responsible artificial intelligence is being implemented across organizations worldwide. In previous posts this year, we have explored AI’s impact on the workforce, what human verification of AI outputs entails, and how to think about agent autonomy.

New research has emerged suggesting AI overuse or overreliance risks degrading our cognitive and moral judgment, resulting in “epistemic de-skilling” and reduced critical thinking skills. Evidence-based concerns about AI’s impacts on children’s cognitive and emotional development, in particular, led New York City to introduce the nation’s first ban on generative AI in schools.

Given the increasing salience of these worries, as a final provocation in this series, we asked our panel to react to the following: The defining challenge of responsible AI will be preserving trustworthy human judgment in the face of capable machines. The majority (65%) of our panelists agree or strongly agree with the statement, contending that human judgment is a muscle worth preserving, even if it is increasingly at risk of atrophy. However, a significant minority (19%) disagree or strongly disagree, and a further 15% remain neutral (neither agreeing nor disagreeing). The experts diverge on what human judgment means and how it differs from human oversight, and cite its inherent limitations. But beneath the surface, broader agreement appears to exist around this concept: The goal isn’t to preserve human judgment for its own sake but to design human-machine systems that keep trustworthy human judgment available when it matters most.

Below, we share panelist insights and offer practical recommendations for organizations looking to create systems that build in room for meaningful human judgment as AI grows more pervasive and capable.

Human judgment is a muscle worth preserving. Most of the experts believe that human judgment is worth preserving while machines’ capabilities increase. Ben Dias, IAG’s chief AI scientist, says, “The defining challenge of responsible AI will most certainly be preserving trustworthy human judgment as increasingly capable AI systems operate at greater speed and scale.” Similarly, Alyssa Lefaivre Škopac, Alberta Machine Intelligence Institute’s director of AI trust and safety, believes that “judgment, critical thinking, and the ability to evaluate AI outputs and actions will be the most important and differentiated skills going forward.”

For many of our panelists, human judgment is critical for providing context that machines lack. United Nations University rector Tshilidzi Marwala says, “AI can process information and generate recommendations at extraordinary speed, but capability does not equal wisdom, context, or legitimacy.” Consultant Pierre-Yves Calloc’h explains: “AI systems live in digital worlds, which are a partial representation of reality.” He adds, “Only a human living in the real world can assess whether an answer or recommended action generated from a prompt and the model’s training material is truly fit for purpose.” For Automation Anywhere’s Yan Chow, “human judgment matters most where machines are weak — novel situations, ethical trade-offs, context not captured in training of large language models.”

Many experts also contend that preserving human judgment is essential for maintaining human agency and accountability. “Human judgment, grounded in expertise and experience, is what lets us steer, verify, question, make trade-offs, and stay accountable when working with AI,” says consultant Linda Leopold, and serves as “the foundation of human agency.” But she also cautions that “judgment and critical thinking are the skills we risk losing if we rely too heavily on AI, which is why preserving them is essential.” Marwala agrees that the defining challenge of responsible AI will be ensuring that “humans retain meaningful judgment, agency, and accountability” as machines’ capability expands.

The answer is not simply to keep a human in the loop. Many experts also recognize the limits of human judgment. As GovLab chief research and development officer Stefan Verhulst notes, “Human judgment was never inherently trustworthy.” Chow explains that “humans face fatigue, bias, lack of experience or expertise, and poor calibration under uncertainty.” In fact, Instalily.ai CEO Amit Shah cautions that “keeping humans continuously in the loop can actually degrade quality by reintroducing inconsistency and error into decisions machines may be better equipped to make.” Going further, Yasodara Cordova, distinguished member of the investments committee of the Co-Develop fund, cautions that “human judgment has never been uniformly trustworthy, and building machines to defer to it risks encoding existing hierarchies of whose judgment counts as legitimate.”

Given the limits of human judgment, TÜV AI.Lab CEO Franziska Weindauer argues that “a human in the loop does not automatically make a system trustworthy.” For RAIght.ai co-CEO Richard Benjamins, “Human judgment is a bigger idea than oversight alone,” and what actually needs preserving is judgment that “serves humanity as a whole, not just the interests of particular groups.” Taking this a step further, Stanford CodeX fellow Riyanka Roy Choudhury argues that “neither human nor machine judgment should be presumed trustworthy.” For Choudhury, trustworthiness means that “no decision maker, human or artificial, should be the judge, jury, and final appeal.”

Growing AI reliance threatens to displace and erode trustworthy human judgment. Despite a consensus around the worth of preserving human judgment, our experts worry that a growing reliance on AI threatens its displacement and erosion due to “incentives to adopt, automate, and optimize,” as Škopac puts it. Dias observes that “as AI production outpaces the capacity for meaningful human oversight, human review becomes the bottleneck, [and] people are progressively removed from decisions.” And Renato Leite Monteiro, vice president of privacy, data protection, AI, and intellectual property at e&, worries that because “AI is anthropomorphized and agreeable by design,” we will “quietly stop valuing the judgment of the people we ask to supervise [it].” The risk, says Ranier Hoffmann, chief data officer at EnBW, “is that organizations design judgment out of the work while leaving humans nominally accountable.”

Experts also worry the trend of overreliance threatens moral and cognitive capacity. As Hoffmann observes, judgment “develops through consequential decisions, including wrong ones.” When machines decide, “they remove the ground on which judgment was trained,” he adds. Škopac agrees that “as more entry-level tasks get offloaded, so does the learning, muscle memory, and scar tissue that accompany these ‘tedious’ tasks that we understandably see an opportunity to automate.” Dias cautions that “prolonged human disengagement erodes the expertise needed to recognize failures and intervene effectively.” Researcher Kirtan Padh hopes we can use AI to improve “productivity, without losing our learning and critical thinking skills.”

Trustworthy systems focus on human-machine collaboration. In the end, Shah says, “the defining challenge of responsible AI is not preserving human judgment for its own sake.” Instead, as AI grows more capable, ADP’s chief product owner Naomi Lariviere believes, the challenge will be figuring out “where human judgment actually matters.” For Shah, “human judgment belongs primarily upstream: setting objectives and guardrails, defining thresholds for autonomous action, determining when uncertainty or out-of-distribution cases require escalation, and assigning responsibility for outcomes.” For Wharton School professor Kartik Hosanagar and others, it matters most for “mission-critical tasks” and in “high-stakes settings.”

For many experts, preserving trustworthy human judgment when it counts is a matter of design. Because “humans and machines excel in different scenarios,” Chow contends we must build “systems that recognize that difference.” Verhulst agrees we should design “human-machine judgment environments in which each compensates for the weaknesses of the other.” Škopac is optimistic that “after a few scraped knees, it’s going to become clear that completely offloading to AI results in worse outcomes, low morale, and more embarrassing headlines.” If we strike the right balance, Dias believes, we can “preserve human accountability, competence, and authority, while maximizing the value extracted from AI.”

Recommendations

Considering the above, we offer the following recommendations for organizations seeking to design systems for human-machine collaboration that preserve the opportunity for trustworthy human judgment:

1. Identify where human judgment matters. Not every decision requires human review or intervention. Organizations should identify where human judgment matters in their specific decision-making processes and workflows, and focus on preserving the opportunity for human judgment in the context of high-stakes and low-precedent decisions.

What’s more, human judgment will continue to be essential in determining the philosophical bases on which human-AI machines operate. What is the purpose of a given application? What actions do we allow the system to take autonomously? What is appropriate evidence for a decision — that is, when do we know enough to make a decision? These questions should be foundational to the design of every AI system, presenting the earliest embodiment of human judgment and accountability for AI. As Shah remarks, human judgment is critical to these “upstream” concerns.

2. Create the opportunity for human judgment by design. Where human judgment matters, the opportunity to exercise it must be designed into the system, not assumed that it will exist. Simply keeping a human in the loop is not enough here. A human asked to review or approve an outcome requires the context and evidence necessary to make an informed judgment. And human review cannot occur after the fact when there is little practical ability to change the outcome. Organizations should build the opportunity for real human review and intervention into the system, process, or workflow from the outset, so that when judgment is needed, there’s an actual point at which it can be exercised.

3. Use it or lose it. Judgment is a skill that will atrophy if it isn’t exercised, and the high-stakes, low-precedent decisions where human judgment matters most are too rare to provide the only training ground. Organizations should deliberately build in opportunities to exercise judgment outside of high-stakes moments, through simulations, case studies, and postmortem or retrospective exercises. They should go a step further, deliberately keeping some lower-stakes decisions in the hands of humans instead of offloading to AI as practice for when it counts. Retaining human decision-making will carry a small near-term opportunity cost but will have net benefits over the long term by building critical organizational capacity.

4. Calibrate training by level. How judgment is honed should look different depending on experience level. Junior team members need chances to build judgment they haven’t yet developed, while senior decision makers have established judgment but fewer opportunities to exercise it when they rely more on AI-assisted tools or delegate routine decisions. Organizations should design distinct pathways for individuals of different levels of experience.

5. Build feedback loops. Neither human nor machine judgment should be trusted by default. Organizations should track outcomes on both sides, such as when the human override of an AI recommendation was right or wrong, or when deference to the AI output was right or wrong. This record can help organizations recalibrate where responsibility should sit, since human-machine collaboration is something that will require ongoing adjustments as AI capabilities evolve. It will also be critical to assess whether human oversight is effective.

Do human reviewers blindly follow the AI recommendations or do they exhibit true judgment? This analysis will identify areas that may require redesign, additional training, or other interventions.

Jakob Ingebrigtsen, one of the best middle-distance runners ever, has won gold medals for Norway in two Olympics, won multiple world championships, and set five world records at distances between 1,500 meters and 2 miles. Watch him and you’ll see that he’s often aided by someone running with him for part of the race — a pacer. If Ingebrigtsen runs too fast, his pacer helps slow him down; if too slow, the pacer speeds him up; if he’s on the right pace, the pacer just relieves Ingebrigtsen’s mental energy by allowing him to follow without continuously monitoring his timing. Elite athletes themselves, Ingebrigtsen’s pacers have the sole goal of helping him win. When the pacer’s job is done during the race, they simply step off the course and stop.

Similarly, marathoners and distance swimmers like Diana Nyad often use pacers so that they can focus mentally not on pace but the myriad other details that lead to outstanding performance. The right pacer can also distract from or help minimize an athlete’s mental or physical anguish and help them endure the challenging experience. Pacers, in short, reduce the cognitive load of the performer, help the performer better manage their energy, and prevent second-guessing and anxiety that can affect both short-term performance and long-term well-being.

Leaders, too, often create significant performance problems for themselves, their teams, and their organizations by running too fast or too slowly. Leaders may spend far too much time and cognitive energy monitoring one factor at the expense of others that are critical to success. And leaders, too, can suffer from the mental and physical effects of running at the wrong speed, such as losing confidence in their own timing-related decisions.

Nonetheless, many of us focus primarily on what leaders do without paying sufficient attention to the pace at which they do it — and rarely treat the latter as a distinct dimension of leadership effectiveness. As a result, leaders are seldom encouraged to examine whether their natural tempo fits the demands of the moment or to consider who or what could help them recalibrate their personal pace. In our experience coaching senior leaders, we have repeatedly seen the problems that stem from this oversight. Getting the pace of leadership right can be as consequential as getting the direction right.

To explore the related issues for leaders and their teams, we surveyed 125 employees in a vast array of professional and leadership roles. What we learned is that pacers of various kinds perform the same essential roles for leaders, with benefits similar to those experienced by athletes. A pacer can rein in leaders who move too fast too early or, alternatively, speed up decision-making for leaders inclined to be slower and more cautious.

Do you need a pacer? Let’s explore some warning signals and questions to ask, as well as advice on utilizing this type of resource.

What Pacers Provide for Leaders

In our survey, we asked respondents to consider a current or former leader of theirs and describe why they think that leader tends to run too fast, too slowly, or at just the right speed. For leaders who ran fast or slowly, we asked whether the leader had an identifiable pacer and who the pacer was. Then we asked everyone about the positive and negative consequences for the leader and their team.

Nearly all respondents (more than 93%) described a leader whose pace was often at odds with the needs of the situation. Among that 93%, a notable three-quarters said their leader tends to run too fast while one-quarter said they run too slowly. How many of these leaders had pacers? Just 40%. That means a full 60% of these leaders, whom we’ll call “caffeinated squirrels” and “tentative tortoises,” respectively, had no one to help counterbalance the limitations of their natural pace. So the leaders’ natural advantages — being quick, bold, and decisive for the squirrels, or careful, thorough, and thoughtful for the tortoises — often went from strengths when moderated to weaknesses when unchecked.

Perhaps most interestingly, the benefits afforded by pacers were largely the same for both groups, helping leaders slow down or speed up, or shift gears more regularly to fit specific circumstances — for example, sprinting in a true crisis, slowing down when high uncertainty required exploration, and maintaining their pace during sustained implementation phases. Our findings reveal that pacers help leaders:

• Make better decisions and learn more from prior actions.
• Stick to strategic goals and related metrics and timelines.
• Manage others’ energy levels, motivation, and commitment.

Making Decisions and Learning

A pacer can help fast-moving executives slow down enough to appropriately deliberate before committing to high-stakes action. One survey participant explained that his leader’s pacer helped ensure that projects were “well-thought-out and rarely chaotic while moving forward at a sustainable rhythm.” Another respondent likewise said that the benefits of their leader’s pacer included “not making decisions without being generally aligned and understanding of the risks.” Pacers who help fast leaders slow down also allow time for group learning, the absence of which leads to bad decisions being repeated because they “happen so fast that they aren’t processed before repeating” and “no one has time to solve problems or learn for themselves,” according to survey participants.

For leaders naturally inclined to move slowly, pacers help break through analysis paralysis by convincing them to accept decision deadlines and “good enough” thresholds for action, respondents said. They do so by helping leaders see who or what is closing the gap on the organization’s position — illuminating blind spots such as an advancing competitor, an evolving technology, or macroenvironmental changes creating headwinds or tailwinds.

In other words, pacers help their leaders avoid decision-making and learning failures that would occur at the person’s natural speed. Pacers create feedback loops and action guardrails that otherwise wouldn’t exist. They channel information related to decision pace and outcomes from people who are afraid to share it with the leader directly. They can also help a leader set (and follow) processes for adjusting their natural pace.

Sticking to Strategy

Without a pacer, a leader’s behaviors may frequently be at odds with the organization’s core strategy and objectives. For example, fast movers can be so consumed by speed that their actions don’t ultimately add up. Said one respondent, “The leader demands progress, but because there’s no strategy, everything accomplished is a tactical gain, and the tactical actions are not tied to operational or strategic outcomes.” Unchecked by a pacer, such leaders may react to every new signal, leading to strategic drift and emotional whiplash and confusion for the team.

On the other hand, slow-moving leaders, daunted by the magnitude of the race, may repeatedly delay leaving the starting blocks or fail to speed up despite warning signals about progress. One respondent described watching “the consequences playing out in real time as we struggle financially and fail to meet year-end targets.”

Here is where a moderating influence could make a real difference: As one participant whose leader had an effective pacer noted, “It’s harder to ignore or change performance metrics in real time when you’ve agreed to hold to a pace.”

Energy Management, Motivation, and Commitment

By far, the most common factor respondents mentioned was the unchecked leader’s effect on the energy, motivation, and commitment of people around them. Without a pacer, fast movers continually burn out those around them — not only because of the pace itself but also the emotional fallout from frustrating last-minute changes and unnecessary mistakes. One respondent noted the depressing realization that “everything becomes a competition, and everyone is trying to outshine the next person.” Such a reality leads to turnover. In the words of another respondent: “Burnout. Cash-flow heart attacks. I quit my job despite loving the team, culture, mission, and my job.”

A fast mover may also cause broader misalignment problems among groups and units that don’t buy into the leader’s requests or aren’t equipped to move at the leader’s pace. A leader who fails to adapt to the different schedules or priorities of other units creates missed team deadlines and many “moments of avoidable rework,” one person noted. But a pacer can push a leader to consider and integrate another group’s pace into decisions about operating speed. This dynamic, explained another respondent, “keeps decision-making grounded, prevents burnout, and creates more consistent alignment across the organization.”

Meanwhile, naturally slower leaders don’t burn their people out: They bore them out. Respondents described top talent becoming frustrated by continuously missed opportunities and cynical about the future in the absence of a pacer who could help spur faster action. The result? Many people leave voluntarily; others get jaded and give only a fraction of their best.

How to Pace(r) Yourself

As obvious as the problems caused by fast movers and slow movers may be to the people around them, it’s often hard for a leader to connect the dots. If you want a more well-rounded view, seek input from a few people you believe are most likely to tell you the truth.

You can also take a look in the mirror and consider these signals that you’re running too fast:

• Your team regularly asks for clarification on priorities you thought were obvious.
• You’ve reversed a major decision within 90 days more than once in the past year.
• Key people describe your organization as “exhausting” in engagement surveys or exit interviews.
• You feel frustrated that others “can’t keep up.”
• You’ve launched initiatives that died from neglect because the next thing had already arrived.

And look for these signals that you’re running too slow:

• Competitors often launch what you’re still “studying.”
• Your default response to a new idea is to commission more analysis.
• Your team uses phrases like “when leadership decides” or “still waiting for direction.”
• You have more than three “pending” strategic decisions older than six months.
• Those above you ask you what’s taking so long regarding new initiatives.

Choosing the Right Person

If you believe that you might benefit from a pacer — or that you need a new one — ask yourself some questions about what you need from them, beyond the tempo adjustment itself. For example, do you need help with your general leadership pace, or with domain-specific matters like financing, exit decisions, or how to pace technology adoption or growth? Do you need an ongoing relationship with a pacer, or is your need limited to a specific time frame (say, the first six months of a new role or the first year post-merger)? Do you need a pacer whose primary role is to be your operational metronome, or one whose role is more strategic?

Having decided what you need from a pacer, determine which qualities are nonnegotiable. High on your list should be evidence that they will be trustworthy — not just protecting information but speaking the truth, whether you want to hear it or not. This may mean that the pacer you choose is structurally independent from you: They stand neither to gain nor lose by being candid, and you’re confident that they don’t want your job. You must, in short, believe that the pacer’s intention is to help you win, not to pull ahead themselves.

You must also have confidence in the pacer’s competence and credibility, or you’ll be unwilling to follow their suggested pace. That typically requires that the pacer has enough of their own experience to know what the “right pace” looks like in practice, not just theory. They also need the relevant contextual knowledge. Pacing in a vacuum is useless: Your pacer must understand your team’s bandwidth, organizational capacity, and industry pace if they are to help make decisions worth trusting. An ongoing pacer will likely need to modulate speed — slowing things down, speeding things up, or keeping things coasting, as the situation evolves. This means that they need strong emotional intelligence and relationship-building skills befitting someone who can get a good read on how you and others are doing.

Where do the best pacers — those who are trustworthy, competent, and complementary — come from? Our data suggests that the answer varies. Your pacer might be senior to you in position or experience, a peer, or a subordinate. They might be within your chain of command or in a broader advisory role within the organization. They might be only tangentially connected to the organization if they are an executive coach. Here are some of the more common pacer job types and the associated strengths and limitations we observed.

Drawing on our advisory work with leaders, we have found that the right place to look for a pacer depends on both your stage of leadership and the kind of pacing challenge you face.

If you’re a rising leader, start close to the work. Look for a former boss, experienced mentor, board member, customer, or peer who has already operated successfully at the level you’re approaching and has seen you perform under pressure. If you’re starting a new role, you might ask that person to serve as your pacer for the first 90 days, meeting every two weeks to discuss three questions:

• What am I moving too fast on?
• What am I delaying unnecessarily?
• What consequences of my pace might I not be seeing?

The best pacer is not necessarily the most senior person. It is someone experienced enough to recognize the appropriate pace, independent enough to speak candidly, and familiar enough with your context to offer more than generic advice.

If you’re a veteran leader, begin by asking who can still tell you the truth. Direct reports may understand your pace better than anyone, but they may not feel safe challenging you. Consider an independent board member, a trusted peer outside your chain of command, a retired executive, or an external coach. You can also look for clues to help you determine whether you need someone who will slow your urgency, accelerate your action, or help you distinguish between the two, and whether that person is likely to come from within or needs to be an outsider with different experience. Look for these signals in existing data, such as employee engagement comments, exit interviews, missed deadlines, repeated strategy changes, stalled decisions, and initiatives that were launched but never sustained.

Reaping the Benefits of a Pacer

While acting as a CEO, one of us (Rahul) experienced the benefit of having a pacer firsthand. Rahul was accustomed to moving quickly, carrying multiple responsibilities, and treating every promising opportunity as a reason to accelerate. One of his board members became his pacer.

When he brought forward a new initiative, the board member would ask, “What are you willing to stop doing to make room for this?” That question helped him distinguish between urgency that the business genuinely required and urgency that reflected his own operating style. By prompting Rahul to consider organizational capacity, strategic sequencing, and the consequences of acting too soon, the board member became a trusted counterweight. Rahul remained responsible for the decisions but made them with greater awareness of pace and what the organization could realistically absorb.

Over time, the best pacers make themselves less necessary. Their purpose is to strengthen the leader’s ability to sense pace, adjust deliberately, and develop better internal judgment. The relationship succeeds when the leader no longer needs to be continually reminded to slow down or speed up but has learned to recognize the signals and respond before the organization pays the price.

The point of a pacer isn’t to let someone else determine your fate or run the whole race for you. It’s not to eliminate your need for self-awareness or self-discipline, or to shield you from personal growth or accountability. Taken too far, using a pacer can just lock in your own limited adaptability and make you overly dependent on the judgment, trustworthiness, and pacing skills of someone else.

But when used correctly, a pacer can help you perform better by enabling you to manage your own mental and physical energy over time. They can free you up cognitively to focus on fewer but more important decisions and help you preserve and expend energy at a sustainable pace. When the pacer worries about the tactics and the splits, you can win the race.

The digital strategy team at Stakewise (a pseudonym), a New Zealand-based investment platform, watched a new channel — referral traffic from large language models (LLMs) — grow steadily in their 2025 analytics data. By midyear, ChatGPT was generating roughly 700 referral sessions to the site each month, up from only a handful a year earlier. Visitors arriving through ChatGPT also had lower bounce rates and viewed more pages per visit than those arriving from other referral sources. The site was performing strongly by conventional SEO measures, including keyword rankings, impression volume, and organic click-through rates.

But the team could not answer an important question: What were AI systems actually telling customers about their business? Conventional analytics showed the traffic arriving, but not the conversations happening before the click. Finding out meant auditing the answers themselves.

When the Stakewise team began systematically querying ChatGPT with the questions real customers were asking — about fees, fund choices, and how the platform compared with competitors — they found responses that were confident, fluent, and frequently wrong. One widely repeated incorrect claim stated that the platform offered five (not the actual seven) funds in its KiwiSaver product. Fee structures were mischaracterized. A protective investment safeguard was described as restrictive. Features from the investment platform were conflated with a separate product. For a financial services provider, this meant prospective customers were receiving inaccurate information at exactly the moment they were researching retirement savings options.

Two Key Challenges of GEO: Visibility and Accuracy

As generative AI search proliferates, new LLM search algorithms are giving rise to what practitioners now call generative engine optimization (GEO), also referred to as answer engine optimization, or AEO. GenAI search introduces two strategic challenges for leaders. The first is visibility: ensuring that the brand appears when customers ask questions relevant to its offerings. The second is accuracy: ensuring that those systems represent the brand correctly, consistently, and with up-to-date information. A company can win on one count and still lose on the other — a costly outcome, and an increasingly common one.

The SEO-GEO distinction may sound subtle, but it is profound. Without SEO, there is nothing for AI to find. Without GEO, what AI finds may no longer be true. Traditional SEO asked, “Can customers find us? Where do we rank?” GEO adds a new question: “What is AI saying about us in answers we cannot see?”

Organizations with strong SEO foundations are generally better positioned for GEO than those without them, because the content that ranks well in traditional search is also, broadly, the content that AI systems draw on. Strong SEO makes it more likely that an LLM will use a company’s content. But for the comparison and discovery queries that increasingly drive purchase decisions — category comparisons, product alternatives, and “best X for Y” searches — the LLMs may examine many sources. But only a smaller, highly ranked subset — the sources chosen by the platform’s own constantly updated black-box algorithm — are treated as trustworthy and credible, which ultimately shapes the shortlist of results, the comparison set, and the user’s first impression. Many lower-ranked options are thus omitted, so a potential customer may never learn about a company or reach its site.

Visibility: Building On and Beyond SEO

Practitioners with experience in both SEO and GEO describe GEO as the most rigorous version of what modern organic strategy should be: content that is structured, semantically clear, contextually authoritative, and genuinely useful to the person asking the question. As such, GEO adds meaningful new demands.

The distinction between being found and being cited is more than semantic. Being found and ranking well enough to be retrieved by an LLM is necessary, but not sufficient, for visibility. While SEO helps AI find you, GEO shapes AI’s responses. One digital marketing practitioner told us, “We’re now writing for two audiences: the human reader who still values depth, narrative, and context, and the AI assistant. The real challenge is balancing those two so that you don’t lose authenticity while still making sure AI can ‘see’ and surface your expertise.” Content that cannot be cleanly extracted because the claim is buried, the structure is loose, or the evidence is vague doesn’t make it through this second filter, regardless of how well it ranks.

GEO heightens the importance of extractability. Beyond being persuasive or visually appealing, content must also be easy for AI systems to parse. Headings need to communicate substance, not just brand tone. Key answers must appear clearly rather than being buried in long prose. FAQ sections, structured data, and page architecture matter more when systems are synthesizing answers rather than simply listing links.

GEO also expands the arena of authority beyond owned web properties. As one interviewee put it, “AI doesn’t just crawl links. If your brand shows up in discussions on Reddit, Quora, LinkedIn, or niche forums, that builds semantic relevance.” Reputation management, community presence, and external validation now influence whether and how brands are surfaced in AI-generated responses. Appearances in well-ranked “best of” lists, strong review profiles, and credible customer examples can also shape citation outcomes.

Additionally, GEO requires organizations to treat technical access as an operational issue. Many enterprise sites still inadvertently block AI crawlers through legacy robots.txt settings. Structured markup may be missing or inconsistent. Content priorities may be unclear to LLMs processing the site. In these conditions, a brand’s site cannot be accurately read at all. While SEO remains the foundation — elevating the brand into the set of sources AI can draw on — GEO determines whether those sources are clearly understood, credibly cited, and accurately synthesized for customers. To be represented accurately in AI answers, companies need to manage both SEO and GEO.

For many companies, the economics of digital marketing long favored immediacy. Paid search delivered quickly attributable results. Organic search worked over longer horizons, with returns that were harder to trace and easier for skeptical leadership to dismiss. The result was a familiar asymmetry: large paid media budgets alongside organic search teams struggling to justify modest investments in structural content improvement.

The consequences accumulated in the background. Websites became visually polished but semantically thin. Landing pages communicated well to human readers but offered little clarity to machines. Content strategies emphasized branding and campaigns more than extractable answers. Sites grew larger and more incoherent over time, with no systematic commitment to making the organization legible to engines and models.

Generative AI makes these existing vulnerabilities visible. In effect, AI search is exposing the structural quality of a company’s organic foundations. The brands struggling most in AI-generated results are not victims of a new technology so much as they are confronting the accumulated consequences of decisions made years before generative AI arrived.

This is also why the current moment is such a rich opportunity. GEO has created boardroom urgency to support marketing leaders, who until now have generally struggled to secure such support for content audits, schema markup, or technical SEO to advance their AI visibility strategy. The strategic case for a visibility strategy has changed with the introduction of AI, even if much of the underlying work has not.

Accuracy: When AI Gets Your Brand Wrong

Visibility is just one half of the new GEO challenge. Accuracy is the other. The Stakewise team had been seeing LLM-referred traffic in analytics, but until they audited the LLMs’ responses themselves, they had no systematic way of knowing whether the information customers were receiving was correct. Once the team began testing queries, it discovered that AI platforms were drawing on a wide array of outdated and inaccurate sources that had suddenly become commercially consequential. Accuracy, not visibility, it turned out, was the core problem. One stakeholder said, “We can’t rely on the user clicking through to double-check the source, so the AI’s version becomes the truth.”

The company developed a diagnostic process that traced specific AI-generated inaccuracies back to their source documents. Corrections were made at the source level through content updates, structured data, and outreach to third-party publishers.

Most organizations we studied were aware, in the abstract, that AI-generated inaccuracies posed a commercial risk. But very few had operational processes for identifying what AI systems said about them, tracing those claims back to source materials, and correcting root causes. That gap between awareness and operational discipline is where many businesses now sit.

Closing that gap requires a more systematic approach. A disciplined audit process has four steps.

1. Systematic querying. Ask AI platforms the same questions that customers and prospects are likely to ask about your products, category, and competitors. Stakewise, for example, developed 30 prompts with questions spanning those topics. Each prompt was individually tested on ChatGPT in a fresh incognito session with no prior conversation history. Testing was conducted through the web interface because that version of the tool surfaces the source citations that enable individual inaccuracies to be traced back to specific URLs.

2. Source tracing. Identify what content is feeding LLM platforms incorrect or incomplete responses, distinguishing between owned and third-party sources. Stakewise found that AI platforms were reporting inaccuracies drawn from a wide set of sources: third-party directories; personal finance publications and review aggregators pulling from outdated, now-incorrect information; and, in some cases, the company’s own older help documentation. By following the citation trail in ChatGPT’s web responses, the audit traced each inaccuracy to the specific URL cited.

3. Targeted correction. Update the source material itself through content restructuring, structured data, refreshed listings, or third-party corrections. After Stakewise’s audit of the third-party platforms, it pursued corrections of the sources themselves. The process included direct outreach to publishers, with URL-specific correction requests; updates to the company’s online help documentation; and revisions to structured data. Stakewise’s own content that predated a product change was revised or redirected so that AI systems would not treat it as current. Within three months, the corrections had propagated: The KiwiSaver fund count, fee structure, and safeguard descriptions were accurate across tested queries, and the audit had become a repeatable process rather than a one-off fix. Organizations like Stakewise must engage in ongoing efforts to update active content and to revise, archive, or remove outdated pages that could continue to influence AI-generated answers.

4. Ongoing monitoring. Repeat the audit process regularly as LLM behavior and the web ecosystem continue to shift. Stakewise’s follow-up audit three months after its first one used the same methodology, allowing the team to verify which corrections had propagated and which had not. GEO work is fundamentally a recurring discipline without a completion date, because products evolve, sources shift, new third-party content is published, and the algorithms behind the LLM’s responses change continuously. Keeping information current is therefore a managerial responsibility, not merely an editorial task. Content left untouched for too long can remain visible even when it no longer accurately represents the company.

This is where GEO moves from hype to management practice.

New Metrics for Zero-Click Search

Traditional search measurement was built around clicks. The logic was straightforward: Stronger rankings generated traffic, traffic moved through conversion funnels, and conversion outcomes could be tied back, however imperfectly, to organic search efforts.

Generative search disrupts that model. A customer may ask a detailed question, receive an AI-generated recommendation that satisfies their intent, and move on without ever clicking through to the source material. There’s no funnel at all. In that scenario, the brand that appeared in the answer gained influence, and the brand that did not appear in the response lost it, but neither outcome is visible in conventional web analytics.

Attribution remains incomplete across the board. ChatGPT and Perplexity append referral parameters that can identify their traffic, and analytics tools now natively recognize several major assistants. Yet a substantial share of AI-referred sessions (by some estimates, between one- and two-thirds) arrive with the referrer header stripped, particularly from mobile apps and copy-pasted links. These sessions are therefore classified as “direct” or unattributed traffic, regardless of platform.

The deeper data gap sits one level above. For traditional search, Google Search Console reports the actual queries users typed before reaching a site. No equivalent exists for stand-alone conversational LLM platforms. Google has begun surfacing AI Overview and AI Mode impressions in Search Console, but they reflect its own search-integrated results, not those of stand-alone chat products. No major LLM provider publishes what users actually ask. As a result, companies cannot see, even in aggregate, the prompts customers use to research their category, compare their products, or form opinions about their brand.

Every measurement of GEO performance today therefore relies on simulated prompts that practitioners believe customers are likely to ask, tested in controlled conditions, with the understanding that actual user behavior may differ. This does not mean that GEO cannot be measured, only that its measurement must be understood as directional rather than absolute.

Measuring Change Over Time

Three approaches — share of voice, citation patterns, and AI-referred engagement — are useful in combination for measuring visibility and accuracy.

• Share of voice tracks brand visibility through how often and how prominently a brand appears in AI-generated responses to a defined set of queries. In Stakewise’s audit, for example, share of voice across the 30-query set was 6.6%, compared with leading competitors at 12% to 17%. The number is not a populationwide estimate; it is a comparative measure of the brand’s visibility relative to competitors within the tested query set. When repeated over time, it can also reveal changes in competitive standing.
• Citation patterns track which sources AI platforms credit when generating answers. A brand that is rarely cited has an authority problem, regardless of how highly its website ranks in traditional search. The sources AI cites are the sources AI trusts.
• AI-referred engagement measures how the traffic that does arrive from AI platforms behaves: conversion rate, depth of visit, and apparent purchase intent. But because the available data driving AI referrals is uneven, this measure is imperfect. As one digital strategy lead told us, “We are seeing AI users arriving with clearer intent and better product understanding, even amid lower organic search traffic overall. The traffic this metric does capture is therefore often high-intent and worth understanding on its own terms.”

Each of the three measures has limits. LLM responses vary from session to session, even using identical prompts. Query sets are necessarily proxies. Attribution is incomplete. But together, share of voice, citation patterns, and AI-referred engagement can give leaders a clearer picture of how their brand is being represented in AI-mediated discovery.

Companies can begin measuring visibility and accuracy by selecting one commercially important topic area, building a fixed query set around it, and tracking these three measures regularly across the main AI platforms. More important than perfect precision is consistently applying a method that shows whether a brand is becoming more visible, more accurately represented, and more competitively positioned over time.

What Managers Should Do Now

Four actions matter most for leaders getting started in optimizing for GEO.

1. Audit what AI platforms currently say about your brand. Don’t assume that strong owned content means accurate AI answers. Query the major platforms using the questions customers are likely to ask. Record factual errors, omissions, and competitive comparisons. This is the ground truth everything else should be built on.

2. Fix your content for extractability. The most common response to GEO initiatives is to produce more content. In many cases, the more urgent need is to make existing content easier for AI systems to parse and cite. Rewrite vague headings. Surface direct answers more clearly. Strengthen the FAQ structure and schema markup. Confirm that technical access issues aren’t blocking AI crawlers. To paraphrase a digital content manager, transform your content strategy from “search by keywords” to “search for answers that get cited.”

3. Manage offsite authority as actively as you manage your website. AI systems learn from the broader digital ecosystem — reviews, rankings, directories, case studies, and public discourse. Third-party content shapes brand representation at least as much as owned content does, and often more so. Treat external sources where your brand appears as something to actively manage.

4. Measure what you can. Traditional SEO reporting is no longer sufficient on its own. Build a query set that approximates what customers are likely to be asking about your category and competitors, and track share of voice, citation patterns, and AI-referred engagement within it. These measures are proxies, not precise readings, but they reveal change over time. The goal is no longer just to drive traffic; it’s also to influence how the brand is surfaced, compared, and described.

Rather than optimizing everything at once, focus on one high-value topic area where AI visibility matters commercially and then build a repeatable process around it. That fosters in-house capability, sharper measurement, and a more credible basis for scaling.

On zero-click AI platforms, organic visibility is no longer merely being found — it’s being found, cited, and described accurately. For brand strategy, this reframes organic visibility as a question of reputation infrastructure rather than a channel-performance metric. The organizations that build an accurate, authoritative, and consistently updated AI presence are not merely managing search rankings; they are shaping the information environment in which their brand is evaluated, compared, and chosen. In markets where AI increasingly mediates first impressions, that capacity is a durable source of competitive advantage — one that compounds with investment and erodes with neglect, much like brand equity itself.

In 2021, Barcelona-based delivery platform Glovo was faced with a law that threatened its business model. Spain had passed the Rider Law, which introduced a presumption of employment when a digital platform organizes, directs, or controls couriers’ work, including through algorithmic management. Glovo came up with what seemed like a clever fix to avoid having to classify its couriers as employees: emphasizing courier autonomy by redesigning its app so they could log in freely, reject orders without explicit penalties, and choose a daily rate multiplier.

Three years later, however, it was clear that the fix had failed. Glovo and its parent, Delivery Hero, announced that they would hire roughly 15,000 couriers as employees in Spain, incurring a significant financial hit. The workaround had simply delayed the reclassification the company had sought to prevent. During this period, courier protests over lower earnings, along with mounting employment-related penalties and a criminal case against Glovo, increased its exposure.

Glovo reached for a common leadership response to unwanted regulation: minimal compliance. This involves making the narrowest changes a company believes will satisfy a rule while preserving how the business operates. But as algorithms become embedded in how companies deliver services and manage workers, that response is becoming a trap.

Leaders often default to workarounds or temporary fixes because they preserve the existing business model while avoiding a costly redesign. But this leads companies into what we call the algorithmic compliance trap. As Glovo’s case demonstrates, in algorithmic businesses, such fixes expand the audit surface, increase internal complexity, and invite scrutiny that can push the company toward the very redesign it was trying to avoid.

That said, not every regulatory change or instance of enforcement calls for a complete redesign. The hard question for senior leaders is when simple fixes are sufficient or a more significant rethinking of the business model is warranted.

We have developed a practical diagnostic for making that call. Drawing on five cases of algorithmic businesses facing major regulatory challenges in Europe — Glovo, Deliveroo, Airbnb, Uber, and Meta — we identified four signals that can help leaders judge whether workarounds are likely to stabilize a business model or accelerate a costly redesign.

Algorithmic Minimal Compliance: What It Is and Why Leaders Are Drawn to It

When regulation arrives, every business faces the same basic choice: absorb the cost of full compliance or find a way to meet the formal requirements while preserving the existing model. Historically, compliance could be layered onto the organization through policies, training programs, and designated controls while keeping core operations largely intact.

For algorithmic businesses, that separation is harder to maintain. Compliance often has to be built into the product itself: into decision logic, ranking systems, access gates, pricing rules, and automated enforcement. Changes in one part of the system can affect outcomes elsewhere and must continue to work at scale as the product evolves. That can make even genuine, good-faith compliance expensive and disruptive, and it creates a strong pull toward minimal compliance: modifying code, contracts, and workflows just enough to satisfy the formal requirements of a rule while preserving the underlying business model and what the system optimizes.

Algorithmic minimal compliance tends to appear in a few recurring forms:

• Contract terms designed to pass a specific legal test.
• Interface redesigns that signal autonomy while the platform continues to coordinate outcomes through incentives and information.
• Data collection and reporting tools built for recurring obligations.
• “Human review” steps that satisfy oversight requirements without shifting real decision authority.
• Compliance programs that produce reports and process artifacts while the system continues to evolve.

None of these moves is necessarily made in bad faith. When rules are new and enforcement standards are still taking shape, a workaround can keep a business operating while leaders learn what regulators will actually test for.

Algorithms make these workarounds harder to contain for two reasons. First, code changes, contract amendments, and workflow redesigns leave auditable traces that a revised policy memo does not. A workaround intended to satisfy a formal requirement becomes a record and, in adversarial settings, may serve as evidence. Second, workarounds in complex, interdependent systems rarely stay contained. A change in one part of the logic can have ripple effects elsewhere, creating new inconsistencies and new exposures. What begins as a targeted fix can expand the audit surface faster than it closes it.

Four Red Flags That Suggest Minimal Compliance May No Longer Hold

Across the five cases, four signals help explain when minimal compliance remained viable and when it began to compound the problem.

1. Tightening legal clarity. Minimal compliance thrives in gray zones. As laws, rulings, and guidance become more specific, loopholes close and the design space for a workaround shrinks.

2. Increasingly visible harm. When social costs become visible to workers, users, communities, journalists, or public officials, regulators become less willing to accept fixes that change the optics without changing outcomes.

3. Intensifying regulatory scrutiny. Regulators often allow room for experimentation while new rules take shape. As they shift from guidance to information requests, investigations, coordinated enforcement, formal proceedings, and demands for proof, the room to rely on a favorable interpretation narrows quickly.

4. Increasing internal complexity. A workaround may begin as a narrow change. But when compliance touches multiple models, markets, and workflows, fixes begin to cascade. Each adjustment creates new inconsistencies, requiring further fixes and expanding the company’s exposure.

These signals reinforce one another. Tightening legal clarity, visible harm, and intensifying scrutiny raise the bar for acceptable compliance, while growing internal complexity makes that bar harder to meet through additional fixes. As the gap widens, the workaround begins to shape the product road map instead of protecting it, and leaders lose control over the timing and scope of redesign.

Consider these rules of thumb when determining your organization’s next move:

• Zero or one red flag: A workaround may hold. Monitor and reassess as it meets operating reality.
• Two red flags: The window is closing. Start redesign planning.
• Three red flags: Escalation is likely. Fund a durable redesign path now.
• Four red flags: Forced redesign is likely. Treat further workarounds as a cost multiplier and move to redesign.

Now let’s consider how those red flags played out across five cases.

Deliveroo and Glovo: Similar Services, Different Tests

U.K.-based delivery platform Deliveroo provides the clearest contrast to Glovo because, despite being in a similar business, its workaround held whereas Glovo ultimately abandoned its contractor model in Spain.

Deliveroo: A Narrow Fix That Held

In 2016, the Independent Workers’ Union of Great Britain petitioned the U.K.’s Central Arbitration Committee (CAC) for collective bargaining recognition on behalf of Deliveroo couriers. Under the statutory definition applied by the CAC, the dispute turned on “personal service”: couriers could qualify for union recognition only if they were required to perform deliveries themselves.

Shortly before the CAC hearing, Deliveroo introduced new contracts granting couriers a broad right to send substitutes and did not tightly police its use. The CAC examined the terms and actual practice, treated the right as genuine, and rejected the union’s claim. The U.K. Supreme Court later held that the couriers were not in an “employment relationship” for Article 11 trade union rights in that specific context. The decision was limited to the claim before it; it did not settle every employment-status question.

Deliveroo’s minimal-compliance response remained stable because all four red flags remained relatively weak.

• Legal clarity: It was a narrow legal test focused on “personal service,” which Deliveroo could satisfy through a genuine substitution right.
• Harm visibility: The dispute was technical and did not trigger broad public salience.
• Regulatory scrutiny: The CAC accepted the contractual boundary as dispositive for this specific claim.
• Internal complexity: The fix remained contained and did not require extensive reconfiguration of the platform’s core algorithms.

Glovo: When the Red Flags Stack and the Workaround Touches Core Economics

Glovo faced a broader legal question: whether couriers coordinated through its software were genuinely independent. Spain’s Rider Law was applied to examine whether the platform organized, directed, or controlled couriers’ work, including indirectly or implicitly through algorithmic management.

Glovo redesigned several parts of its app to make courier independence more visible while retaining its contractor model. Couriers could log in freely, reject orders without explicit penalties, choose a daily rate multiplier, and exercise wider substitution rights. Performance rankings were removed, and monitoring was reduced. With those levers loosened, the platform relied more heavily on pay parameters, bonuses, and information flows to balance supply and demand.

Glovo’s minimal-compliance response eventually encountered all four red flags.

• Legal clarity: The Rider Law left little room for interpretive maneuvering because it reached mechanisms at the center of Glovo’s operating model. Unlike the CAC’s narrow personal-service test, which Deliveroo could address through one contractual change, the law required Glovo to demonstrate genuine independence across the entire operating relationship. Its product design therefore became evidence of organization, direction, and control.
• Harm visibility: Free log-ins shifted supply-balancing risk to couriers through longer waiting times and greater earnings volatility. The rate multiplier also exposed couriers to price competition. Reports of lower earnings and courier protests made those effects visible.
• Regulatory scrutiny: Authorities found that the platform still organized, directed, and controlled couriers despite the new autonomy features. They treated the redesigned model as continued false self-employment rather than a good-faith adaptation.
• Internal complexity: Glovo loosened access and performance controls but still had to match volatile supply and demand. Service problems prompted further adjustments to pay parameters, bonuses, and incentives. Each correction made Glovo’s continuing authority over courier pay and access easier to trace.

As these signals reinforced one another, Spain’s Labor Inspectorate issued its first sanction against Glovo’s post-Rider Law model, and prosecutors opened a criminal investigation into whether the company had continued to deny couriers employment rights. In December 2024 — one day before CEO Oscar Pierre was due to testify in the criminal case — Glovo announced that it would move to an employment model in Spain. The shift would cover roughly 15,000 couriers, and Delivery Hero projected a 100 million euro (about $113 million) impact on Glovo’s adjusted EBITDA business in Spain for 2025.

The diagnostic may have been able to separate the two cases’ paths before the outcomes were known. Deliveroo remained in the zero or one range: Its change was narrow, genuine, and unlikely to spread through the operating system. Glovo entered implementation with two red flags because the law reached core coordination mechanisms and the response required changes across several parts of the platform. Visible harm and intensifying scrutiny then raised the count to four, well before the employment model announcement.

Leaders do not need perfect foresight. They need to count the warning signs early and count them again as their workaround meets operating reality.

Three Design Choices That Make Compliance More Durable

The Deliveroo-Glovo comparison shows why a rule’s reach and a response’s containability matter most. Three shorter cases show what durable compliance requires under different forms of digital regulation: building recurring obligations into the product, giving human reviewers real authority, and making performance continuously verifiable.

Airbnb: Compliance That Becomes Product Architecture

Some obligations are easier to contain because they are specific, recurring, and separable from core marketplace decisions. The European Union’s Directive on Administrative Cooperation in Taxation (DAC7), in effect since Jan. 1, 2023, requires platforms to collect and verify taxpayer information and report host and transaction data annually. The EU’s short-term rental data regulation, in effect since May 20, 2026, adds recurring registration number and activity data obligations where national registration systems apply.

Airbnb’s DAC7 process turns the obligation into a workflow: Collect taxpayer information, and notify hosts when data is missing. If a host still does not provide the information after being notified, Airbnb freezes payouts until the host complies. The newer regulation extends the pattern: Display and check registration numbers, remove listings when ordered to by authorities, and transmit activity data monthly. Because these controls apply to modular functions — identity verification, listing eligibility, payouts, and reporting — rather than core marketplace logic such as ranking, pricing, and allocation, they limit internal complexity and the audit surface.

Specific, recurring duties can become reusable product capabilities when the requirements are stable enough to be standardized across markets. Airbnb applies the same design logic to the newer short-term rental data regulation, although it is too early to judge the outcome. When a rule reaches allocation, pricing, evaluation, or control, reusable infrastructure can support compliance but cannot resolve the underlying operating issue.

Uber: When ‘Human Oversight’ Is Treated as Symbolic

Uber’s GDPR case shows why a human-review step cannot serve as a procedural workaround when the law tests whether oversight is real. Under the General Data Protection Regulation, people generally have the right not to be subject to decisions based solely on automated processing when those decisions produce legal or similarly significant effects, such as permanently losing access to work through a platform. Four drivers from the U.K. and Portugal challenged permanent account deactivations for suspected fraud. Uber argued that members of its operational risk team had manually reviewed each case. The Amsterdam District Court initially accepted that account, but the drivers appealed.

In 2023, the Amsterdam Court of Appeal reached different conclusions for the four drivers. For three, Uber had not shown how reviewers influenced the decision, what information they considered, or whether they had the competence and authority to change the outcome. The court described their intervention as “not much more than a purely symbolic act.” For the fourth, a personal interview before deactivation was sufficient to establish meaningful human involvement.

Uber did not establish meaningful human involvement for three drivers, and the court ordered it to provide information about the logic, significance, and consequences of the decisions within one month, subject to a daily penalty of 4,000 euros ($4,675) for noncompliance.”

In August 2026, the Dutch data protection authority fined Uber nearly 825 million euros ($951 million), finding that it had made fully automated decisions to deactivate drivers and had not adequately informed them. Uber appealed, and the authority said that the violations had stopped.

Meta: When Compliance Must Keep Working

Meta’s experience under the EU’s Digital Services Act (DSA) illustrates why launching a compliance tool may be insufficient when regulation depends on ongoing external scrutiny. The DSA requires very large platforms to provide eligible researchers with access to public data and to address systemic risks to civic discourse and elections. Meta launched its Content Library and API in late 2023 and subsequently announced that CrowdTangle — which enabled real-time data monitoring by researchers, journalists, and civil society groups — would close in August 2024. Access to its replacement required an application and was initially limited to researchers from qualifying academic and nonprofit institutions. The question was not whether Meta had created a replacement but whether it provided adequate access and real-time functionality.

In April 2024, the European Commission opened formal proceedings against Meta after raising concerns that the company planned to close CrowdTangle without an adequate replacement for real-time civic discourse and election monitoring. Researcher access was one part of a broader investigation. Meta added functionality to the Content Library and API but closed CrowdTangle as planned. In October 2025, the commission preliminarily found that Meta’s access procedures were burdensome and that the data made available to researchers was often incomplete or unreliable. The finding remained preliminary, and the proceeding was still open as of August 2026.

The test was practical: Could the replacement provide adequate access and functionality as Meta’s products and policies changed? Launching a tool establishes a process; it does not prove that the process works. When regulation requires ongoing external scrutiny, durable compliance depends on maintaining timely access to usable, reliable data as the product evolves.

What to Do Next: Five Operating Moves That Preserve Management’s Options

The algorithmic compliance trap’s greatest cost is the loss of choice. The diagnostic tells leaders when to escalate; the five moves below can be taken to change how the company evaluates, governs, and implements compliance before the outcome is known.

1. Start with what the rule actually tests. Translate the legal requirement into a concrete operating question. Deliveroo faced a discrete personal-service test; for Glovo, Spain’s broader test implicated nearly every mechanism the platform used to coordinate delivery. That breadth was an early warning. Map the requirement across contracts, algorithms, incentives, interfaces, workflows, and decision rights. Then ask what would have to change under strict, consistent enforcement. If compliance would weaken a core source of control or economic advantage, scope and fund a redesign alongside any limited response.

2. Stress-test whether the change will stay contained. Before scaling, use pilots and scenario tests to assess the combined effects on allocation, pricing, earnings, service quality, fraud, appeals, and adjacent workflows. At Glovo, changes to courier log-ins, rate multipliers, and rejection rights affected several outcomes and prompted further interventions. Treat new bonuses, exceptions, or controls as evidence that the response is spreading. Escalate it as an operating model decision before local corrections become embedded.

3. Set a stopping rule before approving the workaround. Define in advance what will trigger a redesign: a formal proceeding, a limit on financial exposure, sustained deterioration in pay or service, rising complaints, or repeated compensating interventions. Glovo’s successive interventions show why monitoring is not enough unless the threshold is set in advance. Each correction can look cheaper than redesign while cumulative exposure grows. Set the trigger, decision owner, and transition funding when the workaround is approved.

4. Build the proof that the compliance claim requires. Match evidence to the claim. Deliveroo’s substitution right was accepted in the CAC proceeding because couriers could use it in practice. Uber did not establish meaningful human involvement in the deactivation decisions involving three drivers; reviewers need the information and authority to change outcomes. The proceeding against Meta tests whether qualified researchers can obtain timely, usable, and reliable data. Build that proof alongside the response: Record changes and overrides, track affected outcomes, and rehearse what an external reviewer would ask the company to demonstrate.

5. Turn recurring, separable duties into product capabilities. Airbnb’s DAC7 process shows what reusable compliance infrastructure can include: tax data fields, user notifications, payout controls, and reporting workflows. Build such components for reuse only where requirements are genuinely equivalent, with shared specifications, testing, and version control as rules evolve. Reusable components can reduce local inconsistency when duties are separable from core decision logic. They cannot resolve rules that directly constrain allocation, pricing, evaluation, or control; those require an operating model decision.

Alongside these five operating moves, companies can seek clearer rules or interpretations through consultations, coalitions, standards setting, litigation, and engagement with policymakers. Channels vary by jurisdiction. These efforts should complement operational preparation, not delay it. Once authorities request records, inspect operations, or open proceedings, influence cannot replace a verifiable response. A company can challenge an interpretation while preparing to comply should the challenge fail.

Keep the Choice in Management’s Hands

A limited workaround may be sensible when a requirement is narrow, separable, and verifiable. Although these cases involve platforms, the diagnostic applies wherever connected systems shape consequential decisions — when banks allocate credit, insurers price risk, employers screen applicants, or retailers set prices. As generative AI enters these workflows, leaders must identify the model’s role, trace how its output shaped the decision, document the controls applied, and show how the result can be reviewed or challenged.

The risk begins when a limited response becomes part of the operating model. Glovo changed course after its workaround had spread and exposure had grown. Earlier action preserves a genuine choice among a contained fix, reusable compliance infrastructure, and deeper redesign. Once compensating changes shape core operations, redesign is no longer a future possibility; it is already happening. Management can direct it early or let accumulated workarounds dictate its shape.

It’s a familiar scenario: An organization implements digital tools intended to improve a particular process but fails to obtain the desired result. Employees who are meant to use them complain that the tools don’t solve the problem or that they add inefficiencies. Managers who investigate problems with the tools chalk the failure up to poor adoption but don’t look more closely at what caused the outcome.

When those leaders do seek to analyze failed digital investments, they make a common error: They evaluate data quality and system usability as a single variable. However, those factors operate through entirely different mechanisms, and they fail in different ways; improving one does not improve the other.

The cost of that error is specific and recurring. When a digital investment underdelivers, organizations almost universally attribute the failure to implementation problems, insufficient training, or change management breakdowns. These explanations are structurally inevitable when leaders don’t distinguish between a data quality failure and a system design failure. But without that distinction, the postmortem cannot identify the real cause. The result is a corrective investment that addresses the wrong problem, followed eventually by another underperforming system, followed by another misdiagnosis.

Empirical research that I and others conducted across one of the most cognitively demanding decision environments available for study found that data usability and system usability affect cognitive load through entirely different pathways.1 They aren’t two dimensions of the same problem. They are two distinct problems that require distinct diagnosis and distinct investment. The findings are directly actionable for any organization managing knowledge workers whose performance depends on digital systems.

Along with Kalle Lyytinen and David Aron at Case Western Reserve University and Michael R. Cauley at Vanderbilt University Medical Center, we surveyed 564 practicing physicians across 32 medical specialties to examine how data quality, system design, and information overload jointly shape doctors’ cognitive load when using electronic health records (EHRs) in high-stakes decision-making. We chose clinical medicine because it’s the highest-stakes, most extensively documented context for knowledge work under cognitive pressure and time constraints. The cognitive mechanisms that the research reveals are not unique to medicine, though.

EHR data usability, which encompasses the quality, completeness, and clinical relevance of patient information, increases germane cognitive load. Germane load is the productive cognitive effort associated with deep reasoning and meaningful engagement with complex information. When knowledge workers encounter high-quality, well-organized, contextually relevant data, they engage more deeply with it. That deeper engagement is the mechanism through which good judgments are made. It is not a symptom of overload or fatigue. It is the condition that produces decision quality. Better data makes physicians think harder about what matters. That’s the investment paying off.

We hypothesized that EHR systems that are highly usable — that is, those whose interface design, navigation structure, and workflow alignment support user interaction — reduce extraneous cognitive load. Extraneous load is unproductive cognitive effort generated by poor design: excessive navigation steps, misaligned workflows, alert fatigue, visual clutter, and documentation requirements that consume mental capacity without contributing to the decision at hand. Better system design eliminates that waste and redirects cognitive capacity toward the reasoning that produces accurate judgments. The problem it solves isn’t worker well-being. It addresses decision quality degradation caused by avoidable structural friction.

The research confirmed both effects with statistical precision. Data usability demonstrated a strong direct positive effect on cognitive load, with a standardized path coefficient of 0.597. System usability partially — and negatively — mediated that relationship, with an indirect effect of negative 0.571. Information overload mediated both pathways.

Why Correctly Diagnosing System Failures Matters

Consider what happens when an organization invests heavily in data quality without a proportional investment in system design. Data usability improves. Germane cognitive engagement increases. Knowledge workers are able to reason more deeply about better information. But if the system requires more navigation, more clicks, more workflow friction to access that information, extraneous cognitive load increases simultaneously. The total cognitive burden on the user goes up even as the quality of the underlying data improves. Decision throughput decreases. Users report fatigue and frustration. Leadership concludes that the digital transformation underperformed, without understanding the structural design problem: The investment improved data quality but didn’t include commensurate investment in reducing the friction of accessing and acting on that data.

The reverse failure is equally common. Organizations focus on system usability improvements, simplifying interfaces, reducing clicks, and redesigning dashboards without addressing underlying data quality. Extraneous load decreases. Navigation is easier. But if the data is incomplete, inconsistent, or poorly organized, the germane cognitive engagement that produces good decisions is not triggered. The system is easier to use. The decisions aren’t better.

Both failure modes are predictable once the distinction between them is visible, but neither can be seen when leaders evaluate digital system performance as a single variable.

Consider a financial services firm that invests heavily in a new data platform, consolidating market intelligence, portfolio analytics, and risk signals into a single source of truth. Data quality improves measurably. But the interface through which analysts access that data, built by a different vendor on a different timeline, requires eight navigation steps to surface a complete company profile and generates alerts at a threshold calibrated for compliance, not decision-making. Analyst productivity stalls. Senior talent starts leaving. The postmortem team concludes that the platform was poorly adopted and launches a training program.

The data investment was sound. The system design was not. The postmortem examined neither independently, so it fixed nothing. Had the two levers been assessed separately from the outset, the failure would have been locatable, correctable, and cheap to fix relative to what the misdiagnosis cost.

The Importance of Information Governance to System Design

The finding that higher data usability reduced perceived information overload was not about limiting data volume. It was about improving signal quality: making sure the information in front of a decision maker was relevant, reliable, and organized. Acting on this insight requires attending to the following four governance practices.

Eliminate redundancy. When the same data appears in multiple places within the same system, users stop reasoning from content and start verifying consistency. That verification work produces no decision value. Eliminating redundancy removes it and increases trust in what remains.

Show your data’s lineage. When a decision maker cannot quickly assess whether a data point is current, how it was collected, or how much confidence it warrants, they spend mental capacity on source verification rather than the decision itself. Embedding simple indicators of reliability — when a measurement was taken, by what process, and with what confidence level — eliminates that overhead and lets judgment begin sooner.

Mandate the fields that matter most. Missing information at a critical decision point doesn’t just slow things down; it breaks the workflow entirely and forces improvisation that shouldn’t be necessary. Structuring data capture requirements around the fields most critical to core decisions ensures that gaps appear by exception, not by default.

Mind your alert thresholds. Alert fatigue — the learned tendency to dismiss notifications because most don’t require action — isn’t a user behavior problem. It’s a governance failure. Every alert threshold is a decision about where to direct a decision maker’s attention. When those thresholds are set too low, attention is trained away from the system entirely. Resetting them so alerts fire only when action is genuinely required is among the highest-return cognitive design improvements available, and it requires no new technology.

Do Better at Diagnosing Why AI Systems Underdeliver

Organizations investing heavily in AI-enabled decision support systems across industries are generating the same postmortem pattern at scale. Leaders evaluate AI performance primarily through model accuracy, adoption rates, and efficiency gains. Those metrics are necessary. They are not sufficient.

AI systems that improve the quality and relevance of information surfaced to decision makers are engaging the germane load lever. They are making the data better. That is valuable. But if the interface through which users interact with AI outputs is poorly designed, if recommendations arrive without sufficient context, if the system requires significant navigation to understand the basis for a recommendation, or if it generates notifications users have learned to dismiss, the extraneous load the interface imposes will erode or eliminate the germane load benefit the AI delivers.

Making an AI investment that improves recommendation quality without making an equivalent investment in the interface design required to act on those recommendations will produce the same misdiagnosis as every prior generation of digital underperformance: The tools will be blamed, the training questioned, the workforce examined, and the structural design problem left in place.

Leaders should look for three qualities in any AI system in their portfolio. First, the data it surfaces must be worth the depth of reasoning it demands from the people using it. Second, the interface design should reduce the structural friction of accessing and acting on that data. And third, the information that is surfaced, when it is surfaced, and any accompanying signals as to its urgency, should be designed to direct judgment toward what matters, not scatter attention.

How Leaders Should Evaluate the Success of Digital Systems

The questions that belong in leadership reviews of digital system performance extend beyond standard metrics of uptime, accuracy, and user adoption. They are questions about cognitive design.

Does the system reduce the effort required to find relevant information, or does it increase it? Are the alerts the system generates ones our people act on or ones they have learned to dismiss? Is the cognitive engagement this system demands producing better decisions, or is it creating structural friction that the design should be eliminating? What signals are being surfaced at each decision moment, and are they the right ones for that moment? Are there signs of decision quality degradation — rising error rates on complex decisions, increased escalation volumes, or shortened tenure among experienced decision makers — that point to a poorly designed cognitive environment rather than a talent or motivation problem?

Alert configurations, interface logic, workflow design, and data governance thresholds are cognitive design decisions with direct performance consequences; they are not IT configuration choices. Leaders who want to improve the returns on their digital investments must gain visibility into these decisions, not cede them entirely to implementation teams, vendors, and IT administrators.

In July 2024, the cybersecurity software company CrowdStrike released a defective content configuration update. In the ensuing 78 minutes, before a patch was deployed, IT systems worldwide crashed. Microsoft estimated that 8.5 million Windows devices were affected. As blue screens of death cascaded across the globe, airlines grounded fleets, hospitals lost access to patient records, and banking platforms went dark. One analysis estimated that the collapse cost Fortune 500 companies $5.4 billion.

Six months earlier, a ransomware group had accessed Change Healthcare, the largest medical claims clearinghouse in the United States, through a single portal that lacked multifactor authentication. Within days, nearly every pharmacy, hospital, and physician practice in the U.S. was unable to process insurance claims. Surgeries were postponed, and the personal health data of as many as 1 in 3 Americans was exposed. Combined direct response costs and business disruption impacts incurred by parent company UnitedHealth Group exceeded $2 billion in the first half of 2024 alone, according to the company’s filings with the Securities and Exchange Commission.

Two incidents. Two different failure modes. Combined documented losses exceeding $7 billion. In both cases, crisis experts wondered, had anyone rehearsed this?

They could have. As far back as 2022, McKinsey was reporting that 70% of C-suite technology executives at large enterprises were exploring and investing in digital twins as a way to optimize operations, model supply chains, and accelerate decision-making. A digital twin is a dynamic virtual replica of an organization’s operations, supply chain, manufacturing lines, IT infrastructure, or distribution network, connected in real time to the data flows that govern its physical counterpart. Unlike a static model, a digital twin updates continuously as conditions change, and it can be queried, stressed, or reconfigured without touching the physical system.

Few companies are using this modeling capacity to rehearse crisis scenarios. They’re not using it to test what happens when a software provider’s routine update crashes their operating environment, or when their largest payment processor goes dark. But they could be. The gap between what digital twins are used for and what they could be used for is not a technology gap. It is a strategic gap that has cost organizations billions of dollars and is risking the loss of much more.

Warning Signals From Practitioners

To test whether this gap was visible at the practitioner level, I conducted a structured poll during a session titled “Rehearsal Intelligence: AI Digital Twins for Crisis-Ready Organizations” at ASIS Europe 2026, a Tier 1 international conference drawing risk professionals from across sectors and geographies. Roughly 40 practitioners attended the session; between 19 and 23 responded to each poll question. While I cannot claim that this convenience sample is statistically representative, its composition — exclusively senior security and resilience practitioners with direct organizational visibility — offers a meaningful practitioner-level signal.

Of the 21 respondents who identified their role, 47% were corporate security directors or heads of security. The remainder were in risk management, security consulting, and C-suite functions.

Asked how often their organization conducts crisis simulations, 68% reported once a year, through a tabletop exercise — typically, a facilitated discussion in which a team walks through hypothetical scenarios and planned responses. A further 15% reported never testing the crisis plan at all. Only 5% reported conducting simulations continuously, using live data or digital tools.

The second question produced the starkest finding. Of 22 respondents, 20 (91%) reported that their organization was not using digital twins in any capacity. Two said that their organization used digital twins for operational monitoring. Not a single participant reported using digital twins for crisis simulation.

The third question identified future impediments. Asked what the primary obstacle was to adopting digital twins for crisis management — and, by extension, where challenges might lie in organizations yet to adopt general digital twin technology — 43% said “insufficient leadership awareness and buy-in.” This was the top answer, above budget constraints, technical complexity, and data integration challenges.

Why Traditional Simulations Are No Longer Enough

An executive might reasonably say, “We have crisis teams, we run tabletop exercises, and we have AI. Is the absence of digital-twin-based rehearsals truly a strategic risk?”

To be clear, tabletop exercises run by crisis teams have genuine value. They align work groups, expose assumptions, and create shared mental models. No serious resilience professional would argue against them. But they carry structural limitations that become more consequential as crisis complexity increases.

First, tabletop exercises operate on pre-constructed scenarios. In other words, the exercise is designed around a crisis someone has imagined. CrowdStrike and Change Healthcare were not included in anyone’s tabletop scenarios. The crisis that will actually test your organization is, by definition, the one that was never in the playbook. Second, tabletop exercises are episodic. The muscle memory that organizations build inevitably decays between quarterly or annual sessions. And third, tabletop exercises test the team, not the system. The exercise reveals how particular crisis management team members think under pressure, but it does not reveal how the actual organizational systems, supply chain, IT infrastructure, customer-facing operations, or financial flows might behave under real crisis conditions. Stress-testing examines financial or technical resilience within defined parameters but is not designed for cross-system cascade failure. The gap between what a team decides and what the organization can execute is precisely where most crisis responses break down.

Artificial intelligence adds some value to planning, but less than leaders might expect. In a peer-reviewed study published in March 2026 in International Studies of Management & Organization, researchers Raphaël De Vittoris and Carole Bousquet analyzed 24 crisis simulations conducted between 2017 and 2024 that tested the predictive capacities of human and artificial intelligence across a range of scenarios.

Each simulation was built around nine critical developments that a crisis group should be able to anticipate. Those points were embedded in the scenarios in advance and validated by internal and external crisis professionals, and performance was measured as the share of critical developments that each group involved in the study actually surfaced.

On that measure, AI operating alone identified 41% of the critical points. (The AI models tested were the three most frequently cited by the crisis management team members themselves — the tools they would plausibly reach for under pressure.) Human teams without any support identified 48%. But the use of sophisticated AI platforms barely added value to crisis-anticipation performance: Human teams that used AI identified 49% of the critical points, a gain of a single percentage point. Human teams equipped with structured information access (that is, Google’s search engine) found 81%.

Those results carry a direct management implication. AI investments on their own won’t yield the best solutions. Instead, preparing teams before pressure arrives requires different options.

In their 2025 article “How to Supercharge Your Crisis Training,” MIT’s Sandra Galletti and Steven B. Goldman argue compellingly that organizations need to move beyond passive crisis planning toward active, experiential simulation, and that most organizations fall significantly short of what modern crisis complexity demands. They’re right.

Digital twins offer a way to inject systems with active simulation. Digital twins can stress-test systems against scenarios no one has imagined, run continuously rather than episodically, and be used to test not just team thinking but an organization’s actual capabilities to respond to disruption. Rehearsal intelligence is not a theoretical construct but the organizational capability that emerges when the digital twin is deliberately repurposed from an operational optimization tool into a permanent crisis-anticipation environment. The technology already exists in increasing numbers of large organizations. The gap between using digital twins for operations and digital twins for crisis rehearsal is an opportunity that requires a governance decision that data suggests is long overdue. What’s missing is the framework to deploy it for this purpose — and, as the poll data from the practitioners makes clear, the leadership attention that could make it systematic.

A Five-Step Framework to Deploy Digital Twins for Rehearsal Intelligence

Repurposing an organization’s existing digital twinning tool for a new class of questions rests on a five-step framework of organizational design decisions.

1. Treat the digital twin as a permanent rehearsal environment, not an episodic exercise tool. Crisis simulation should not be an event but rather an ongoing organizational practice. The digital twin should be accessible for scenario stress-testing as routinely as it’s used for operational planning.

Some organizations are already moving in this direction, though without a shared language for what they are doing. BMW planned and validated the construction of a new plant in Debrecen, Hungary, in an entirely virtual environment more than two years before physical vehicle production began. Its digital twin simulates every production change before any physical modification is made, in part to avoid danger and damage. For instance, what BMW calls its “virtual factory” allowed faster insight into collision checks — simulating the movement of a new vehicle design through the production line to make sure it wouldn’t scrape up against anything. Using the digital twin reduced what would typically take almost four weeks to just three days. In June 2025, the company announced that it would be scaling the use of digital twins across 30 of its production facilities worldwide and projected that production planning costs will fall by up to 30%.

Similarly, Walmart has deployed digital twins across many of its stores and distribution centers. Brandon Ballard, group director for real estate at Walmart US, said last year that digital twins have helped the company detect and remediate potential equipment failures up to two weeks before they occur, reducing emergency alerts by 30% and cutting refrigeration maintenance costs by 19%, according to CNBC. In one implementation across 20 stores, a digital twin project proactively identified and addressed 842 potential failures in Walmart’s systems, such as refrigeration, in a six-month period, enabling the company to avoid an estimated $1.4 million in downtime costs.

2. Position AI as a structured sparring partner, not a decision maker. The De Vittoris and Bousquet study offers a warning here. Deploying AI casually added a single percentage point to team performance. The best results came from teams that used tools they had mastered, in a structured and deliberate way. Among the top-scoring crisis cells, those that used AI engaged with it through multiple contextualized prompts, in real conversations, explicitly asking for the anticipations that a team emotionally affected by the event might miss. The worst-scoring team to use AI submitted one short, uncontextualized question. The researchers’ conclusion is that organizations need a genuine prompt culture that includes a library of adaptable prompt templates built before a crisis arrives. That literacy comes from deliberate practice, not from buying an AI platform.

Here’s a suggestion: Build one structured anticipation session into your next crisis team meeting. Give the team a specific disruption scenario and ask them to regard your AI environment as a structured challenger to generate three alternative scenarios, identify second-order consequences of each, and test the assumptions your current crisis plan rests on. This is precisely the structured engagement that separated the best-performing teams in the study from the rest. It’s available to your team today.

3. Embed structured anticipation into organizational culture rather than delegating it to a crisis unit. The performance differential in the De Vittoris and Bousquet study came from anticipation mechanisms that were formalized and actually used when the crisis hit, not from titles or org charts. It’s a matter of organizational design, not staffing.

At Walmart, the digital twin infrastructure that monitors refrigeration temperatures could be queried against disruption scenarios: What happens across 200 stores if the monitoring systems fail simultaneously? What is the cascade if three distribution centers lose connectivity during peak demand? These are questions that general managers, not just risk managers, can ask.

The first step toward building this culture is to run one crisis rehearsal scenario in your digital twin environment. You don’t need a comprehensive program or a multiyear transformation. You just need one scenario, representing a realistic disruption event relevant to your sector, with a structured debrief within 72 hours. The goal is not a perfect simulation. It’s to harness the organizational learning that comes from running one.

4. Build scenario templates and structured question sets that enable crisis teams to engage the digital twin environment rapidly when a scenario emerges. In financial services, digital twins that model trading infrastructure and liquidity flows can be queried against settlement failures, counterparty cascades, or regulatory intervention scenarios. In pharmaceutical manufacturing, production-line twins can be stress-tested against supply disruptions or batch-failure events before they reach patients or regulators.

In energy and utilities, grid topology and distribution twins can simulate cascade failures before they reach physical infrastructure. FedEx has built a digital twin of its global logistics network, which it uses to forecast, identify, and minimize disruption.

The point is, the technology footprint and focus will vary by sector. Repurposing digital twins will call for a new class of questions.

5. Adopt a three-stage maturity model — reactive to structured to anticipatory — that gives leadership a clear diagnostic and a direction of travel. The majority of organizations in my ASIS Europe poll were in the first stage, reactive, with crisis plans and occasional tabletop exercises, while digital twins were either absent or siloed in operational functions. The second stage is structured, where the organization has conducted at least one deliberate crisis-rehearsal scenario using a digital twin environment, with assigned ownership of the rehearsal function, a defined debrief process, and documented findings. The third stage is anticipatory, where crisis rehearsal is continuous and embedded.

Stage 3 has an underlying prerequisite that no governance decision alone can shortcut: the need for quality data feeding the digital twin. Organizations that reach Stage 3 have made the accuracy, accessibility, and maintenance of their live data environment a leadership priority, not just a technology department responsibility.

For most organizations, moving from the first to the second stage will mean finding out whether currently deployed digital twins within specific functional domains can be accessed, combined, and stress-tested by a crisis anticipation team. This does not require a technology project. It requires that you assign one person to audit which digital twin environments your organization currently operates and then have a two-hour conversation with your technology leadership.

The organizational capability that CrowdStrike and Change Healthcare revealed to be absent in 2024 has a name: rehearsal intelligence. Although it is urgently needed and the technology to build it is being adopted by more and more large enterprises, most organizations don’t have it in any systematic form.

Organizations can realistically aim to get to Stage 3, where they are in a continual state of anticipation, and crisis rehearsal is ongoing and entrenched. The return extends beyond crisis avoidance. Organizations operating at Stage 3 report faster decision-making (because the decision architecture has been tested against disruption before a real crisis arrives) and more precise capital allocation (because leadership understands which system vulnerabilities carry the greatest operational risk). They have stronger AI utilization because teams engage with AI as a structured reasoning partner rather than a retrieval tool, and because they have a measurable recovery-speed advantage over competitors encountering the same disruption for the first time.

As Jürgen Wittmann, head of innovation, virtual factory, and virtual commissioning in BMW’s production department, told CIO magazine last year, “Thanks to the digital twin, we know exactly the current situation and can immediately see the impact of changes.” Why shouldn’t all companies use that same tool for a wider look at the horizon?

The distance between where most organizations are today and where rehearsal intelligence can take them is not measured in technology investment or budget cycles. It’s measured in the decisions made, or deferred, right now. The CrowdStrike and Change Healthcare outages didn’t announce themselves in advance. The next disruption won’t either.

Businesses working to manage their spending on energy were dealt a blow in late February following military action by the U.S. and Israel against Iran. A predictable consequence of the conflict was a dramatic increase in the price of crude oil, the primary benchmark for global energy costs. The price of a barrel of crude oil shot up from $67 in early March to $112 in early April; by early June, it had come down to $80, still a roughly 25% year-over-year price increase.

Geopolitical conflict can bring exceptional volatility to energy costs, but steadily rising expenditures have been the underlying trend for companies since 2022, putting pressure on their profit margins. Energy is a central input in energy-intensive industries such as cement, chemicals, and metals, which account for roughly three-quarters of industrial energy demand. Companies that rely on global supply chains, and even companies that consume relatively less energy in their operations, like food processors, are feeling the effects of rising fuel and power costs.

The pressure to reduce costs is likely to drive more corporate efforts to improve energy efficiency. While that’s good for sustainability, our research suggests that energy efficiency can also deliver measurable cost savings, enhance resilience, and fuel growth. Attention to energy efficiency can lower per-unit manufacturing costs, reduce exposure to energy-price volatility, and sustain margins during energy shocks. The companies we studied have cut their energy use at some facilities by up to 50% and reported efficiency gains of 10% to 20% in their core industrial processes.

Upgrades to digital systems, manufacturing equipment, and facilities all played a role in those outcomes. The key to success, however, came not from any upgrade made in isolation but from having a comprehensive strategy that linked improvements together. Company executives made energy efficiency a clear priority, with defined accountability and performance targets.

From our surveys of Swiss manufacturing companies and interviews with executives across the globe, we identified three ways in which the most energy efficient among them manage their energy costs and take advantage of digital capabilities in that work.

First, they invest in collecting and analyzing data about energy use from everywhere in the company and its ecosystem and making that data visible to managers. Using sensors, dashboards, and centralized data platforms, they enable real-time energy management and cross-site benchmarking. Second, they redesign their operations and systems to optimize energy use, eliminate waste, and reduce dependence on energy-intensive inputs. By embedding advanced analytics and AI into daily workflows, they can improve continuously. And third, they are more likely to create products and services that support their customers’ energy efficiency goals.

Here, we’ll explore energy efficiency practices among the companies we studied and provide insight into how managers can adopt similar approaches.

Make Energy Data Visible Everywhere

Enterprise-level energy efficiency becomes feasible when the energy that business operations consume is visible in near real time. In our survey of Swiss companies, those in the top quartile on energy efficiency rated themselves more highly, on average, than those in the bottom quartile on integrating data collection, analytics, and reporting technologies across their operations. Executives we interviewed at those and other companies reported that they use those capabilities to systematically and continually track and analyze information on energy usage and share it with internal and external stakeholders. Specific practices included translating data into actionable insights for operations teams, frequently sharing energy metrics with senior leadership, and communicating successes to customers, investors, and regulators.

At Swiss specialty chemicals manufacturer Clariant, digital technologies are central to measuring and sharing energy usage data. One of its early initiatives was to install sensors to capture data on energy used by plant machinery. It also invested in operational dashboards for its sites to consolidate their data and uncover opportunities for efficiency improvements. The sustainable operations team now tracks high-frequency data from 80 manufacturing sites and stores it in a central data repository.

Increased visibility has helped to improve energy efficiency on several levels. Plant managers and site operations teams use dashboards to monitor energy consumption, emissions, and operational efficiency across Clariant’s production facilities so that they can identify potential improvements. Business and regional operations leaders use dashboards for cross-site benchmarking, revealing where they may be lagging behind comparable operations. The sustainable operations team monitors energy usage across plants and works with managers at each site to determine how to reduce their energy costs.

Clariant’s savings have been significant. At one plant in China, the energy used to produce finished goods was cut by about half from 2019 to 2021, and even further in subsequent years.

The company’s generative AI-based Clarita platform, introduced in 2023, is used across production sites to monitor energy consumption, identify inefficiencies, and recommend actions to optimize processes while continuously learning from operational data. At a production plant in Germany, local operations and energy efficiency teams used Clarita to identify abnormal steam consumption in a heating system by comparing real-time usage with historical baselines. The results led them to adjust operating parameters, which significantly reduced steam demand. Additionally, a mining site in Indonesia used the system to save energy by improving routines for scheduling and maintaining its equipment. (See “Can AI Save Energy?”)

Cut Energy Waste

Despite having improved visibility into energy use, many companies we studied still struggled to reduce their energy costs because they did not use the insights from their data to change their production processes or to manage raw materials and fuels differently.

Those reporting the most action on energy efficiency in our sample closed this gap by embedding energy and resource management into routine operations and decision processes. For these companies, energy efficiency became part of daily performance tracking, with energy use measured and managed alongside productivity and cost metrics rather than treated as a separate sustainability concern. Then they deployed new technologies, such as equipment that manages its own energy use, along with heating and cooling systems that employed waste-heat recovery technology. They also redesigned manufacturing processes to use fewer materials and less energy.

Some companies are also moving to replace fossil fuel inputs (and their associated price volatility) with electricity and renewable sources, which are often more efficient.

Take, for example, Swiss industrial group Georg Fischer (GF), which makes equipment for transporting liquids and gases. The company has focused on designing operations to cut energy waste as part of a broader energy efficiency strategy, said Oliver Hilbrand, a plant manager. “Our goal was to significantly increase efficiency in all areas — from production to logistics to energy use,” he said.

At the Seewis, Switzerland, site where GF manufactures valves and actuators, it has deployed a combination of energy-saving technologies, including higher-efficiency machinery, smart lighting, waste-heat recovery, and better insulation. Additionally, the site gets 100% of its electricity from renewable sources. As of 2023, the plant had reduced its emissions from fossil fuels by about 63% compared with a 2019 baseline. While not a direct measure of fossil fuel use, a reduction in emissions indicates that a company is less dependent on it — and less exposed to fluctuating fossil fuel prices.

Meanwhile, GF is shifting to renewable electricity sources companywide. It is installing solar panels on factory roofs and buying certified renewable electricity from regional hydro and wind sources. These steps have not only reduced GF’s dependence on grid power generated by fossil fuels; they have also made it easier to predict future energy costs. Although renewable energy generation fluctuates across days and seasons, the cost per kilowatt hour of solar power is less volatile than that of oil and can be managed through local measures, such as storing solar power in batteries for later use during peak-rate hours. By 2025, GF was sourcing 62% of its total electricity from renewable sources.

Further, the company has applied circular-economy principles to reduce its consumption of energy-intensive raw materials, such as plastic and steel. If it can maintain quality and comply with regulations, the company mixes its scrap with raw materials to reduce raw material inputs. In 2025, it reported that it was recycling 68% of its waste, including scrap from production.

In addition, the company conducts life-cycle assessments to evaluate its energy footprint across the entire value chain, including raw material processing, product manufacturing, transportation, product use, and disposal. It discloses that data through Environmental Product Declarations verified by third parties, enabling it to benchmark against global standards and identify targeted opportunities for further improvements.

Create Energy-Smart Offerings

Among the companies we studied, we saw a third area of focus in energy efficiency emerging: developing products that use fewer resources and enable customers to reduce their own energy consumption. Those reporting more activity in this area were significantly more likely to be partnering with other companies to reduce energy use and environmental impact.

By offering more energy efficient products, companies can create more value for customers. For example, when GF customers upgrade their equipment, the company is able to quantify energy savings and other sustainability benefits they can expect to achieve. Meanwhile, ABB, a global provider of electrification and automation systems, has partnered with E.ON, an infrastructure services company, to offer energy efficiency services, such as appraisals, system design, and financing, with guaranteed savings. Those offerings reduce upfront barriers to investing in new equipment and shift performance risk away from the customers.

Companies can also capture value indirectly by embedding energy efficiency into their products, providing customers with self-service data, and using their strong internal energy efficiency performance to attract and retain customers.

ABB has developed a portfolio of digital tools that help customers manage their energy use more effectively. Analysis of the company’s installed base showed recurring inefficiencies in motors, pumps, and process equipment, along with limited visibility into the energy that equipment consumed. In response, ABB introduced a service that combines monitoring, analytics, and AI to identify inefficiencies and recommend actions. These systems connect equipment, track energy flows in real time, and coordinate operations across assets, enabling continuous adjustment. The company reported that its cement industry customers have achieved efficiency gains of 15%-18%. Data center customers are up to 25% more efficient. In addition, these tools enable customers to troubleshoot anomalies in their energy consumption 60%-80% faster.

Similarly, DMG Mori, a global manufacturer of computer-controlled machine tools such as milling and turning machines, developed its Greenmode offering after analyzing how energy was being consumed during machine operation at customer sites. The company observed that customers’ energy consumption was driven not only by machining but also by auxiliary systems, such as machine cooling, feed drives, and compressed air, and by prolonged idle times. Further, customers had limited real-time transparency into how much energy their machine tools used.

To address these issues, the company built features into its machines that help customers see their electricity consumption in real time and operate machine components according to when they are needed during production instead of running them continuously. DMG Mori data that has been certified by TÜV Süd, an independent testing and certification organization, shows that these measures have reduced machine energy consumption at customer sites by more than 30%, on average, and up to 40% compared with earlier-generation machines.

Our research points to a reinforcing cycle: Companies that make energy consumption visible across the enterprise gain the insight to redesign operations, and the savings from redesigned operations can be used to fund products and services that help customers do the same. Those that have made an organizational commitment to treat energy efficiency as a business priority set performance goals and manage for them. As long as fossil fuel costs remain volatile, companies that can continuously improve their energy efficiency will have the advantage over those that are forced to react to — and absorb — every shock.

The in-house lawyers at real estate investment firm Alturas Capital Partners used to rely on outside counsel for much of its lease work. Thanks to generative AI, it can now do that work internally, saving the firm hundreds of thousands of dollars in spending and compressing lease work that once stalled for weeks. But for the outside counsel that lost the work, the bigger problem is existential: AI has eliminated the need for their services. This is redefining the boundaries of firms, since some firms are expanding the scope of their work, while others are losing it to their former customers. This effect is also affecting the prices that the remaining providers can justify. Just recently, Wall Street banks pushed big law firms to cut fees because of AI.

This scenario is widespread: Generative AI has lowered the cost of producing legal documents, market analyses, creative assets, and software in situations where a capable in-house team equipped with AI can credibly replicate what an outside provider had been supplying. Service providers must find ways to ensure that there is still a need for their work.

Services firms have traditionally built their value propositions around the specialized expertise and experience of their staff, as well as their proprietary methods. Generative AI has changed that by providing access to a fair degree of knowledge that was once the sole province of human experts.

This is not the first time technology has changed the economics of what companies handle in-house and what work they outsource. In the 1990s, technological advancements in communications and software enabled the expansion of services as it became feasible — and cheaper — to outsource back-office work.

Calculating Customer Costs

Service providers must rethink existing sales strategies that are based on expertise alone, now that generative AI is changing the economics of the decision to outsource higher-level knowledge work and making it cheaper to bring it in-house. They need to focus on how to win based on cost as well. Here are three ways they can do so.

Improve unit economics. Services firms’ first move should target production cost. The instinct for providers under threat is to defend their expertise — to argue that their people produce better work than a client’s in-house team armed with AI tools. That defense is collapsing. The durable advantage lies not in the skill behind each output but in the economics of producing thousands of them. Providers that build the infrastructure to produce work at scale — automated pipelines, open models, costs spread across many clients — can price each finished output below what any single client could match by doing the work in-house. Cheap production pulled the work inside; cheaper production can pull it back out.

One of the world’s largest marketing services groups, WPP, offers an example of how a traditional services firm can win this way. It built WPP Open, an agentic marketing platform that draws on decades of proprietary intelligence, including 30 years of data from the world’s longest-running brand equity study, and behavioral science frameworks from its agency Ogilvy. Coca-Cola is among the brands using it. WPP has also built a self-serve offering through which marketers build strategies, generate assets, and activate media campaigns. While a competitor could acquire the same technology, it would still lack the accumulated expertise that powers the agents — expertise that takes decades to build.

This requires a fundamental shift in how professional services firms think about their business, to service productization rather than bespoke delivery. Generative AI has made that shift both more urgent and more achievable. Providers that have already begun the journey — by standardizing workflows, training models on proprietary data, and packaging expertise into repeatable systems — are best placed to win on unit economics. As one analysis of professional services firms argued, the service providers that thrive will be those that shift from delivering their expertise through people alone to delivering it through people and systems together.

Make buying as easy as asking. The second move targets a different cost: the effort of buying itself. Hiring a provider has never been free of effort. The customer has to find the right expert, negotiate terms, explain its needs, review drafts, and integrate the result into its own systems. Every hour spent on that strengthens the argument for doing the work in-house. Agentic AI lets providers eliminate that friction, with agents that take the request, do the work, and deliver the answer directly into the tools the customer is already using.

Moody’s shows this advantage at work. The credit intelligence firm watched its own customers pick up generative AI and saw that they would soon be able to answer their own credit questions. Its response was to build AI agents that run analyses automatically and deliver the results within the Microsoft applications customers already use, such as Excel. A portfolio manager checking a counterparty’s credit risk now gets Moody’s ratings, data, and research right in the spreadsheet via Microsoft Copilot, the AI tool they would have used to do Moody’s work themselves. The answer arrives in moments.

Law firms are taking the same path. A&O Shearman, for example, has built AI agents — developed with legal AI firm Harvey — that distill the reasoning of its senior lawyers for tasks like antitrust filing analysis and reviewing loan documentation, and makes them available to clients and other firms by subscription.

When buying from a provider becomes as easy as asking, building an alternative in-house stops looking like a worthwhile investment.

Own the operational burden. The third move targets the cost that clients encounter last: quality assurance and maintenance. Bringing work in-house with AI looks easy at first, but someone has to check every output before it can be trusted, fix mistakes, rewrite prompts and workflows as models change, and clear outputs through compliance. The costs accumulate quietly — in staff hours, in rework, in constant maintenance — until they rival the fee the client was paying the provider in the first place. The service provider’s move is to make that burden visible to the customer and be able to carry it for them.

Thomson Reuters illustrates how a traditional knowledge provider can win on this ground. The company’s CoCounsel Legal is an AI research product built on its Westlaw legal platform, which contains decades of curated case law, 35 million legal classifications, and the ongoing work of more than 650 attorney-editors. A firm building its own AI legal research tool would have neither the content infrastructure nor the verifiable citations and audit trails that courts and compliance functions require.

What Service Providers Under Threat Can Do Now

Service providers survive when they win the client’s full make-or-buy comparison: cheaper to run, easier to buy, and less painful to manage. Here’s how to get started on each of the three moves.

1. List which of your deliverables generative AI can already produce just as well as a skilled person, because those are the ones clients will pull in-house first. For each, compare what one output costs you, all in, against what the client would pay in tools, tokens, and staff time to produce it internally. Wherever your numbers win, package the system and sell it.
2. Count the steps and the days between a customer asking you for something and getting a usable answer. Every step is a reason to build instead. Cut those steps to deliver your expertise as close to the customer’s use case as possible. Agentic AI can help here.
3. For every customer considering building in-house, write out what running the work would actually cost their business: the hours checking outputs, the rework, the model updates, the compliance reviews, and the salaries behind it all. Put that total next to your fee. If the customer insources anyway, stay close and reopen the conversation two quarters later, once those costs have appeared in its books.

Generative AI has not made expertise worthless. It has raised the bar for what providers must offer alongside it. The firms that thrive will be those that make buying cheaper, easier, and less burdensome than building in-house.

For years, social media companies have come under fire for promoting divisive, false, and dangerous content in order to monetize user engagement. But in March, when a jury in Los Angeles Superior Court found Meta and Google liable for causing harm to users due to the addictive nature of their products, the verdict put the focus not on content but on the companies’ deliberate design choices.

What stands out to us in considering the implications of the court decision is that even absent a willfully exploitative business ethos, other companies may unwittingly create the potential for harm simply by applying user-centered design methods such as design thinking. These methods prioritize giving users what it seems that they want. They involve engaging with target users to understand their wants and needs and having them test and offer feedback on prototype products. Our research found that such methodologies, used in good faith, can nonetheless lead practitioners to ignore potential harms or give inadequate attention to ethical considerations.

To better understand how such situations arise, we interviewed 27 senior managers and leaders at well-known technology, pharmaceutical, and design consultancy firms based in the U.S. and Europe. We asked them about the ethical challenges they have encountered when applying design thinking and other user-centered approaches, as well as the governance mechanisms they have used — or wished they had used — to mitigate unintended harm.1

We heard about many moments of reckoning. Features that increased user engagement were described as leading to addictive user behaviors, as in the Meta case. Over time, personalization contributed to the amplification of extreme and polarizing content. In each case, the teams had followed user-centered best practices, but unintended consequences repeatedly emerged. Such experiences raise a fundamental issue: How can we better understand and manage the blind spots and ethical challenges that user-centered design may create?

Methods like design thinking have gained popularity as a creative way to identify and address user needs.2 Applied by companies such as IBM, Meta, PepsiCo, and SAP, design thinking is a problem-solving and innovation approach founded on several principles, including collaboration across functions, iteration and experimentation, and the adoption of an empathetic, user-centered orientation.

While there are plenty of examples of design thinking as an effective methodology for developing products that appeal to users, we’ve also seen cases where its use simultaneously resulted in harm to users and other stakeholders. For example, Juul’s electronic cigarettes emerged from the work of two Stanford University product design graduate students, who used design thinking to create more appealing products than those already on the market. While the product was certainly appealing, achieving 70% of e-cigarette market share in the U.S. in the mid-2010s, it had to pay over $450 million to settle legal claims over product harm and false marketing in 2023. A big part of the problem was that efforts to improve the user experience by introducing nicotine delivery pods in fruity flavors appealed to teenage users, making Juul the dominant brand for vaping among high school students. The failure to consider their choices’ potential harms as the business scaled belongs to Juul’s leaders, but the ease with which design thinking practice can overlook such concerns must be noted.

Overall, a clear ethical framework underpinning the application of design thinking and other user-centered approaches is lacking. Our interviews suggest that this gap is systemic. It stems from recurring blind spots in how user-centered innovation is practiced.

When User-Centered Becomes User-Blind

While our interviewees acknowledged that understanding users’ needs is important, they also highlighted four issues that a strong focus on users may create.

1. Leaving the environment out of frame. Even though the natural environment was clearly affected by the design thinking implementations we considered, none of the firms we engaged with explicitly considered the environment as a stakeholder. And when someone raises such considerations, design teams aren’t necessarily responsive. The director of user experience design at a global smartphone manufacturer recalled the team’s negative reaction when, during a design session, the head of environmental policy asked whether the company really needed to release a new device every year, given how many old phones went into landfills despite the company’s recycling program.

Such issues are problematic because the creation, production, and delivery of various companies’ products and services have evident environmental impacts, including the use of rare mined materials in consumer electronics, the disposal of hazardous materials in pharmaceuticals, and the energy consumed to create digital products. In all these instances, design decisions may degrade a healthy natural environment that humans depend on. Most companies appeared to treat these issues as externalities rather than risks to be proactively managed.

2. Overlooking stakeholders who aren’t the primary target users. Concentrating on the target users often restricts product developers’ attention to specific personas, leaving out other relevant stakeholders. These may include vulnerable users, nonusers who could be indirectly affected by the product, and potential malicious users. For example, the CEO of a design services firm described a location-sharing feature in a dating app that was intended to improve matching and enhance user safety. While the design team saw it as a value-adding innovation, many women interviewed during testing compared the feature to stalking. “For features that track users, you need to design with creepers in mind,” the CEO pointed out. In this case, the team had designed for well-intentioned users, not for malicious ones or those more exposed to risk.

3. Sacrificing scrutiny for speed. Like other innovation approaches, such as agile methods and Lean Startup, design thinking entails rapid experimentation with users to develop new offerings. While quick prototyping and pressure to deliver results can be useful in certain settings, speed can reduce consideration of potential harm, especially for more vulnerable users, such as children. A senior director at a design services firm reflected on whether more time could be spent in developing and iterating prototypes with users: “In a perfect world, definitely yes. But we live and work in an imperfect world with tight project deadlines and resources.”

4. Exposing the company to risks through scaling. We found many examples of iterations conducted with a small set of users that produced dysfunctional outcomes when the products scaled and reached more diverse populations of customers, particularly vulnerable users, such as children. There are two reasons for this. First, small-scale tests are not always the best approach for unearthing a wide variety of use cases, especially malicious use cases, which often emerge as user diversity increases. Second, several executives we interviewed described how so-called edge cases were treated as statistically insignificant during development. As one expert noted, even a small percentage can become consequential at scale: “In Facebook’s case, the 1% is 1% of 2 billion users. That’s approximately 20 million people. People. Not widgets.” Features developed around a narrow user profile can behave very differently when used by millions of people with varied motivations, vulnerabilities, and social contexts.

Delivering Innovation While Minimizing Ethical Risks

Effective leadership and deliberate governance are required to address the limitations we’ve identified in the design thinking process. Senior managers need to understand that user-centered approaches can have ethical blind spots, and that they must take an active role in supervising their use. There are three elements in doing this.

Start with purpose, not just user needs and desires. By purpose, we mean both financial objectives and the company’s mission and values, especially as they relate to society and the environment. Potential consequences for multiple stakeholders should be discussed from the beginning, including secondary and indirectly affected stakeholders. However, there are trade-offs to be managed here. In principle, broad inclusion improves foresight, surfaces edge cases, and reduces downstream harm. In practice, it is neither feasible nor strategically prudent to convene a “committee of everyone,” as this can slow development cycles, create decision paralysis, expose confidential information, and dilute competitive advantage. The challenge, therefore, is not maximal inclusion but structured inclusion.

Three key questions should be asked at the outset:

• What broader value are we trying to create beyond addressing direct users’ needs?
• What trade-offs are unacceptable, even if engagement or revenue increases?
• Who benefits, and who might bear unintended costs?

Attend to the larger system within which design takes place. In our study, we found that processes and structures can greatly enable or limit the agency of design thinking teams. For example, in several companies, KPIs, performance targets, and rewards that prioritized time to market constrained design teams’ ability to consider diverse stakeholders and potential harms. Similarly, business models optimized for maximizing online engagement hindered design thinking teams’ ability to consider downstream harms, particularly among young people. To counter such issues, ethical considerations should be built into strategic discussions — about the business model, competitive strategy, and key objectives, for example. Decisions about user interfaces and product features must not be seen as purely operational. Leaders must promote psychological safety for designers and developers so that they feel free to speak up about potential product harms.

When considering the broad context, ask these three questions:

• What behaviors do our KPIs and incentives reward?
• What harms are invisible in our dashboards?
• Can teams raise ethical concerns without penalty?

Pause before scaling. Before a product launch, assess the full spectrum of benefits and harms across multiple stakeholders and then decide whether to proceed. In addition to reviewing the product against its initial purpose, leaders should evaluate it against ethical criteria: Does it respect relevant rights and duties? What are its most consequential downstream effects? Does it nurture context-sensitive relationships and responsibilities?

In our study, several experts noted how difficult it is to anticipate unintended consequences before a product reaches scale. AI tools can assist by simulating edge cases, modeling malicious-use scenarios, and stress-testing assumptions. For example, a team developing a content recommendation feed could, before product launch, generate synthetic user profiles across age groups and other risk factors and stress-test whether the system disproportionately steers some users toward content that is extremist or is disinformation, for example. If it does, the team could require mitigation (such as friction, downranking, or guardrails) as a condition of launch.

Discussions of scaling risk might consider the following questions:

• What happens if a relatively small percentage of users misapplies this feature at scale?
• Have we stress-tested edge cases and categories of vulnerable users?
• Who signs off on scaling risk?

As products, services, and algorithms become more sophisticated and more complex at scale — reaching more people faster — the cost of getting things wrong shifts from inconvenience to measurable harm and legal liability. A more exacting approach may add time and cost upfront but could reduce downstream consequences such as reputational damage, product recalls, in-market fixes, and litigation. The question is not whether leaders can afford diligence but whether they can afford to postpone it. In some industries, those risks are no longer hypothetical — they’re already being tested in court.

In an unpredictable economy and fractured media landscape, marketing leaders are navigating a period of profound transformation and disruption — and as AI technologies evolve, the pace of change will only increase. In response, the highest priorities of chief marketing officers today are shifting, and understanding their concerns is essential to addressing the most pressing growth challenges companies face.

Marketing is at an inflection point. The structures, talent models, measurement frameworks, and influence mechanisms that will define the function for the next decade are being established right now. CMOs who move with intention — building the function, proving its value, leading effectively through disruption — will define what marketing becomes. Those who wait may find that it has already been defined for them.

Recently, I conducted 15 in-depth interviews with global CMOs from a range of industries to identify their most urgent priorities and challenges. I heard five main concerns:

• Embedding marketing as an executive-level growth function
• Transforming with and for artificial intelligence
• Hiring to defuse the talent time bomb
• Reaching consumers across a fragmented media system
• Developing new metrics for zero-click AI platforms

Compared with other business functions, marketing as a whole, along with even the most granular of measurement tasks, is perhaps being transformed most deeply by the rise of AI. While some sectors are establishing new AI workflows to speed up whatever they already do, marketing is working to harness AI not just as an efficiency tool but as a fundamental reinvention of the function.

In some ways, its very task has changed. Zero-click platforms have upended the meaning of what was, until recently, the web’s most basic metric: clicks. Meanwhile, the media landscape is growing ever more fractionalized, rewriting the rules of attention as consumers encounter individualized feeds algorithmically tailored to them. Function leaders face the challenge of developing talent for a future where the skills that mattered yesterday are being rapidly displaced. An additional challenge is to elevate the role and importance of marketing as a company’s growth engine, which will require tomorrow’s CMOs to develop new skills, lead an effective organization, transform the work, and create more pronounced value. This isn’t about learning new tools; it’s about designing more effective work in a more effective organization.

Together, these five priorities describe a function at a critical point — where marketing leaders are clear-eyed about the challenges ahead and increasingly convinced that marketing’s moment to shape enterprise strategy has arrived.

Evolving From Marketing to Growth

CMOs who operate as marketing specialists rather than general managers are often excluded from upstream product and business decisions — a marginalization that keeps them from delivering the impact that marketing should have. “If marketing isn’t involved when a company is identifying new products and services and isn’t helping bring insights to the table about what customers and future customers want or don’t want, then you end up getting a product that you can’t sell,” Amy Martin Ziegenfuss, CMO of Six Flags Entertainment Corporation, said.

CMOs must be part of higher-level conversations to drive their companies’ success and increase their own influence, and to shift the perception of marketing from a cost center to a growth driver. Julie Nestor, executive vice president of marketing and communications for Mastercard Asia Pacific, described a previous employer’s “investment optimization tool that calculated the ROI for every marketing request before it was approved. It removed subjectivity entirely and created a shared language between marketing and finance. Marketing was no longer asking for [a] budget; it was presenting a business case with a projected return.”

As John Costello, a former board member of Ring Inc. and president of global marketing and innovation at Dunkin’ Brands said, “Changing the title from chief marketing officer to chief growth officer would help all the constituents understand what the role is actually for.”

Implementing AI or Altering the Function?

Nearly every CMO holds a dual view on AI: enthusiasm about its potential, and genuine anxiety about their organization’s ability to capitalize on it. Experimentation is ongoing, but effective scaling of AI has so far remained elusive. CMOs pinpoint two primary internal challenges of AI implementation: identifying the tools that can drive efficiency, speed content production, and improve personalization; and getting teams to adopt them.

Sara Mendez, formerly CMO of SC Johnson Lifestyle Brands, said that marketing is the function to be “most transformed by AI, because it is the most complex, the most connected to every other part of the organization, and AI is simultaneously transforming every individual discipline within it.”

Another CMO mapped over 700 marketing tasks and found that more than 80% could be AI-led. But the limits of what AI can do are unknown and constantly changing as AI tools evolve. A financial services CMO put it bluntly: “Execution, machines can do. Creative thinking — yet to be proven.”

The implication is not just greater efficiency but a full reinvention of how marketing functions are organized, staffed, and run. One CMO reduced a 50-step campaign process to nine steps by mapping AI across her organization’s workflows and rewriting every job description accordingly, yet she still pondered what the “right” organizational design would be in 2030.

Transforming Talent Priorities

The skills AI demands don’t yet match the pipeline — a talent time bomb spanning industries and professions. One CMO asked, “Who am I going to hire on my marketing team — HTML coders, content writers, creative designers, copywriters, media planners? Or do I need people who know how to think?”

Teresa Barreira, global CMO of Publicis Sapient, noted, “When we mapped what only humans could do, we landed on soft skills: plasticity, storytelling, being a great spotter of things. Then the question is, how do you teach young people that?”

CMOs in particular are caught in a three-way bind as they confront skill obsolescence, as the channel-specialist model is displaced by AI and demand shifts toward judgment and critical thinking; pipeline anxiety, as execution-level work is commoditized and marketing risks losing its appeal to the best young talent; and the build-for-today paradox, where CMOs are making staffing decisions now without knowing what their organization will need in two or three years.

“CMOs will have to figure out how to manage hybrid teams — human and AI working together. And, critically, how do you manage AI when your people haven’t done it themselves?” asked Michelle Froah, former global chief marketing and innovation officer at ETS. One CMO noted that young marketers can now gain years of experience faster through AI agents but lack the judgment to lead them, because they haven’t learned what “good” looks like.

Reaching Consumers in a Fragmented Media World

The old model of reaching a broad audience efficiently is gone. CMOs are navigating a splintered media landscape — one that includes connected TV, retail media networks, podcasts, influencer ecosystems, and AI-powered search — and face the ongoing challenge of keeping up with new platforms as they emerge. Each platform grabs further attention with individualized feeds algorithmically tailored for each audience member. And integrating the disparate media to provide a superior consumer experience is becoming increasingly challenging.

“Creating a seamless and frictionless and integrated experience for a consumer is one of the biggest challenges. Companies stitch the pieces of communications, commerce, and marketing together and try to portray one single, seamless experience, but it does not look like Michelangelo’s David. It looks much more like Frankenstein,” suggested Luis Di Como, former executive vice president of global media at Unilever.

The emergence of AI agents making purchasing decisions on behalf of consumers compounds the problem and raises a fundamental question: Who, exactly, is the marketer marketing to? As Noha Abdalla, CMO of Choice Hotels, said, “[AI] agents don’t watch TV, so maybe it doesn’t matter anymore, if the agent is picking for you. Are the end consumers part of the loyalty program, or are the agents part of the loyalty program?”

The dissolution of a clear audience, in conjunction with a shifting set of platforms with obscure priorities, has created a pressure chamber for marketers. Abdalla suggested that an unclear media system greatly broadens marketing’s remit. “How, as a marketer, can I influence what is shown to consumers in their agent and LLM interactions? It used to be, I had to create a breakthrough ad campaign. But now maybe my job is to make sure that the product experience, the customer experience, and the reviews are the best they possibly can be, because that’s what’s going to influence the [AI’s] decision-making.”

Measurement Challenges and As-Yet-Undetermined Metrics

Measurement challenges are nothing new to marketers. Even before AI, the ROI on social media was difficult to calculate. Marketing mix modeling (MMM) is a statistical technique for estimating how much each marketing input — TV, digital, social, promotions, pricing, and so on — contributes to a business outcome, such as sales or revenue, while also accounting for external factors, such as seasonality, competitor activity, or macroeconomic conditions.

Andrea Zaretsky, CMO of Morgan Stanley Wealth Management, suggested, “If I could wave a magic wand, I would take away the walled gardens so we can understand the return of our dollars on social media as well as receive comparable measurement and methodology from across the key platforms. Further, MMM takes three months to assess performance — three months in today’s real-time world. There really hasn’t been enough innovation in this area in many years.”

A major lag in the availability of this data dramatically slows decision-making around how a marketing function should adapt its approach. And beyond the perennial marketing measurement question, identifying the ROI of AI is a much tougher challenge. In a zero-click AI landscape, what, exactly, is to be measured, and how? As the field works to establish a meaningful mechanism to measure the reach of AI, marketers must synthesize inconsistent data and standardize measurements — a seemingly impossible task, given AI models’ dynamic black-box algorithms.

Today’s CMOs are building organizations in a dramatically changing business environment and developing talent for a future moving faster than the hiring process. As Greg Stuart, CEO of the Marketing + Media Alliance, said, “They are navigating a world being reshaped by AI, political polarization, and media fragmentation at a speed that makes last year’s playbook obsolete.”

But these CMOs are not waiting for clarity before acting. They can’t afford to. They are experimenting, iterating, and embracing change as an operating condition rather than an obstacle. The ones moving fastest are those who have accepted that speed itself is a competitive advantage — and moving forward imperfectly is better than planning perfectly.

Corporate sustainability is facing headwinds. Net-zero pledges are being quietly walked back. Regulatory pressure is loosening in some parts of the world. Shareholders are demanding stronger business cases. Inside companies, sustainability leaders sometimes spend more time defending their function than expanding it. The familiar question “Where is the value?” has returned with a sharper edge, often accompanied by an implicit either/or framing: profit or purpose.

In most organizations, sustainability transformations don’t dramatically fail. Rather, they lose their edge. Goals get softened; ambition narrows. The original aspiration is often still in the strategy deck, yet the day-to-day decisions begin to look like the ones the company would have made regardless of its sustainability plans. Understanding why this happens in corporations is difficult because the process often unfolds gradually and becomes visible only much later. The same tension between commercial viability and sustainability ambition plays out more quickly and visibly in early-stage sustainable or social impact ventures, where teams face similar pressures to reconcile profit and purpose but have less room to postpone difficult choices.

Over more than two years, in real time, we tracked six startups targeting sustainability or social impact, from the first idea through to either reaching proof of concept or collapsing. The research was published in Strategic Entrepreneurship Journal.1 We expected to find that the difference between success and failure came down to strategy, market timing, or capital. What we found instead points to something less obvious, and potentially useful for managers leading sustainability transformations in established companies.

The Pattern Behind the Drift

The conventional explanation for sustainability transformations losing momentum is that conviction wavers under pressure, such as the impetus to make quarterly numbers. There is some truth in this, but it overlooks how often the retreat is structurally encoded long before the pressure arrives.

In our study, three of the six ventures we followed ended in collapse. The founders rarely identified a specific moment of compromise that led to failure. The CEO of a failed mental health venture traced it to the team’s early idealism: “We have a very new product, which was launched with very high standards and high idealism, without saying ‘we grow on the go’ and [starting] with a simple, small product to make money first.” The realization that there was no viable business model came late. By the time it arrived, the venture was past the point where it could simplify its approach.

Strikingly, the same fate awaited a second team that, on the surface, took the opposite approach. Where the idealists committed too completely to achieving an ambitious sustainability impact, the pragmatists committed too completely to developing a business case before anything else. One of the pragmatic-minded founders told us, “Frankly speaking, we want to become rich with it. I also want to have my island in the Maldives.” The team agreed early to scale first and add social impact later, but they never got to the impact piece. The sequential plan (first survive, then pursue purpose) turned out to be one stop on a multileg trip. The second leg never happened.

Both companies failed, and in both, the team was internally aligned. The problem was that everyone shared a single, one-sided way of framing profit and purpose, with no one to push the alternate view.

The Cognitive Variable

The pattern that distinguished the ventures that built viable, mission-aligned business models from those that didn’t was not primarily strategy, market, or money. It was how the team thought.

The teams that struggled, whether they were idealistic or pragmatic, treated profit and purpose as a trade-off. They tended to think in either/or terms, which led them to tackle first one goal, then the other. In contrast, the teams that built durable models held the two objectives in tension, maintaining both/and thinking (a paradox mindset).2 They didn’t resolve the tension; they worked through it, decision by decision. They were slower to commit to a fixed business model, developed simpler early prototypes, and were more willing to engage stakeholders whose feedback complicated their plans. From the outside, the second group of companies sometimes looked indecisive. But taking time to reason through their decisions turned out to be the source of their later flexibility.

Why does this matter for established companies pursuing sustainability transformation? Research on corporate sustainability has long argued that managers’ cognitive frames shape how they perceive and respond to sustainability tensions.3 Our findings suggest that this is not just specific to individual managers but also a property of their teams. The cognitive composition of the team running a sustainability transformation, typically a small group reporting to the C-suite or with a C-level member at its core, is the corporate equivalent of a founding team. They make the early structural decisions that determine the transformation process. They define what the sustainability function does and doesn’t do. They set the metrics and ambition level. They choose the language. And whatever logic dominates their thinking gets imprinted into the organizational architecture they build.

This early founding logic is what most companies underestimate. In our cases, the early choices the teams made based on how they thought (what to commit to, how complex to make the offering, which trade-offs to accept) became increasingly hard to reverse as their venture moved forward. Each implementation step locked in the assumptions of the previous one. The way these early conditions get baked into the structure and persist is what management researchers call structural imprinting.4 By the time financial pressure or stakeholder feedback signaled that the model wasn’t working, the team had less room to maneuver than they had realized. The structural lock-in was already in place.

The same dynamic often plays out in corporate sustainability functions. The KPIs chosen and the targets set in the beginning often frame what counts as progress for the coming years. The reporting structure that’s set up can influence which conversations happen and which don’t. The team that’s assembled affects what gets noticed in the first place. Most of these early decisions feel small at the time. They are not. And all of these small, early choices are, to some extent, a function of how a leader thinks: either/or or both/and.

Why Today’s Risk Is Asymmetric

In principle, both ends of the spectrum are dangerous. A purely idealistic transformation team can over-engineer the program, refuse useful compromises, and produce something the rest of the business cannot absorb. A purely pragmatic team can quietly hollow out the more aspirational ambition until what remains is business as usual, just rebranded.

The gravity in most established companies pulls in one predictable direction. The ambient pressure is to soften, defer, and de-prioritize the purpose part. With sustainability under sharper economic and political scrutiny, and with the regulatory and reputational tailwinds of the late 2010s replaced by harder questions about return on investment, the dominant internal voice is the one asking for the business case. That voice is not wrong to ask; the questions are legitimate. But if it is the only voice within the transformation team, the early imprinting tilts toward pragmatism. And pragmatism, once imprinted, doesn’t tend to course-correct toward purpose. Our data on the sequential “first survive, then pursue purpose” path suggests that the second part doesn’t reliably arrive.

The implication is that stronger conviction at the top alone is not enough to protect against drift. Divergent thinking, conflict, and productive friction within the team driving the work are also needed.

Three Things Managers Can Do

Drawing on the patterns we saw in the ventures that managed to get off of failing paths, we identified three practical implications for sustainability transformation leaders and the teams driving them.

1. Select team members for cognitive friction, not consensus. The most viable teams in our study held real tension internally. They were not comfortable, but they were productive. One health-tech CEO described a recurring debate between a psychologist on the team who insisted that the assessment instrument they were offering needed to be more rigorous, and a colleague who insisted that it didn’t matter as long as it sold. Neither of them was entirely right, but together, they kept the model honest.

That means a critical question to ask when assembling a sustainability transformation team is “Does this person disagree to some extent, productively, with the rest of the team about what we are here to do?”

2. Build in reality checks before pressure forces them. In our research, the teams that managed to find alternatives to failing paths almost always did so because of input they had not asked for. A new colleague joined and reframed an assumption. An accelerator coach pointed out that the commercial model didn’t work. A stakeholder responded to a prototype in a way the team had not expected. These interventions could be considered because they came before structural lock-in had narrowed the team’s choices.

Corporate transformation teams likewise should create opportunities for reality checks that don’t depend on a crisis to surface. They should consult standing advisory groups and customer or stakeholder panels early, while they are still developing plans, rather than going to them only later to seek approval. These interventions might feel inefficient when things are going well, but they could be the inefficiencies that buy flexibility later.

While it can seem like more governance, these reality checks add the kind of friction that an external-pressure event would otherwise force on the sustainability team under worse conditions (often when it is too late).

3. Keep early architecture simple enough to flex. The teams in our study that survived shared a pattern: While they embraced complex both/and thinking, they were strikingly cautious about complexity in the beginning. They prototyped simply. They committed late. They added scope only when stakeholder feedback warranted it. In contrast, the teams that failed (both idealists and pragmatists) went big early. They built elaborate offerings before they understood what worked. In other words, the teams that could hold the most complexity in their thinking were often the most disciplined about keeping their early moves simple.

For corporate sustainability functions, this translates into a heuristic: Build the simplest version of the program in the first place that still credibly addresses the core mission. Resist the urge to ship a flagship initiative that bundles five priorities into one. Each priority that’s added early constrains the next decision that can be made.

The pressures on corporate sustainability won’t ease in the near term. But the three levers we learned from observing early sustainability and social impact ventures (selecting team members for cognitive friction, building in early reality checks, and keeping initial architecture simple) can facilitate both/and thinking in a transformation team when pressure pulls toward either/or — either profit or purpose. Taking team composition and its internal dynamics seriously from the start helps a transformation keep its edge.

Businesses are striving to adapt ever faster to keep pace with rapid change, and yet B2B marketing practices have remained surprisingly static. Sure, the tactics have shifted to digital executions, and the use of data has made targeting B2B buyers more precise, but marketers remain rooted in fundamentally flawed assumptions about the corporate buying process.

For far too long, marketing leaders have operated on the belief that B2B purchasing decisions are almost entirely rational, driven primarily by product-feature superiority and competitive pricing. However, this conventional wisdom supposes that buyers live a Dr. Jekyll and — let’s say — Mr. Spock existence. In their private lives, they are fully formed human beings subject to all the cognitive and affective influence that B2C marketing wields to shape personal buying decisions. But when they step into the office, they become emotionless, like the famous half-Vulcan first officer of the starship Enterprise, and all of their decisions are informed solely by logic. This assumption undergirds nearly all B2B marketing efforts.

Nevertheless, the truth is that B2B buying is far more emotional and socially influenced than C-suite leaders have accepted. The same emotional triggers, social forces, and cognitive biases that influence us in our personal lives remain in play in our professional lives as levers that marketers can use to shape B2B purchasing decisions.

This shift in understanding isn’t just relevant for marketers in exclusively B2B settings; it’s a crucial insight for CMOs across industries, who often have at least one segment of business customers. Netflix must simultaneously attract individual subscribers and court advertisers; Google markets consumer-facing search along with enterprise cloud solutions. Even educational institutions like the University of Michigan’s Ross School of Business, where I teach, must appeal to both individual students for its MBA program and corporate partners for its executive education offerings. In each case, the underlying principles of how to influence decisions are similar, but marketers are unlikely to use similar tactics.

Of course, marketing and go-to-market strategies are going to be different for B2B than for B2Cs, even if they work on similar leverage points, because the buying processes are so different — and, as we’ll discuss, more relationally complex for B2B. To uncover the human drivers that influence purchasing decisions, I partnered with Mimi Turner, head of marketplace innovations at LinkedIn, and Jann Schwarz, senior director of marketplace innovation and strategies at LinkedIn and founder of its B2B Institute, on a research study. We surveyed 750 senior B2B buyers responsible for large-scale, enterprise purchases about their purchasing decisions. Forty-three percent were at the vice president level or above, and 41% were in organizations that had over 10,000 employees and were engaged in billion-dollar deals. Our exploration focused on what gets purchased and why once a short list of brands has been narrowed down to a final consideration set.

Our research revealed that buying decisions are heavily influenced by individuals who are not the nominal purchaser. They may be executives, such as the COO or CFO, or staff members from procurement. While B2B marketers typically target the known prospect or lead who has technical experience and domain expertise relevant to the purchase, these other individuals typically lack nuanced understanding of the product value propositions under consideration. These hidden buyers don’t download white papers or attend webinars and are more or less invisible in terms of general B2B marketing signals. They don’t turn up on lead sheets or pipeline trackers. Yet their perception of brand contributes to up to 50% of the buying decision, according to our research. Understanding and swaying these hidden influencers is critical to driving B2B purchasing.

We also found that a contributing factor to B2B purchasing decisions was how defensible the decision was. For instance, if a product bought from an established, reputable company performs poorly, the buyer is unlikely to be blamed, whereas if a purchase is made from a young startup, the buyer’s judgment may be questioned. As the old business adage went, “Nobody ever got fired for buying IBM.” That means the more defensible option often wins, even if a competing vendor offers a better value proposition. According to additional interviews we had with buyers, it’s also easier to get the more defensible option through procurement. This means that brand reputation matters, perhaps even more so than performance claims, once the short list has been narrowed to vendors that meet the buyer’s technical requirements. In fact, 81% of respondents said that in purchasing situations, almost everyone with a voice in the decision knew the brand that was ultimately selected. That makes brand marketing, which typically focuses on emotions and values, essential to B2B sales.

Lastly, we found that the confidence to make a purchase in a B2B context — what we call buyability — is significantly influenced by the extent to which the vendor is culturally aligned with the purchasing company. B2B buyers want to make sure that the company they’re buying from is not just good but also aligned with the working styles and priorities of their company. For instance, Patagonia, known for its strong environmental stance, discontinued its business of emblazoning its merchandise with the logos of hedge funds and tech companies, because it regarded them as contributors to planetary degradation. That means B2B marketers have to communicate not only product specs and value but also how the organization sees the world and how it acts within it.

The upshot? C-suite leaders must rethink transactional B2B marketing approaches that focus on product features and take steps to understand purchasing culture at buyer organizations. They must speak not only to nominal buyers but to unseen stakeholders that ultimately drive buying in a B2B context, and they must understand what those stakeholders value and how to influence them. Doing this well is a more significant determinant of who wins a deal than whether a product is cheaper, faster, or more efficient. Investing in brand marketing that communicates a company’s culture and values matters as much in B2B as it does in B2C.

Dry rot. A Potemkin village. The Wizard of Oz.

What do those things have in common? In each case, they may look good on the surface, but it’s only an illusion. Wood afflicted with dry rot looks just fine until the tree it’s in topples down. Grigory Potemkin is said to have tried to impress Catherine the Great by creating a prosperous Crimean village constructed only of building facades, masking the real town’s crippling poverty. The Wizard of Oz looked great and powerful, until the cowering man behind the curtain was revealed.

Could artificial intelligence do the same to organizations — create an impressive, seamless exterior even as capability completely falls apart inside?

This alarming possibility was raised by the AI experts we interviewed for a joint Anthrome Insight-Axialent study on AI’s impact on behavior and culture inside organizations. Even as those experts uniformly cited AI’s potential to transform work for the better, they cautioned that AI adoption risks unintentionally creating an organizational mirage: workplaces that appear highly capable while people’s real skills quietly erode beneath polished AI-generated output.

Perhaps even more worrisome: When we can no longer reliably tell who truly knows what, the interpersonal trust that teams depend on to function also crumbles.

What are the early signs that the mirage effect is already forming in your organization — and what does it look like when it takes hold? How can leaders avoid the mirage effect and make sure that in the AI age, their people and organizations are as capable as they appear?

Capability Mirages: The Early Signs and Troubling Possibilities

An AI-driven illusion of competence, hiding the absence of genuine understanding, is already appearing in some organizations, the AI experts told us. Stephanie Antonian, founder and CEO of AI product development company Aestora, explained how this can play out in dangerous ways for leaders and organizations: “The upside [of AI tools] is that everyone can produce a level of work that’s pretty good for basic tasks. … It looks pretty good. But then you don’t know what’s underneath it, how resilient that piece of work is, or whether it’s going to give you an additional liability.”

After all, AI is not necessarily an improver of work but, rather, an amplifier. As AI Business Impact CEO Gábor Szórád observed, “At the end of the day, if you are a fantastic software engineer or a great manager, AI allows you to do more with the same energy. If you’re a bad one, you’re just going to create more crappy instructions, longer ones, more bad ideas. It just magnifies whatever you put in.”

The tech vendor marketing narrative stating that AI makes individuals more capable just compounds the mirage issue. In some situations, AI does enable the production of better short-term outputs — but it also breaks the historical link between strong output and strong capability, causing teams to lose sight of who actually possesses strong skills.

Moreover, AI itself doesn’t know when it’s out of its depth, said Amir Michael, professor of accounting and deputy executive dean for executive and professional education at Durham University. Likewise, Antonian observed, people who aren’t capable in a particular subject area can’t spot where AI-generated work has gone wrong — unlike people with subject-matter expertise. They may pass on bad output to others who also can’t tell the difference.

The mechanism through which professional mastery has always been built — productive struggle, error-based learning, the slow accumulation of genuine judgment — may be quietly bypassed before many people realize what is being lost.

But there is a second, less visible problem that, in the long run, may be the more dangerous one: Not everyone is self-aware enough to notice their own skills eroding. There is no single individual — among managers, colleagues, or clients — who can do an accurate, real-time read on how capable an organization is, overall. Individual skills and collective capability could weaken long before anyone notices, and there is absolutely no guarantee that even the best-functioning AI could begin to fill the gap. Leaders should be especially concerned about losses in the sophisticated, critically necessary “muscle of critical thinking,” said Albert Durig, cofounder and partner at Triviam Consulting.

This situation results in a direct and damaging consequence for organizational trust. Teams have traditionally functioned based on knowing, or at least being able to calibrate, who knows what. That calibration determines whose judgment should be relied upon, how managers identify who is ready for greater responsibility, and how organizations know what they are truly good at. When AI makes that calibration unreliable, trust erodes with it. When AI-assisted work is later discovered to have been misrepresented, the trust collapse tends to be swift: “The impact on trust is 0 [doubt] to 100,” said Elisa Farri, vice president at Capgemini Invent Management Lab.

Concerned yet? We all should be.

But if leaders act now, they can avoid organizational dry rot — and maintain true capability fueled by humans and machines alike.

How to Preserve Human Capability and Team Trust

Let’s explore what the experts we spoke with had to say.

1. Choose purpose and business goals over an AI-first mentality.

It’s trendy to announce that your company is thinking “AI first.” But it’s not what these experts would recommend. Antonian put it directly: “When you go AI-first, you have already told your organization it’s not human-first.” That signal, once received, is hard to unsend — and at a moment when people are already anxious about their relevance and job security, it can quietly erode the trust that makes teams function.

One potential outcome of an AI-first future is an unpleasant inversion of the roles of human and machine, said one senior AI executive at a Fortune 500 company. They see a concrete risk that people might let AI do the reasoning, interpreting, and responding only to become “transactional tools” themselves.

“AI can be the lead,” Michael noted. “That’s the problem. As long as AI is your follower — it follows your requests, your orders — we’re fine. The time that AI jumps to be your lead, that’s the downturn.”

AI-first thinking, in the view of our experts, causes people to lose perspective on the utility of AI as a tool — and to deploy AI in comically inappropriate settings. Remember the old saw “If you’re a hammer, everything looks like a nail”? That applies to AI-first: “You don’t walk around the house, holding the biggest drill that you have, asking people if they need their coffee stirred,” said Andrea Jones-Rooy, a data scientist, organizational researcher, and visiting associate professor at New York University’s Center for Data Science.

Even when AI is used for more seemingly appropriate ends, such as measurement and KPIs, Durig noted, it can cause a dominance of measurement over meaning. That leads to “a performance culture without purpose,” he said.

On the flip side, when purpose comes first, enabled by AI tools, our experts see the potential for true progress and even stark disruption. With AI enablement, “Small groups of people that get together for a specific purpose may outperform corporations because they are more nimble, flexible, fast-moving,” said AI entrepreneur Thierry Kahane. The key, in Kahane’s vision, is cohesion around a goal versus a technology.

What concrete leadership steps fuel this mission? First, you should check your own rhetoric. If discussion of AI is eclipsing dialogue around business outcomes, the conversation is framed incorrectly; people are more likely to become passive, let their skills slip, and quietly lose faith in their own relevance. Similarly, if the only voices you’re hearing in the AI conversation are those of people who are passionate about the technology, the balance of business-purpose versus tool is likely off.

Finally, it’s critical to routinely audit how people are operating AI tools on the ground. Is AI making decisions that humans should be making? To ensure that an organization is operating “business first,” leaders and teams must exercise constant vigilance around day-to-day AI use. Otherwise, human capability is destined to slide … and we won’t know until a black swan event happens that AI, trained on typical data, is ill-equipped to handle it.

2. Leaders should model AI usage specifics.

If you don’t want people to switch off their judgment, critical thinking, and intrinsic motivation, you must teach them how to engage with AI as a sparring partner rather than as a delegation tool, Farri said. People need to engage in active, back-and-forth interaction with AI —interrogating its suggestions, pushing back, and cocreating outputs.

And this behavior needs to start at the top. When leaders model active, curious, judgment-led AI use and do so visibly, this signals what the organization actually values far more powerfully than any guidance document can, Szórád said. “The project sponsor needs to be the CEO. The CEO needs to use AI daily,” he said.

That’s important advice at a time when many organizations have given employees only the barest clarity on how to utilize AI day to day. This may be well intentioned on those organizations’ parts; perhaps their leaders don’t want to stifle employee creativity. But thoughtful guidance on AI usage can walk the tightrope of specifying behavior without shutting down exploration. For example, organizations can distribute highly generic but intelligently framed prompts — “When I say _____, what am I not thinking of that I should consider?” — that keep the human in the driver’s seat.

The AI realm is still new enough that people genuinely need to be steered away from the wrong behaviors: Telling people what not to do is just as important as telling them what to do, Farri said. Caveats must be clearly communicated, she added. For example, you might say “The more flawless an AI output looks, the more ruthlessly you should stress-test it.”

3. Build human capabilities first, AI augmentation skills second.

MIT Sloan School of Management postdoctoral researcher Isabella Loaiza offered a deceptively simple principle that organizations are widely ignoring: “You need to learn first and then use a tool to supercharge your abilities.” One executive asked us to imagine what could happen when the sequence is reversed: “When someone with 20 years of professional experience uses AI to amplify their impact, that works. But what happens when a 20-year-old’s first interaction with work involves AI from day one? How do we develop that person at the same speed and depth? We will create a talent gap that will be hard to close.”

This has direct implications for how organizations design onboarding, early-career development, and role progression. One question you should ask now is not “How can we use AI to accelerate this person’s output?” but “What does this person need to genuinely understand before AI can help them go further?”

Performance measurement systems need to take this phenomenon into account too — not just asking “Is the work good?” but “Does this person understand it, and are they growing through producing it?” Measuring human-centered outcomes requires human-centered metrics, which most organizations have not yet built.

4. Restore accountability, and use it to rebuild trust.

Accountability is challenging in the best of times. Knowledge work, particularly, has long been slippery to ascribe: That’s due to both messy over-collaboration and, frankly, some bad, illegitimate-credit-taking behavior within teams. When you don’t know who does what, it’s hard to hold anyone accountable. Add in AI, and the accountability muddle gets even worse, with direct consequences for both skills and trust.

When a team member takes credit for work that an AI clearly performed, intra-team trust is naturally eroded, several people commented. But worse — and central to the issue of AI mirages — if a team is not sure how work is getting performed (that is, what mix of AI and human capability, exactly, is being deployed), it quickly becomes impossible for colleagues to understand each other’s skills. Without that visibility, the natural ways that teams maintain their collective skills will break down. More seasoned team members and managers won’t know who to coach.

The trust consequences extend further than most organizations realize. Kahane described a dynamic that he said is almost universal: employees deliberately concealing their use of AI to ascribe productivity gains to their own efforts — thus protecting manager and peer perceptions of their human performance. (Researchers are beginning to see this dynamic in academic settings as well. Students are concealing their AI use due to perceived taboos.

Daniel Strode, a professor at the IE School of Human Sciences and Technology, pointed out the vicious cycle this could create: Employees hide AI use; leadership expects efficiency gains that don’t materialize; neither side communicates effectively; and AI initiatives collapse under the weight of accumulated mistrust. The irony is sharp: The very tool meant to help organizations perform better becomes the source of the opacity that prevents them from understanding how they are actually performing.

Now you might say, “Shouldn’t that be OK? If the AI can do the work capably, who cares?” The issue, though, is that the AI can do the work capably until it cannot. And as previously noted, neither the humans involved nor the AI know those specific capability limits, which are masked by the polish of AI outputs. Losing early-warning signs of capability gaps only makes the eventual emergencies more dire. The solution here can be classified as “simple but not easy.” Cedric Wells, head of IT innovation and new technologies at Gorilla Glue, framed it practically: “Set clear ownership and verification norms.” Who produced the work? What role did AI play? Who is accountable if it is wrong?

Today’s AI technology has raised important questions about the basics of how we get work done — especially around who (human or technology) has the capability to do what. Individual-skills erosion, and collective capability erosion, could damage organizations irreparably. If AI tools are both contributing to human deskilling and concealing that fact, they could speed the collapse of many organizations. The experts we spoke to were alarmed, but they were also hopeful that leaders will step up to the challenges.

If leaders coach teams to use AI thoughtfully, accountably, and with real purpose, the mirage could become the reality: Organizational capability could meaningfully increase and human skills could grow too. We can be our best selves if we lead technology and are not led by it.

A vice president of product opens her laptop on a Monday morning to find that the AI model her team had worked with for the past six weeks to build a customer workflow has been leapfrogged by a cheaper, faster alternative. Again. Her Slack feed is blowing up with links to the announcement. The CEO has already forwarded an article about what a competitor is doing with the new tool, with the subject line “FYI.” She hasn’t finished rolling out the last integration, and now she’s wondering whether to scrap it. She is not resistant to AI. She is worn out by it.

Most leaders look at this picture and see an execution problem: The organization wasn’t moving fast enough. The cautionary tale that reinforces that instinct is Chegg, the education company whose market capitalization collapsed when the launch of AI-powered alternatives rendered its core tutoring model obsolete. The lesson everyone has drawn is obvious: Move fast or die. So leaders push harder, with more pilots, more mandates, and a constant drumbeat of urgency.

But that lesson, taken too literally, backfires. Bracing only against the danger of moving too slowly, leaders managing AI adoption underestimate a quieter risk: that they will wear out their organizations by racing toward a finish line that does not exist. The old playbook was built for disruptions that end, and its instincts (move faster, push harder, wait for things to settle) become liabilities when there is no end state. Leaders who optimize for speed alone will lose to those who build for endurance as well.

What follows is a reframing and a set of emerging practices for leading through an AI disruption that will not settle.

From Process to Permanent Condition

Research on disruption has been circling this problem for years. One influential strand that one of us (Rory) developed with Clayton Christensen and Michael Raynor pressed on the point that disruption is a process, not an event. The recurring incumbent error is to judge the threat by where it stands rather than where it is heading. Yet, even correcting for this carries a quiet assumption that the threat’s trajectory has an ultimate destination. After Netflix disrupted Blockbuster, streaming became the new normal. Each wave of new technology ran turbulently for a while and then hardened into arrangements a company could see and plan around.

AI changes this dynamic. With previous technologies, the entrant’s advantage grew because something outside it improved: Components got cheaper, networks got faster, supply chains got better. AI is increasingly self-improving. Each generation helps train and build the next, so the distance between waves keeps shrinking. There is no settled position to plan toward, because the core keeps extending and the old barriers to disruption keep falling. We have come to call this condition steady-state disruption: a context in which capability shifts arrive continuously and accelerate one another, with no equilibrium in sight.

Managers and scholars already have language to describe turbulent environments. They talk about VUCA (volatility, uncertainty, complexity, and ambiguity) and about the dynamic capabilities a company needs to sense change and adapt. But that vocabulary assumes that the turbulence eventually breaks: A period of upheaval is followed by a return to relative calm. Steady-state disruption is the condition in which the calm never comes.

If disruption is a process rather than a sequence of separate shocks, then the AI capabilities landing inside an organization are not a series of discrete events to be handled one at a time but an ongoing process. Companies that treat a continuous process as a string of episodes fatigue their employees and ultimately struggle to adapt.

The pace itself shows no sign of letting up. Stanford’s 2026 AI Index shows AI systems posting steadily higher scores on standard industry benchmarks. New frontier models appear every few months, and an industry investor reported that the leading model often holds its position for only a few weeks before a newer one or an open-source rival takes share.

This cadence is especially hard to absorb because the work never reaches a stopping point. As Airtable CEO Howie Liu has observed, AI adoption is unlike the move from desktop to mobile or from on-premises to cloud computing. Each of those shifts was a single, fairly foreseeable change in form, but with AI, every model release brings new capabilities and new patterns that have to be learned more or less from scratch. Even if this progress were to hit a sudden plateau, organizations would still need to spend years folding existing capabilities into their products, workflows, and decision-making.

Addressing the Human Cost

For a lot of people, the early excitement has curdled into something that’s harder to sustain. Recent research makes the cost concrete. One analysis found that AI tends to intensify rather than lighten individual workloads, piling on cognitive demand faster than it strips away drudgery. And Deloitte’s Global Human Capital Trends survey saw the same thing at the organization level. As collaboration with AI deepens, so, too, do burnout, loneliness, and overload.

These findings document strain on individuals, but the cause sits above the individual level. When a playbook written for episodic disruption no longer works, the organizational machinery that once absorbed shocks now transmits them directly to workers instead.

So, what should a change management toolkit for the age of continual AI disruption look like? A handful of practices that are provisional but useful are taking shape at leading companies. Instead of placing the burden of absorbing change onto individual employees, these practices move some portion of that burden onto the structure of the organization.

Practice 1: Build a Permanent AI Unit

Since 2022, a common response to the rapid rise of AI has been to form an AI committee — a group of people asked to advise, set direction, and evangelize on AI, usually on top of their existing jobs. Committees of this kind tend to add work rather than soak it up. Members are stretched thin, the rest of the organization gets only intermittent guidance, and the committee’s own queue keeps growing. Steady-state disruption calls for a sturdier, more permanent approach. Whether it is a full-time team at a large company or a carved-out slice of a few people’s time at a smaller one, the work of tracking, translating, and triaging AI’s churn should be somebody’s actual job rather than a standing favor.

Microsoft offers an illustration. The AI Center of Excellence inside Microsoft Digital began as an ordinary advisory group in 2023. But the group’s leader, Qingsu Wu, recalled that as adoption spread, so did duplicated effort, uneven governance, and gaps between strategy and implementation. The question, Wu said, shifted from “How do we help teams try AI?” to “How do we turn AI into consistent, measurable outcomes at scale?” The center became the place where AI work is coordinated, with a single idea intake pipeline, a hub for upstream architecture and security decisions, and the ability to see patterns where individuals and siloed teams cannot.

Once such a group has enough depth, the frontier-scanning activities and scrap-or-scale calls that used to land on scattered individuals become the standing remit of people equipped to handle them.

Practice 2: Run Two Clocks, Not One

Most organizations keep time on a single clock. Plans are quarterly, budgets are annual, and the all-hands meeting lands on its dependable schedule. Onto that steady rhythm, employees are now also being asked to ship AI experiments by the week, keep last quarter’s integrations running, follow a frontier that shifts constantly, and explain to leadership what any of it means for the business. Companies have always lived with some gap between fast work and slow work. AI has widened it past the point where one person can comfortably hold both ends.

The strain shows up first in product organizations, where the distance between weekly model releases and quarterly road maps is hardest to ignore. Airtable CEO Liu watched AI-native competitors shipping major capabilities every week while his own teams followed quarterly road maps. So he split the product organization into two groups, borrowing a distinction from psychologist Daniel Kahneman: A fast-thinking group ships AI capabilities on a near-weekly basis, while a slow-thinking group takes on the deliberate infrastructure bets — the kind of work that, as Liu put it, you cannot ship in a week via a “hacky prototype.” The two are meant to feed each other. The fast group surfaces new possibilities, and the slow group turns the promising ones into things the company can rely on. Without such a deliberate split to protect the slow clock, the faster clock becomes the standard against which everyone is measured.

Practice 3: Teach in the Flow of Work

Under constant change, most corporate training approaches, such as annual certifications or one-off workshops, fall behind almost as soon as they are delivered. Employees are left to keep pace with new developments on their own and end up concluding that the frontier is simply outrunning them. The feeling only deepens when the learning is stacked on top of a job that is already full.

Learning should instead be made continuous and specific to the role, delivered inside the work itself. There is good evidence that knowledge sticks better through short, repeated exposures spread over time than through one-off intensive sessions. Salesforce adopts this approach with its AI-powered internal platform, Career Connect, which reviews an employee’s existing skills, identifies the gaps between those skills and their aspirations, and serves up tailored learning opportunities (such as courses, stretch assignments, or mentorship) through Slack, where they already work. The organization’s Agentforce Learning Days add a recurring, companywide push on AI skill development specifically. What emerges is an architecture for AI fluency and ongoing development that is steadier and closer to the work, better aligning with the realities of steady-state disruption.

What complicates this approach when it comes to AI is that employees who fear being replaced by the technology have little incentive to engage with it seriously. CEO Greg Case at insurance and risk advisory firm Aon has addressed this fear directly. His bet is that AI will widen what the firm’s roughly 60,000 employees can do rather than substitute for them. He has built Aon’s investment in continuous AI fluency around that framing and has credibility with his employees for leading Aon through the pandemic without layoffs. Continuous learning requires continuous buy-in, and buy-in requires workers to believe that getting better at AI benefits them, not just the organization.

When training stops being a place employees go and becomes part of how they work — small and constant rather than disruptive and periodic — it keeps a workforce current without asking people to absorb the frontier on their own time.

The vice president of product we mentioned at the beginning was faced with a new-model announcement, a forwarded article in her inbox, and a newly built workflow that was already obsolete. This steady-state disruption is exhausting because the weight of all this change rests on her and her coworkers, with nothing in the organization’s design built to help them carry it.

A generation of managers learned that disruption was a phenomenon that eventually settled. AI shows no sign of settling. Leaders who keep treating it as a series of episodes will keep piling that weight onto their people. Those who build organizations designed to carry it, through permanent AI infrastructure, split cadences, and work that has learning embedded into it, will be the ones who endure.

How does a fossil fuel company become the world’s largest developer of offshore wind? How does a software company written off for missing the mobile revolution become one of the world’s most valuable companies in the age of AI? Ørsted and Microsoft have faced a puzzle familiar to many leaders: When technological, regulatory, and societal shifts redraw the basis of competition, reinvention is no longer optional, but the path forward is far from obvious.

Reinvention has been a key concern in business strategy. Stories of startups disrupting markets are compelling and often romanticized. But established companies do not have the benefit of a clean slate. Existing frameworks illuminate different parts of the problem: Disruptive innovation explains why incumbents get trapped, value shows how companies create new markets, and work on renewal and reinvention highlights the organizational barriers that make change so difficult.1

Leaders are faced with a dilemma in strategizing their way forward. Some companies double down on their known strengths: They keep innovating around the core and extend existing products, technologies, or business models into adjacent spaces. Kodak’s unwavering attachment to film demonstrates the risks of sticking to the legacy formula. Starting from a clean slate with a pivot to an entirely new business may seem the better option. But in doing so, a company risks discarding capabilities, relationships, and resources that could offer comparable advantage in new markets while also alienating internal and external stakeholders.

Consider Ørsted. In 2009, the oil and gas company, then known as DONG Energy, was also Denmark’s largest utility and biggest CO₂ emitter, with 85% of its power coming from fossil fuels. But its business was under pressure from multiple directions, including exposure to volatile swings in fossil fuel prices, and public opposition to building new coal-fired power plants.2 At the same time, the global energy transition created new opportunities. To DONG’s leadership, it was clear that the existing business would not be viable long term, environmentally or financially, and that renewables were the path forward.

The question was no longer whether to transform but how. DONG’s leadership looked inward. The company’s portfolio included the world’s first offshore wind farm. More importantly, decades of extracting petroleum in the North Sea had enabled it to build something harder to replicate: deep capabilities in large-scale infrastructure engineering, complex project development, and the logistics of operating in harsh marine environments. This became the foundation for renamed Ørsted’s 85/15 Black-to-Green strategy for inverting its portfolio mix to 85% renewables and 15% fossil fuels, scaling its offshore wind business, and achieving a 30-year target in only a decade.

Ørsted’s remarkable transformation exemplifies a recurring pattern that we observed in our research on corporate reinvention stories. Microsoft’s trajectory tells a similar one. After failing to respond to the mobile shift and then making an ill-fated $7.2 billion acquisition of Nokia’s devices and services business, Microsoft emerged from what many have called a lost decade to become a leader in enterprise cloud computing. Like Ørsted’s, this reinvention was anchored around something deep, durable, and transferable: Microsoft had the trust of enterprise CTOs, and its products were in almost every Fortune 500 company. Its enterprise business competencies, combined with its nascent cloud computing business, became the foundation for a growth story that resulted in a tenfold increase in the company’s valuation under CEO Satya Nadella.

This is the pattern we have observed among successful corporate reinventions. Companies that navigate structural disruption do not typically build from scratch, nor do they simply protect the core. They uncover a deep, embedded capability that already exists inside the organization, often built for one context but carrying latent value that can be directed toward a fundamentally new strategic purpose. These capabilities, which we call kernels of reinvention, are powerful anchors around which new businesses can be built.

Finding a Kernel of Reinvention

Evolutionary biology has a name for the process Ørsted, Microsoft, and similar companies have undergone: neofunctionalization, where a gene that evolved for one function acquires a novel function. In evolution, new functions do not always arise from entirely new structures. Organisms facing shifts in their environment can benefit from repurposing existing genetic material for new advantage under environmental pressure. For example, millions of years ago, as the Southern Ocean cooled, Antarctic zoarcid fish evolved an antifreeze protein gene through the neofunctionalization of another gene, allowing the fish to survive in waters where other creatures would have frozen.3

Sometimes the kernel of reinvention is technical or scientific: Fujifilm’s future did not lie in film itself but in the chemistry, materials science, and precision capabilities involved. BYD’s kernel was a deep competence in battery electrochemistry and power electronics that became central to electric vehicles and broader energy applications. Kernels can also be commercial and relational, such as Microsoft’s enterprise business capabilities and trusted relationships with enterprise customers.

Four characteristics distinguish a genuine kernel from wishful thinking about legacy assets:

Deep. The kernel is not what the company sells, builds, or is known for but something that sits beneath it — underlying capabilities or resources that make today’s business possible. It can be genuinely hard to see because companies need to look beyond their strategic position, through their products and architectures. In photography, the kernel was not film but the chemistry and precision-coating science beneath it. For example, the same science that kept film stable and protected from ultraviolet light could also be used to keep skin moisturized and protected from UV damage. Fujifilm harnessed it to move into new markets, such as cosmetics.

Generative. Kernels enable a company to pursue new value and competitive positions. In e-commerce, Walmart’s capabilities in grocery retailing, including perishables logistics, and a dense retail network that used its stores as last-mile fulfillment nodes, became a launchpad for online grocery sales — a segment that pure-play digital players had yet to crack at scale. Walmart integrated this kernel with new e-commerce capabilities to transform it into an omnichannel model that served as a wedge. Later, it borrowed from the Amazon playbook by layering on a third-party marketplace and advertising business.

Defensible. Kernels need to give a company something that rivals in the new market will struggle to build quickly. Microsoft’s enterprise advantage in cloud and AI rests on trust and operational dependency built over decades with Fortune 500 customers, the security integrations embedded in its procurement and compliance processes, and the developer ecosystem that compounds with every new product. New cloud entrants could match the underlying compute, but they could not come to market with a web of institutional relationships and embedded dependencies similar to what it had taken Microsoft years to build.

Coherent. Perhaps least obvious, a kernel should create a visible, unbroken narrative thread from the legacy business to the new direction. This helps make the reinvention story more credible to employees, investors, and partners because it builds from a position of advantage rather than competing from scratch in a completely new field. Fujifilm explicitly positioned its beauty product line as based on the science developed for film.4 Similarly, Ørsted faced resistance to moving away from coal among its own employees, so it framed offshore wind as building on transferable skills and capabilities.5

In some cases, the kernel may already be being expressed through a nascent side business, an overlooked capability, or an underleveraged growth platform. (See “How Companies Build on Kernels of Reinvention.”) Ørsted had fortuitously inherited a modest wind portfolio through a 2006 merger with six Danish energy companies. In 2003, BYD had acquired small Chinese automaker Qinchuan, which it ran quietly in parallel with its existing business for nearly a decade before electric vehicles became the dominant story. Microsoft similarly had a nascent but growing cloud business under Nadella that became the foundation for its growth.

The Reinvention Playbook

Recognizing kernels in retrospect is easier than identifying them in real time. Successful reinvention is not a single decision but a sequence of decisions and continuous adaptation over years. Ørsted’s transition took a decade, while BYD’s took two decades — from the time it acquired an automotive business to becoming the world’s largest electric vehicle maker. The playbook below maps five steps toward kernel-based reinventions.

Step 1: Diagnose the disruption. Start outside the company. What technological, regulatory, social, or economic shifts are changing the rules of the game? Which parts of the current business are being weakened, commoditized, or rendered less legitimate? If the basis of advantage is moving, leaders need to confront that early. DONG’s leadership saw that fossil fuel volatility, public opposition to coal, and a strengthening renewables agenda were not isolated signals but a moment when converging trends made action both necessary and credible. Taken together, they suggested a structural change in the energy market’s direction.

It is important to guard against misdiagnosing structural shifts as cyclical headwinds. Such misdiagnoses can lead companies to build on kernels that are too close to the legacy product or business model. Consider whether the change is threatening current performance or the company’s underlying relevance.

Step 2: Excavate and validate latent kernels iteratively. Leaders must be prepared to dig, looking beyond the core products or technologies it currently sells to uncover deep, latent capabilities and assess their value in new markets. These are deeply embedded capabilities, such as subsurface engineering in the oil and gas industry, which could readily transfer to wind energy initiatives. They could already be being expressed through peripheral businesses that are relatively small in scale and not considered core, such as DSM’s penicillin capabilities or Ørsted’s modest wind-power business.

Leaders should be wary of identifying false kernels at the level of current core products — for example, film versus chemicals and materials. Frameworks such as VRIO (value, rarity, inimitability, and organization), which invite leaders to consider whether a capability or resource provides sustainable competitive advantage based on if it is valuable, rare, difficult to imitate or substitute, and how well the business organized to exploit it, can help. But the key is to look beyond today’s products and assess value in new contexts through a deep, generative, defensible, and coherent lens. Kernel identification is not a single act of insight. It is an iterative loop of excavation, validation, market testing, and refinement.

Step 3: Lock in the advantage in a new growth engine. A kernel is a strong foundation for growth, but it is not a business on its own. Once leaders have identified a credible kernel, a new growth engine and competitive moat need to be built around it. That may require new partnerships, business models, capabilities, and talent. Ørsted built an offshore wind business around its core capabilities, including developing innovative new project development and financing models, supply chain partnerships, and an operating model designed for scale.

This step often requires a dramatic shift in competitive logic. For example, one of the biggest shifts in Microsoft’s reinvention was to embrace open-source software and ecosystem collaboration with competitors. That made its cloud offering more relevant and credible but was a dramatic break from its more closed, Windows-centric posture of the past. Given that a significant portion of cloud workloads were running open-source software like Linux, Microsoft needed to ensure that the Azure cloud computing offering would be competitive and to regain trust and relevance among developers.

Step 4: Manage the transition. In biology, neofunctionalization often begins with duplication: One gene copy continues the old function while the other is free to adapt under new environmental pressures. Companies need to follow a similar logic: Kernels must be given space to develop and adapt in a new context where they will encounter new competitive pressures and metrics, but they must also maintain ties to the existing business to take advantage of corporate assets. For example, Ørsted created a new wind power business unit but invited the engineers from the fossil fuel business to join the new endeavor.

Reinvention also does not mean abandoning the old business overnight. While startups have the advantage of clean slates, established companies have the benefit of channeling legacy cash flows to fund the pivot. Leaders must manage the overlap: sequencing exits and investments and preparing investors and employees for short-term dips in service of long-term growth. Reinvention fails when the new business is suffocated by the old, but it also fails when leaders cut away the old before the new is strong enough to stand on its own.

Step 5: Craft a narrative of continuity. While leaders may spot the need for change, other stakeholders, including investors, employees, customers, and partners, need to be brought along on the journey. A compelling reinvention story must signal a radical departure from the nonviable past while also maintaining continuity. The story must also be made tangible through visible proof, such as a new product, business unit, market entry, or operating commitment. This is where kernel-based reinvention shines, since there is a genuine connection to something deep that made the company great. Validation from outside also matters: Customers, partners, and analysts telling the new story makes the reinvention real in ways that internal communication alone cannot.

If the story sounds like “everything we were is now obsolete,” stakeholders may resist or disengage. On the other hand, if it sounds too similar to business as usual, the shift may not seem convincing enough. The best narratives make clear what must change and how the company is building from an existing advantage to gain a strong competitive position in the new arena. Done well, this gives stakeholders an opportunity to credibly reframe where the company is heading, on a path grounded on past strengths. Ørsted’s transition worked in part because it was framed as necessary for the company to stay financially and environmentally relevant yet still built on internal strengths honed over decades.6 This also gave employees a path toward transferrable skills in an exciting new sector that aligned with the aspiration of creating a cleaner, more sustainable future.

Finally, it is important to remember that reinvention is not a one-shot, linear path but rather an iterative journey of learning, discovery, and scaling what works.

It is tempting to view reinvention as a choice between defending the core and starting anew. But the strongest reinventions rarely fit either extreme. They begin by separating the legacy business from the deeper capabilities that made it possible. The former may need to shrink, be sold, or disappear. The latter may become the kernels of the next growth story.

Companies that thrive in the face of disruption will not be those that preserve the past intact nor those that discard it wholesale. They will be those that can identify their kernels of reinvention and give them new strategic functions in a changed world. For leaders, the hardest decision is not simply whether to reinvent. It is what to carry forward.