The consulting manager took a call at 7:30 p.m., while volunteering at her son’s soccer practice, from an employee who felt “on the verge of quitting.” Later that same week, she responded to texts sent at 2 a.m. from team members who could not sleep amid corporate restructuring and AI uncertainty. On Sunday, she sent notes of encouragement before the workweek resumed.

This is the reality of a climate in which expectations for leaders to show humanity, compassion, and empathy have intensified. Across industries, employees are feeling stressed, worried about economic headwinds, and unsure how AI will reshape their jobs. Organizational fear has always existed, but it’s becoming more visible as the pace of change accelerates.

Leaders are expected to steady anxious teams, absorb emotional fallout, and respond to employees’ increasing mental health needs. These expectations are redefining leadership roles. Yet the burden is being shared unequally: Women are carrying a disproportionate amount of caring tasks at work, often at the expense of their own well-being.

When we polled more than 350 professional women in managerial roles as part of our research, 81.6% told us they spend at least 30% of their workweek on caring tasks, such as listening to colleagues’ anxieties, offering encouragement, or monitoring how people around them are feeling. That’s more than a business day’s worth of work in a five-day week. Increasingly, such work is no longer incidental. It’s becoming part of how organizations function. This level of emotional labor is equivalent to a part-time job layered on top of a person’s existing formal responsibilities. These findings mirror what we’ve consistently heard in one-on-one interviews and group sessions.

We call this the empathy tax, or care tax: the invisible emotional toll women leaders pay when they shoulder most of an organization’s caring labor. This labor causes care fatigue — exhaustion that stems from constantly absorbing people’s stress, frustration, and anxiety. Care fatigue is rarely discussed in leadership circles, yet many managers recognize it immediately when it’s named. It’s the slow accumulation of small stabilizing acts: calming a worried employee, translating a confusing strategy shift, reassuring a team after another round of change.

To be clear, compassion is a valuable component of leadership; when employees feel seen and supported, that’s a good thing. Compassion has positive organizational impacts, including increasing trust, engagement, and resilience. But when women are expected to shoulder an outsize share of caring work, it undermines their well-being and feeds burnout, exposing companies to higher risks of attrition and disengagement among women in managerial roles.

Care creep — the expansion of emotional and support work that becomes expected but not formally recognized — also tangibly hits organizations as women spend more time on caring work. That’s time that would otherwise be spent on core responsibilities and advancing organizational goals.

An Increasing Burden

The caring burden is clearly growing, according to our research. When we asked women how their time spent on caring tasks had changed since the previous year, 20.1% of respondents said they were spending “much more” time on caring tasks, and 38.8% said “somewhat more.” In other words, nearly 59% of women reported an increase in emotional labor at work. Our findings suggest that, at a time when workplace stress and uncertainty continue to rise, it’s women who are increasingly being called on to absorb the emotional energy of their teams.

People may experience additional or different expectations related to race, ethnicity, and cultural and organizational contexts; this article focuses on the gendered pattern that first prompted our research.  

What about the men? In our early conversations with professionals of all genders, men did not describe feeling the same pressure to provide emotional support. In many cases, men didn’t even observe such work happening around them — whereas women described it as commonplace.

That dynamic shows that emotional labor often goes unnoticed. So to surface its scope and impact, we asked women about the extent to which they were performing emotional care at work. We heard many stories like the one from the consulting manager at the beginning of this article.

Why are the empathy tax and care fatigue hitting women so hard? A large body of research in social psychology and management has found that women are expected to demonstrate warmth and caring in the workplace and are viewed negatively when they fail to do so. Gender norms that associate women with caring, compassion, and warmth are deeply ingrained. A notable 76% of respondents reported that emotional and caring work in their organizations is performed mostly by women, while only 10.6% said it is shared equally and just 1.7% said it falls primarily on men. These findings underscore how deeply gendered expectations shape the distribution of emotional labor, amplifying the pressures on women leaders.

This dynamic isn’t just statistical; it plays out in everyday life. Researcher and author Brené Brown described being stopped by strangers eager to share their most painful and traumatic stories, a dynamic her fellow academic, Adam Grant, said he hasn’t experienced. Despite having similar platforms, they’re expected to demonstrate empathy in very different ways. When one of us shared this example on LinkedIn, dozens of women responded with similar experiences.

Research shows that in occupations where emotional labor is high, women in senior roles report feeling more overwhelmed than their male peers. This dynamic isn’t new, but as the load increases, the labor is spreading. Caring work has long been expected in functions with a high percentage of women, such as human resources and communications. But as societal stress and mental health challenges rise, especially among Generation Z and younger workers, empathy has become a broader organizational imperative and companies are leaning on a larger group of women.

Three Ways the Empathy Tax Shows Up

Here are some of the invisible ways women leaders shoulder emotional labor at work.

1. Absorbing others’ emotions. Gender norms that cast women as naturally warm and attuned to others’ feelings create an expectation, conscious or not, that women will provide support and compassion when colleagues raise concerns or share their challenges. Research has shown that women in managerial roles are acutely aware of these gendered expectations and work to meet them.

These expectations result in female leaders spending significantly more energy listening to others and soothing and managing their emotions, such as stress, worry, and frustration, than their male counterparts do. Some women reported to us that they were expected to have an endless well of emotional availability and the capacity to constantly absorb others’ stories and stresses. The overloaded consulting leader we mentioned earlier said she spent hours holding space for employees after layoff announcements, but her male counterparts weren’t asked to do the same.

This work doesn’t just take time. Whether it’s absorbing stories about difficult or painful experiences, sitting with a crying or angry employee, or being present for someone processing hard news, these types of moments evoke emotions in the listener. It takes a toll on their energy. This work doesn’t end when the conversation does. Leaders carry the emotional residue of these interactions — such as sadness, frustration, and anxiety — with them. As one nonprofit leader put it, “I wasn’t trained in trauma or therapy. I leave these conversations emotionally exhausted and unsure how to set the boundary and not absorb it all.”

2. Getting graded on compassion. Gendered expectations that encourage employees to look to women leaders for emotional support also limit how women can respond and how much space they have to deal with their own feelings. In our respective books, we each interviewed multiple women who spoke about having little room to process their own sadness, stress, fear, or other difficult emotions during times of organizational or societal turmoil.

CEOs may make the formal statement or claim that “the buck stops here” when an unpopular choice is made. But, in practice, they are often insulated from front-line reactions and are rarely expected to engage deeply with employees’ emotions. Managers — particularly women, as our data shows — take on the organizational work of demonstrating empathy, allaying fears, and reassuring employees that they are being heard.

Also, female managers face backlash when they’re seen as insufficiently warm. As a result, women’s performance of caring becomes central to how they are perceived and valued as leaders. They are, in effect, graded on how much and how well they show compassion.

3. Sacrificing time. The different behaviors that make up empathic labor at work, from listening to offering pragmatic support, take not only effort but also time. When women are expected to take on primary responsibility for expressing care within an organization, they must dedicate a meaningful portion of their work hours to meet this demand, our data shows. One problem is that the time women spend filling this role is time taken away from core job responsibilities.

Additionally, many of the women we surveyed said they were often “volunteered” by others on their team to take on caring responsibilities. A finance vice president shared that she was late to a client meeting because an employee became very upset in her office and her colleagues thought it best that she stay until HR arrived.

Like “office housework” or secondary roles such as leading employee resource groups, organizing team morale programs, and mentoring, the emotional caretaker role isn’t formally recognized or rewarded with concrete benefits like higher pay or high-profile assignments. Yet this work plays a critical role in group functioning and supports the common good.

Balancing these caretaker demands with the typical leader’s roster of meetings, emails, and core tasks can quickly lead to overwork. More than a third of our respondents (35.8%) said that the caring work they do in the workplace increases their likelihood of leaving their current role. That is a tangible consequence of the empathy tax for organizations. Moreover, caring tasks limit women’s bandwidth for other leadership work, such as advancing organizational goals and priorities.

The combination of overwork and emotional strain can lead to underperformance, burnout, and, ultimately, attrition — reinforcing the very pressures organizations hope to avoid by calling on women to provide emotional support.

How Women and Organizations Can Address the Empathy Tax

Emotional labor rarely appears in strategy documents or performance metrics, yet it is often the quiet infrastructure that allows organizations to function at all.

For women, the first step is containment. Before reducing your care burden, you need to recognize and reject the harmful narrative that you have to prove your worthiness for leadership by sacrificing your own well-being to meet others’ emotional needs.

Persistent gender norms that expect women to hold endless emotional capacity are the backdrop to this myth, which means that pushing back can be uncomfortable. Remind yourself that your needs — including time to process your own emotions, rest, and pursue your goals — are as important as those of people asking you for care.

This reframe doesn’t mean devaluing care. Indeed, many women we’ve heard from have noted that emotional IQ and warmth are useful leadership traits, even “superpowers” at times. It means deploying your compassion in ways that don’t deplete you.

Next, define and set boundaries that will enable you to feel centered and focused, not drained and scattered in too many directions. These boundaries are going to look different for everyone, depending on role, team dynamics, and personal circumstances. The important part is reflecting on what boundaries will support your success at work and then practicing them consistently.

Remember that your “no” can be a “not now” that respects your time by shifting a conversation to a day or time when you’re not on a deadline. Or your “no” can direct someone to a resource that meets their needs so you can step away. Your boundary might be a clear limit on how long you can talk through a team member’s feelings about a new initiative.

It may help to talk to your peers about their level of care fatigue and collectively reinforce the value of protecting your time and energy. In an organizational culture where many women are taking on a high caring load, one person shifting their approach is swimming against the tide, but a group can create real momentum toward change.

However, the change can’t come just from women: Organizations have their own work to do in taking action to prevent care fatigue from spiraling out of control. First, they should actively assess the extent to which care fatigue is affecting women in leadership roles, whether broadly or in specific pockets of the company. Every manager should be curious about how much time the women on their team — especially women managing others — are spending on caring work. This means both asking directly and observing interactions as they occur.

Managers should also visibly support the boundaries women set in attending to others’ emotional needs — for example, by redirecting any pushback regarding warmth and ensuring that women aren’t positioned as the default emotional resources.

Organizations can also disrupt a culture that leads to the empathy tax and care fatigue in the first place. One approach is rewarding men who step up in a caring capacity. Gender norms often mean that men are criticized for emotional expression or may be seen as weak for exhibiting caring behaviors that are valued in women. By prioritizing empathy and compassion as core leadership qualities for everyone, organizations can reap the benefits of building multifaceted leaders, without placing the burden on women alone. Organizations can help everyone develop a broader leadership toolkit.

Care fatigue is real, and it can derail hard-won career progress for women and for the organizations that want to retain them. The challenge is not whether empathy belongs at work. It is whether organizations are willing to recognize and share the labor required to sustain it. There are ways to address the empathy tax:

• Name it. Acknowledge that the empathy tax and care fatigue exist in your organization.
• Normalize limits. Remove the stigma around a person expressing limits to caring work or seeking support. Leaders should normalize that emotional labor is real and requires intentional management and that caring, empathic leadership is a strength, regardless of someone’s gender.
• Redistribute workload. Design team practices so that all team members share the workload of providing empathy and support. This helps women perform at their best and organizations realize everyone’s potential.

Jean-Christophe Jaunin became CEO of Nespresso North America, the Nestlé unit that sells coffee brewing machines and capsules, on Jan. 1, 2026, after having served as global chief customer and technology officer. At the NYU Stern Center for Sustainable Business’s annual practice forum in March, MIT Sloan Management Review spoke with Jaunin about Nespresso’s commitment to sustainability. This interview has been edited for clarity and length.

How do you make sure sustainability targets don’t get sidelined as the pressure to deliver financial growth intensifies?

Jean-Christophe Jaunin: It’s the foundation of the quality we promise. Every time you’re drinking an espresso from Costa Rica, it will taste like Costa Rica. Yet the inputs — the soil in which the coffee tree grows, the environment — are changing rapidly. Sustainability here means going deeper into taking care of the soil, the climate, the environment in which the coffee grows, so that we can proactively manage change and future-proof our business.

More than 20 years ago, we started to identify the risk that conventional agriculture posed to coffee quality. Traditional farming practices were aimed at maximizing productivity. When mass production of coffee began, the thinking was to get rid of all other plants and just put in coffee trees. What happened is that the soil got poorer and poorer. Poor soil means drier beans, and the whole taste profile suffers. So we started putting back trees to see how a mix of different plants would stabilize the soil. Birds, insects, and other plants come back. This creates a new kind of compost that nourishes the soil, and by enriching the soil, the coffee quality gets better.

How do you convey the value of these changes to farmers who may be used to doing things the traditional way?

Jaunin: We need to create loyalty with them. The more than 150,000 families who are part of our Sustainable Quality Program are independent business owners who joined voluntarily. We have trained more than 600 agronomists to provide farmers with technical assistance and cultivate a direct relationship rather than going through brokers and intermediaries.

With traditional agriculture, if the coffee market was bad, there was nothing else. Now, with biodiversity, they have bananas, they have avocados. A couple of years ago, we partnered with expert beekeepers in Colombia and helped farmers put back beehives. The bees pollinate the coffee [plants], but they also create additional revenue for the farmers through honey. By giving farmers the chance to diversify their revenue, we create a more resilient economic model for them. And that resilience ultimately protects our supply.

Nespresso has taken on significant costs to manage the end of life of its aluminum coffee capsules. How do you justify that?

Jaunin: It is costly, but it’s core to our business model. We made the choice to use aluminum because it lets us vacuum-seal the coffee’s freshness for a very long time. In addition, aluminum can be recycled indefinitely. But because we made the choice to use this material, we need to take care of it.

There are more than 30,000 municipalities in the U.S., so we need to work with local authorities, regional authorities, businesses, recyclers, and composters. There’s the mail-back program, where we prepay the return for customers. In New York, we’ve invested in equipment at a waste management facility in Brooklyn that separates the aluminum from the coffee grounds so customers can simply drop capsules in the regular recycling bin. In Texas, we’re currently testing a pick-up-at-home model: The postal delivery person delivers your coffee and goes back with your empty capsules. It takes time and investment, but we are committed to ensuring 100% of our capsules can be recycled.

Create Generative AI Value at Scale

Kevin Schmitt, Gregory Vial, and Ivo Blohm

Key Insight: Organizations are expanding their GenAI use by implementing coordinated cross-functional structures that draw on domain expertise and user innovation.

Top Takeaways: Companies that establish a new kind of internal AI organization that researchers have dubbed the “AI spine” are better positioned to expand the scope of use cases, continually improve them, and identify the ones that will improve processes and create real value for the business. The spine model facilitates greater sharing of knowledge and innovative ideas across business units by connecting resources — including users and cross-functional experts — to a flexible technical core. Disciplined project governance keeps resources focused on the areas where generative AI is most likely to have a positive impact.

Read the article

Scaling AI With Adaptive Governance

Gianvito Lanzolla, Margherita Pagani, and Christopher L. Tucci

Key Insight: Organizations must implement a new approach to AI governance across a system’s life cycle to manage risks at scale.

Top Takeaways: As organizations adopt AI systems across business functions, they need to manage increasingly complex risks not only during the development process but also after deployment. Leaders should start by identifying the risks their organization faces and the controls needed to manage them. Then, by adopting adaptive AI governance practices, they can continually realign AI with organizational needs as those systems scale. Organizations that embed risk controls into operations, overcome cross-domain barriers, and institutionalize continuous learning and improvement will have an advantage over those that don’t.

Read the article

Why AI Isn’t Transforming Finance Yet

Stijn Viaene, Kristof Stouthuysen, and Bjorn Cumps

Key Insight: CFOs must adapt their leadership approach to balance finance’s traditional role with the use of AI to help shape organizational strategy.

Top Takeaways: Finance offices have been slow to meaningfully adopt artificial intelligence, often due to a narrow perception of the function’s role as a steward of discipline and consistency. When finance leaders and their teams realize how AI can help them stay alert to changes in the business environment, experiment in the course of their work, think differently about the future, and embed new practices in their everyday processes, they will begin to see opportunities for using AI as a tool that supports broader organizational change.

Read the article

Why Businesses Should Experiment With Quantum Computing Now

Avi Goldfarb and Florenta Teodoridis

Key Insight: Quantum’s benefits won’t materialize overnight. Companies that start experimenting today can gain a competitive edge.

Top Takeaways: Companies shouldn’t wait until quantum computing technologies have reached maturity to invest in them. As an enabling technology, quantum requires hands-on experimentation, feedback loops that support incremental learning, and co-invention cycles between producers and users — over time — to identify practical use cases. Investments in quantum today may see near-term payoffs, but the focus should be on active learning and the potential for breakthrough innovations over the longer term.

Read the article

Level Up Your Crisis Management Skills

Rick Aalbers, Killian McCarthy, and Arjan Groen

Key Insight: Leaders can become more adept at responding to crises by developing stronger skills in seven critical practice areas.

Top Takeaways: People who have successfully managed crises in governments and large organizations aren’t innately better at it. They’ve learned to apply critical crisis management practices. Interviews with high-level leaders in a variety of industries found that organizations with strong crisis management capabilities have invested time and effort to develop maturity in seven key areas researchers have dubbed the “7C’s”: contingency planning, cross-functional coordination, transparent communication, compassion, confrontation of hard truths, control, and continuity.

Read the article

Data Transformation Is the CEO’s Business

Barbara Wixom, Ogi Redzic, Brandon Hootman, Joaquin Rodriguez, Gabriele Piccoli, and Cynthia Beath

Key Insight: Caterpillar’s data overhaul shows the essential transformation work that CEOs and senior leaders must commit to for AI readiness.

Top Takeaways: A multiyear data transformation project at Caterpillar that made the heavy-equipment manufacturer AI-ready provides an exemplary case for what leadership commitment to such a technology project involves. CEOs must go beyond communicating abstract intentions by setting a tangible, strategic business goal that the transformation will support; giving teams realistic time horizons and adequate resources; and assigning meaningful, instrumental roles to members of the leadership team.

Read the article

What It Takes to Scale Value-Based Industrial Solutions

Johan Frishammar and Vinit Parida

Key Insight: Manufacturers can successfully build upon value-based sales pilots by using a framework centered on six core capabilities.

Top Takeaways: Industrial equipment manufacturers moving to a value-based sales model often find that delivering initial solutions on a one-off basis is relatively straightforward. The real challenge lies in scaling those solutions to more customers, which requires structured, repeatable processes and strong, entrenched capabilities. New research points to two important phases of capability building — scaling prerequisites and scaling execution — and identifies the organizational skills, processes, and relationships that successful companies assemble.

Read the article

Gain Consumer Insight With Generative AI

Neeraj Arora, Ishita Chakraborty, and Yohei Nishimura

Key Insight: Large language models can transform marketing research by enabling faster concept testing, qualitative research, and data analysis at scale.

Top Takeaways: Typical marketing research efforts can cost tens of thousands of dollars and take months to complete. LLMs are starting to change the industry by compressing timelines from months to days. How? By enabling the development of synthetic consumer “digital twins” for rapid concept testing, the use of AI-moderated interviews for qualitative research at scale, and the ability to conduct powerful analyses of unstructured data. These LLM-based AI tools allow smaller research teams to conduct larger studies while maintaining quality, thus enabling more frequent testing and experimentation.

Read the article

How Leaders Can Move Past Personal Obstacles

Katherine W. Isaacs and Richard C. Schwartz

Key Insight: Leaders can overcome conflicting motivators that hinder their effectiveness by applying psychotherapeutic tools while managing others.

Top Takeaways: Professional growth involves acknowledging and releasing beliefs and behavioral patterns that have been interfering with good decision-making or strong working relationships. A leadership development expert and psychologist explain how simple techniques drawn from the Internal Family Systems psychotherapy approach can help leaders shift persistent attitudes and behaviors through greater self-awareness and cultivate greater compassion, curiosity, clarity, creativity, calmness, confidence, courage, and connectedness.

Read the article

Resolve the Conflict Between Efficiency and Resilience

Vishal Ahuja, Yasin Alan, and Mazhar Arıkan

Key Insight: Fine-tuned buffers and adjustments to performance metrics can strengthen operational resilience without sacrificing efficiency.

Top Takeaways: Studies of the airline industry show that achieving resilience doesn’t have to come at the cost of efficiency. Managers in a variety of industries can meet both objectives by ensuring that operational performance metrics reflect true customer priorities; using predictive analytics and data-driven insights to allocate system buffers where they generate the most meaningful resilience benefits; and shaping the options offered to customers to improve the organization’s resilience to disruptions.

Read the article

It’s safe to say that most people who rise to the top of their companies like to win. A healthy competitive streak is energizing and motivates individuals and teams to do their best — to find their edge and sharpen it. But sustained, long-term success and industry leadership often rely on the ability to look beyond your self-interests and see where, in the bigger picture, your contributions to building up your sector can make your own company stronger.

That theme emerges from several articles in this issue, although in different contexts. We see the benefits of early-stage engagement with emerging technology, the buffering effect of business ecosystem connections in a crisis, and how one company models sharing lessons learned with a community of practitioners.

On the emerging-technology front, Avi Goldfarb and Florenta Teodoridis issue a call to managers to get off the sidelines of quantum computing and contribute to developing practice. The authors remind us that with all enabling technologies — as was the case with the internet and electricity — value is cocreated by early users of the technology. Users’ experiments contribute to feedback loops that identify promising application areas and clarify how the complementary ecosystem around the enabling technology needs to evolve.

Goldfarb and Teodoridis emphasize that in quantum computing, this is a collective effort among organizations that apply the technology and those that aim to supply it. Practitioner involvement now, even though it’s early days, is crucial to the development of the ecosystem that must grow up around this technology to make it broadly usable by businesses. Companies that get involved in and contribute to this development will certainly be helping others, maybe even their competitors, but they will also be demonstrating that they are leaders in their sectors.

Cultivating the ecosystem around your company can also improve your resilience in a crisis. Rick Aalbers, Killian McCarthy, and Arjan Groen interviewed senior leaders across government, multinational companies, and the military to understand what made them successful at crisis management. From their research, they distilled not a set of heroic personal traits but rather a framework of organizational capabilities that those leaders had put in place well before a crisis arose. Those leaders were able to weather storms because they could rely on well-prepared people and well-tested processes when disruption occurred. One subtle but significant element that isn’t part of the framework but does show up in crisis stories: Having good working relationships with other players in the broader business ecosystem clearly gives leaders more options when the chips are down.

Finally, one of the most significant ways to contribute to the health of your business ecosystem is to share your own hard-won lessons. That’s not the subject of our article about how CEOs need to engage deeply with data transformation initiatives. But the article itself is a result of Caterpillar’s willingness to allow researchers (led by Barbara Wixom, principal research scientist at the MIT Center for Information Systems Research) to follow its massive, multiyear transformation in great depth. The company permitted executives involved in the initiative to join the researchers in writing about their experience for MIT SMR in order to share what they learned with the wider practitioner community. That kind of generosity contributes to collective learning via an ongoing conversation among practitioners and scholars that, over time, improves management practice to everyone’s benefit.

Artificial intelligence was supposed to dramatically change the corporate finance function. Forecasts would become more accurate and more frequent. Closing cycles would shorten. Risks would be identified earlier. Scenario planning would evolve from an occasional exercise into a continuous capability. On the basis of those optimistic predictions, many finance leaders have invested heavily in the technology.

However, when CFOs speak in private, a different story emerges. There are proofs of concept that never leave their sandboxes. Models that looked promising in pilot sit unused when the pressure of the quarter hits. Dashboards are produced and refreshed but rarely shape the decisions that matter most. Finance is undeniably busier and more automated but not obviously more forward-leaning in how it helps the organization adapt.

It is tempting to blame the shortfall on technology issues: The data quality is not there yet; the tools are not sufficiently integrated; the models are not trusted; the vendors overpromised. All of those factors matter, but they do not explain why similar AI technologies, introduced under broadly comparable conditions, lead to very different outcomes in finance than in other corporate functions.

After several years of working closely with CFOs and their teams as they tried to apply AI in practice, another explanation became hard to avoid: In many organizations, the technology is moving faster than the way leadership actually works inside the finance function.

When new tools arrive, people tend to talk, decide, and behave much as they did before. Attention gravitates toward getting the close done, explaining variances, defending a single forecast number, and treating deviations as errors to be corrected rather than signals to be explored. AI is introduced into that environment and expected to transform it. Most of the time, it does not.

To understand why, and what might be done differently, it helps to look less at technology adoption and more at leadership practice. (See “The Research.”) In studying the question of AI adoption in finance, we took a simple but demanding view of leadership: Leadership is not a job title or an individual trait; it is the work through which people help their organization adapt under uncertain and changing conditions. In finance, that work shows up every time someone reframes a question, tries a different way of seeing the numbers, surfaces an uncomfortable signal, or helps colleagues adopt a better routine.

Viewed this way, leadership does not sit only with the CFO or a small circle of direct reports. It can be exercised by an analyst who notices something unusual and asks, “What might this mean?”; by a controller who proposes a trial of a different forecasting driver; or by a planning manager who brings several futures into the conversation instead of converging on one.

Leadership becomes visible in the way practices are introduced, tested, and shared. It also becomes visible in who feels empowered to initiate and sustain that work. This view aligns with broader work on digital transformation that frames leadership less as top-down control and more as the orchestration of attention, accountability, and learning across the organization.1

This way of looking at leadership has a sobering implication for AI in finance. If leadership is understood primarily as the CFO’s personal competence or as the formal hierarchy’s right to decide, then the function’s capacity to experiment, learn, and embed new ways of working will always be limited. If leadership is instead understood as shared work in practice, then AI becomes an opportunity to reshape how that work is done. However, the very nature of finance work itself can raise challenges for AI adoption.

When Finance Is Pulled Into a Paradox It Did Not Choose

Finance has always lived with a tension between control and change. Its core mandate is to ensure reliability: accurate reporting, regulatory compliance, and disciplined stewardship of capital. Over time, the function has built processes, controls, and habits designed to reduce surprise. A great deal of finance’s professional ethos is shaped by the imperative of not being caught out.

AI introduces a different dynamic into this environment. It allows finance teams to see more, and earlier. It makes it possible to scan wider sets of signals, test alternative assumptions at low cost, and explore uncertainty in ways that were previously impractical.

The result is that finance is pulled more deeply into a paradox it did not choose. It remains responsible for being the organization’s safe pair of hands while at the same time being asked to become more curious, experimental, imaginative, and adaptive. Finance must protect what is working, even as it helps reinvent what may soon no longer work.

Some finance functions have learned to live with this paradox. They develop ways of working that keep discipline and exploration in constructive tension. Others fall to one side or the other: They either protect the familiar and treat AI as an efficiency add-on, or they embrace every new tool and struggle to make anything stick.

What makes the difference is not simply the tools they buy but the pattern of leadership work that emerges inside the function.

Across many engagements, we saw four recurring activities that particularly mattered for finance teams learning to work with AI: staying alert to what is changing, experimenting in practice, thinking differently about the future, and embedding what proves useful. These are not stages in a process. They are different ways that leadership shows up in everyday finance work. Here, we will present four vignettes, drawn from our research, that show how leadership work around AI takes shape in everyday finance practice. Details have been anonymized and, where necessary, combined to protect confidentiality, but each vignette reflects patterns we observed repeatedly across multiple organizations, rather than isolated or exceptional cases.

When Vigilance Becomes Shared Work

At a European manufacturing company, the central finance team had invested in a sophisticated data platform that provided access to a wide range of external market and supply chain indicators. Over time, the volume of available information increased, but much of it remained in the background. The data was technically accessible, yet it was rarely featured in the conversations that shaped plans or decisions.

That began to change when a financial planning and analysis manager proposed a small adjustment to how the team worked. Each Monday morning, two analysts were asked to bring one external signal they considered potentially important to a short discussion with colleagues. The conversation always started from the same question: “If this were the first sign of something bigger, what might it be?”

AI made it possible for the team to scan a much broader range of signals than before and to narrow that field to a manageable set for discussion. The more significant shift, however, was behavioral. Analysts and controllers began to see paying attention to early signals as part of their everyday responsibilities rather than as a specialist task. Over time, the team’s discussions started to influence budget assumptions and the way scenarios were framed for business partners.

No one on the team described this as leadership — yet it was. What changed was not the technology but the shared responsibility for noticing and interpreting what might be changing around the business. We observed similar practices across multiple finance teams. Where this shared vigilance took hold, AI was experienced as practical support embedded in everyday work rather than as an abstract promise. Where it did not, AI-driven signals tended to remain peripheral: available in dashboards, discussed in isolation, or quietly ignored when core planning routines took over.

When Experimentation Becomes Routine Rather Than the Exception

At an international consumer services company, the finance director had grown skeptical of large transformation projects that promised to reinvent planning with AI and delivered little beyond stress. Instead of launching another program, she encouraged her teams to think in terms of small, bounded trials that would generate insight without putting core processes at risk.

During one quarter, a business unit controller proposed running an explainable forecasting model alongside the standard statistical one. Using machine learning, the model suggested different revenue drivers for a specific product line. The trial was deliberately constrained: It would run for two sprints, it would not influence formal guidance, and success would be assessed not only by accuracy but by what the team learned.

The experiment did not outperform the existing model. What it did produce was a clearer understanding of which data the team actually trusted, where the conventional approach was more robust than expected, and where there might be room to rethink drivers in the next cycle. Crucially, no one was criticized for having wasted time. In the review meeting, the CFO asked a different question: “What did we learn that we could not have learned otherwise?”

That question quietly reset expectations around experimentation. Over time, teams began to treat trials not as projects to justify but as part of how finance learned. This mattered even more when generative AI tools became available.

In many finance functions we observed, the arrival of generative AI initially led to highly individual and largely invisible experimentation. People tried out tools on their own, unsure of what was acceptable, wary of failure, and sometimes concerned about what the technology might imply for their own role. Learning remained fragmented, insights stayed personal, and little of that experimentation translated into changes in shared routines.

In this team, the dynamic unfolded differently. Because experimentation had already been legitimized as collective work, generative tools were folded into the same discipline. Teams openly tested where AI-generated explanations or variance narratives genuinely improved shared understanding and where they merely added fluent but unhelpful noise. The tools were kept in a supporting role, helping teams reflect on results rather than replacing human judgment.

More broadly, we found that where AI remains peripheral, experimentation is treated as a temporary deviation from “real work.” Where AI begins to change practice, experimentation becomes part of how finance operates: disciplined, bounded trials carried out by people close to the business, using AI to learn what works, what does not, and where judgment must remain firmly human.

When the Future Becomes a Subject for Conversation, Not Prediction

At a regional utilities provider, the leadership team had always expected finance to deliver a single forecast that captured where the business was heading. As volatility increased in energy markets, the numbers they generated became harder to defend. After a series of painful forecast misses, the CFO tried a different approach.

Using AI-enabled scenario tools, the planning team constructed a handful of plausible futures for the next three years, each built around different combinations of input prices, regulatory decisions, and customer responses. Instead of producing one projection, finance brought several stories to the executive table, each with its own numbers and early-warning indicators.

The discussion shifted. Rather than debating which forecast was “right,” executives began asking, “What should we do if this scenario starts to materialize?” Finance’s role changed accordingly. It was no longer expected to predict the future with precision but to help the organization think through alternative futures and make deliberate strategic choices in the face of uncertainty.

AI made it easier to construct and analyze scenarios in greater depth. But the leadership move was to make it legitimate for finance to acknowledge uncertainty and to connect that uncertainty to concrete choices.

We observed a very different dynamic in finance teams that remained focused on prediction alone. There, AI was primarily used to refine a single forecast, optimize existing assumptions, and reduce apparent error. Paradoxically, this often increased people’s defensiveness. When models were challenged, teams responded by tightening assumptions rather than widening the conversation. Uncertainty was compressed into confidence intervals, and alternative futures were treated as distractions rather than inputs to decision-making. AI made the forecast more sophisticated but not more useful.

When Good Ideas Are Helped to Spread

At a large retail group, a finance team in one country had developed an AI-assisted routine for identifying unusual patterns in store-level expenses. It helped the team distinguish more quickly between genuine issues and benign anomalies. For some time, this remained a local success, known mainly to the people directly involved.

The practice began to travel when the organization created space for teams to share how they worked, not just what they delivered. During an internal learning session, a finance manager from another country heard about the routine and asked for the code and the checklist the team was using. Within weeks, a slightly adapted version was running in two more countries.

Crucially, this early diffusion was noticed and reinforced. Several months later, someone in the head office suggested integrating the approach into the retail group’s standard monthly review, with a small number of governance guardrails agreed to jointly with internal audit. That move did not mandate adoption, but it signaled that the practice was legitimate, safe to reuse, and worth building on.

Nothing in this process required a major program. The AI component itself was neither complex nor revolutionary. What mattered was that people saw it as acceptable to borrow and adapt one another’s ways of working, and that senior leaders took an active interest in how a good idea could become normal practice rather than remain a local innovation.

The Quiet Power of the CFO

Across all of these patterns, including shared vigilance, routine experimentation, strategic direction setting, and the spread of local practices, the CFO’s influence is both indirect and decisive.

Some of it is visible: The CFO decides where to invest, which initiatives to sponsor, which skills to hire for, which projects to stop. These choices shape what is possible. But another part of the role is less obvious and, in the context of AI, just as important: the tone they set about what counts as “real work” in finance.

When a CFO consistently asks only about accuracy and speed, people learn that the safest way to succeed is to avoid anything that might introduce uncertainty. When a CFO shows interest in what teams are learning from experiments, or in which weak signals might matter, people learn that thinking and trying are also part of their job. When a CFO insists that every pilot demonstrate a clear return on investment before it starts, experimentation dies in the planning stage. When a CFO is prepared to back a modest trial to see what happens, even if the payoff is not guaranteed, experimentation becomes possible.

Leadership in this sense is not dramatic. It is expressed in questions, in the allocation of a little time here and there, and in the willingness to protect a practice that is still fragile. Over time, those small acts accumulate. They determine whether AI finds a place in the real routines of the finance function or remains stuck in presentations and proofs of concept.

What the Numbers Can Reveal, if Read Differently

Our earlier work on digital finance maturity examined which organizational conditions tend to be present when finance teams make sustained progress with AI, such as ongoing experimentation and clear accountability for decisions supported by models.2 Our follow-up work showed why many teams nevertheless stall as AI expands: Those conditions often fade in day-to-day routines, even when the technology itself continues to improve.3

The evidence in this article adds another layer. It shows how those same conditions are not abstract capabilities but rather the result of everyday leadership work inside the finance function. Whether experimentation, accountability, or learning persists depends less on formal design choices than on how people are encouraged, protected, and listened to in practice — and on how those same people, in their everyday work, help shape and carry new practices forward. The constraint, in other words, is how leadership work takes shape.

Finance leaders are used to interpreting diagnostic surveys on digital maturity or AI readiness at face value. Read that way, such data shows where functions have invested, which tools they have adopted, and how they assess their own progress.

There is another way to read the same data: as indirect evidence of where leadership work is, and is not, taking place.

When external data rarely influences planning discussions, the issue may not be data availability but whether shared vigilance is part of everyday work. When many organizations report running pilots but few of them see changes in core forecasting or planning routines, the problem may lie less in experimentation than in the ability to turn local learning into shared practice. And when finance professionals say they have access to advanced tools but hesitate to surface the uncertainties those tools reveal, the constraint is often one of leadership tone and permission rather than technical capability.

It is important not to overstate what such numbers can prove. They do not establish causal relationships between specific leadership behaviors and AI outcomes — but they do reveal patterns. Read carefully, they make it plausible that differences in how leadership shows up in everyday practice help explain why similar technologies produce very different results across finance functions.

Changing How Leadership Works Inside Finance

AI will continue to advance. Tools will become more accessible. Vendors will refine their offers. Regulatory expectations will grow. None of that guarantees that finance will transform. What will matter is whether finance functions change how leadership works.

If leadership remains concentrated in a few places and focused primarily on protecting existing routines, AI will mostly be used to continue old ways of working, only faster. If leadership is understood as shared work on practice, such as watching the horizon, trying things in small ways, shaping strategy under uncertainty, and helping good practices spread, then AI can become a powerful ally in reshaping what finance does and how it contributes.

CFOs do not need to brand any of this as a new leadership model. They do not need to talk about frameworks at all. They do need to ask themselves some plain questions:

• Who in our function feels responsible for noticing and discussing early signals?
• Where do people feel safe experimenting with new ways of forecasting or analyzing risk?
• How do we encourage thinking about alternative futures rather than defending a single number?
• What do we do to ensure that local innovations become shared practice when they prove their worth?
• What am I, as CFO, doing day to day to make those forms of leadership more likely, or less?

The latter question is perhaps the most important. But until the answers to all of those questions begin to change, AI will keep pressing against the glass of the finance function without fully entering the room. When they do change, the technology that once felt like an external pressure will start to feel more like an instrument that teams can pick up and use as they learn together how to work differently.

Leaders with even a cursory understanding of artificial intelligence know that while the technology can help them improve productivity and capture new opportunities, it can also expose their organization to many risks. Those with a bit more knowledge are aware that surfacing and mitigating those risks requires adopting responsible AI practices. And leaders who are scaling an AI implementation within their organization will quickly realize that ad hoc attention to those practices is inadequate and that they need to develop the capacity to systematically govern AI at scale.

But building that capacity is proving far harder than most executives expect. They know what they need to accomplish; frameworks from governments and regulators define important guardrails and principles, such as transparency, fairness, and accountability.1 But to implement controls and principles into day-to-day workflows and decision-making, organizations must rethink AI governance. They must frame that task not as a compliance obligation but as a strategic, adaptive capability that evolves as AI systems scale, use cases expand, and risks shift over time.

In this article, we will share how leading organizations are doing exactly that. We will also introduce an approach to adaptive AI governance built on two principles: matching governance controls to the type of AI system and risk involved, and embedding those controls directly into workflows, decision rights, and accountability structures.

The Fundamentals of AI Risk

To design effective AI governance, leaders must first understand the multiple ways in which AI can fail and the corresponding risks. The nature and severity of these risks depend on the type of system, its level of autonomy, and the scope of domains affected by its decisions. The central challenge, therefore, is to design controls that anticipate how risks will emerge and that evolve as AI systems operate. Even as conditions, inputs, and expectations change, AI must remain reliable, safe, and aligned with an organization’s values and goals.

In practice, most AI risks emerge at two moments that require very different governance responses: during development and after deployment. Development risks include using biased or incomplete training data, failing to adequately align the model to the task requirements, and following inadequate validation processes. For example, an early credit-limit-increase model at a bank we studied demonstrated that small input changes could lead to unexpected decision shifts.

Deployment risks arise when models interact with dynamic environments and human operators: Sustaining legitimacy, judgment, and accountability once AI systems are operating at scale in real time is a central challenge. Over time, model quality may degrade as the statistical properties of input data change over time, a phenomenon termed data drift. A model may generate plausible but false outputs or be overly trusted by users who lack the means to detect errors. At Nasdaq, AI-driven market-surveillance systems monitor trading activity for suspicious patterns, generating hundreds of alerts per second. Those systems may fail to accurately flag activity, however, because the boundary between abnormal and illicit behavior is often hard to spot; illegitimate behavior may be deliberately designed to pass as compliant by exploiting model learning patterns.

Fit-for-Purpose Controls

The kinds of controls employed depend not only on when risks arise in the AI life cycle but also on what kind of AI system is involved and how widely its decisions propagate. Artificial intelligence systems can be broadly divided into two categories: those based on bounded-learning (or static) models and those that learn and adapt in deployment. (See “Controls in Adaptive AI Governance Systems.”)

Bounded-learning systems operate within a fixed set of rules and parameters. Optimizing how those rules are applied, rather than changing them, is what improves their performance. Credit-scoring models, for example, refine risk estimates based on income or payment history, but they do not alter how those variables relate to one another. Many generative AI models are “pretrained” (static) and do not get updated during use. Contrast that with adaptive learning systems, which evolve by incorporating production data into their training data and by updating internal representations and relationships between variables. Algorithmic trading platforms and dynamic fraud-detection systems illustrate this approach.

Just as salient to the type of control required is the scope of domains affected by AI decisions, shown on the vertical axis of the figure “Controls in Adaptive AI Governance Systems.” This dimension determines how far and how fast risks can travel once a system goes wrong. At one extreme are narrow-scope systems, where errors remain contained within a specific function or task (such as detecting anomalies within a single transaction stream). At the other extreme are wide-scope systems that shape outcomes across multiple functions, geographies, or even industries, such as cross-border supply-chain optimization platforms. The difference is not incremental but exponential: As system reach expands, small errors interact, propagate, and amplify into second-order effects.

Based on our typology of AI systems, we believe that rules-based controls provide the baseline safeguards for all narrow, static AI systems. When such static systems operate at a wider scope, additional propagation-risk controls must be layered on to address broader downstream effects.

For adaptive learning systems, baseline safeguards remain necessary but must be complemented by ex post alignment controls, particularly those focused on explainability and legitimacy. When adaptive systems also have a wide scope, they require the most comprehensive approach: integrated controls that combine baseline rules-based measures with propagation risk management and alignment mechanisms. Let’s take a closer look at how each of them works in practice.

Rules-Based Controls

Rules-based controls are designed to prevent and correct errors in systems that operate within clearly defined parameters. They are particularly effective in narrow decision domains where logic is explicit and outcomes auditable, such as credit scoring, fraud detection, or the use of customer service chatbots. Rules-based controls embed relevant norms (such as ethical guidelines and industry standards) and compliance requirements into models, using them as design constraints. Rules-based controls also include processes such as validation testing or anomaly monitoring.

Consider the credit-limit-increase decision model mentioned earlier. A senior AI leader at the bank explained that it uses a statistical model rather than deep learning so that decisions remain interpretable. Before deploying a new model, the analytics team produces documentation called a model card that covers three aspects of AI risk management. First, data checks indicate whether the training data is complete, recent, and balanced and how the team will detect data drift over time. Next, decision logic and edge cases are checked to see how scores translate into approve/deny decisions; this includes explicit analysis of thresholds where a customer tips from no increase to an increase, so that customers in the “gray zone” are not unfairly treated. Finally, bias and discrimination tests are undertaken to check that the model does not overfit to particular customer profiles or systematically disadvantage certain groups.

The model card undergoes quality-assurance review by an independent model-risk unit, with input from credit-domain and regulatory experts. Internal auditing later verifies that these steps were followed. Only then does the model go live.

Human judgment is central even in rules-based settings. In one organization, each new lending model for midmarket clients underwent sample testing before deployment. Risk teams selected 100 existing client files and ran them through the model. Relationship managers then compared the model’s recommended lending decisions with their own assessments. Where recommendations diverged, the model team investigated whether the model had uncovered a genuine insight or was overfitting to idiosyncrasies in the data. Only once the sample review showed an acceptable level of alignment between model outputs and the judgments of the domain experts involved — and the sources of disagreement were understood — did the bank approve the model for live use. After launch, periodic sample reviews continued as part of the standard risk-and-control cycle.

Rules-based controls are effective because they make critical decision boundaries explicit, reviewable, and contestable across domains. They are adaptive because they can be recalibrated over time. Divergences between model outputs and expert judgment are treated as learning signals, feeding back into updated model thresholds, assumptions, and review routines as data, models, and decision contexts evolve.

Ex Post Alignment

The complexity of advanced AI systems, particularly those based on deep neural networks, render traditional traceability and explainability methods less effective. Rules-based controls depend on the ability to specify decision logic, yet that logic becomes increasingly opaque as models grow more complex. As a result, stakeholders must ensure that outcomes remain reliable, fair, and aligned with organizational and regulatory expectations. When such systems operate with significant autonomy, this need for explainability becomes especially critical, since decisions may be made and acted upon without immediate human review. Generative AI introduces an additional layer of difficulty because of its stochastic behavior, where the same prompt may yield different outputs.

This is where ex post alignment controls come in. They reveal not how a decision was made but whether its outcomes remain legitimate. They assess AI decisions against ethical, regulatory, and domain-specific standards. While some techniques carry over from rules-based approaches, the emphasis shifts from preventing errors upfront to detecting misalignment as systems operate, learn, and scale.

Organizations operationalize ex post alignment through layered evaluation processes that test outcomes against reference expectations. Microsoft, for instance, has developed a structured evaluation pipeline in which high-stakes models are assessed against libraries of expert-defined policies — such as what constitutes a “fair” or “acceptable” outcome. Evaluators annotate model outputs against these policies, while independent reviewers validate where the system falls short. In some cases, these evaluations can be partially automated — for example, when AI systems are continuously assessed against predefined policy benchmarks, fairness constraints, or risk thresholds, with automated monitors flagging deviations for human review.

This is why algorithmic auditing is a critical component of ex post alignment. After its deployment, a model’s behavior is systematically examined to detect hidden risks, evaluate fairness and performance across affected groups, and verify that outcomes align with organizational policy and ethical standards. Auditing proceeds in two steps. First, auditors identify plausible failure scenarios and define the full use case, including whom the system serves, who is affected by its decisions, and for what purpose it operates. They then monitor these risks by assessing decision outputs, input data, and internal logic against predefined criteria. This process helps organizations surface unintended consequences, such as disparate impacts; document recurring risk patterns; and trigger corrective action before harm proliferates. Frameworks for auditing algorithmic risk, such as those articulated in Cathy O’Neil’s work on auditing AI systems, provide practical tools and metrics to operationalize this approach and strengthen accountability.2 In this way, auditing functions as both a diagnostic mechanism and a foundation for continuous improvement.

A key part of ex post alignment is ensuring that people do not treat AI outputs as unquestionable truths. Because many AI recommendations are inherently probabilistic, organizations need to train users to interpret them as informed signals rather than final decisions. Helping managers understand when to rely on the system, when to challenge it, and how to spot unexpected or biased outputs is essential for keeping AI use legitimate, accountable, and aligned with organizational values over time.

Managing misalignment at scale can be a particular challenge, especially for systems that are designed to filter, prioritize, and escalate alerts in real time. Nasdaq’s AI-driven market surveillance, for example, monitors trading activity for irregularities — such as unusual volumes, price anomalies, or potential manipulation — and can generate hundreds of high-risk alerts per second. Cross-functional teams of compliance officers, data scientists, and domain experts review flagged activity through structured case workflows. Each alert is assessed to determine whether it reflects genuine market manipulation or a false positive triggered by unusual but legitimate trading behavior. Investigators document the rationale for their conclusions, and these outcomes are fed back to model developers to recalibrate thresholds, refine detection features, and reduce recurring noise in future alerts.

Escalation committees intervene when investigations suggest the involvement of coordinated bad actors or when anomalies indicate broader systemic risk. Audit trails capture key elements of this process, including the original alert, supporting data signals, the human decision taken, and any subsequent model adjustments made. Periodic governance reviews are conducted to evaluate patterns of false positives and missed detections to ensure accountability, regulatory compliance, and continuous improvement of surveillance rules. Even so, surges in alert volumes can place severe strain on teams, overwhelming response capacity and increasing the risk of error.

One effective approach to managing the impact of high volumes of alerts is to redesign workflows around AI outputs. This approach is well illustrated by a global bank’s experience with AI-driven fraud detection. Executives found that the main challenges did not stem from errors in the model predictions but from breakdowns in how fraud alerts were interpreted, routed, and acted upon across teams. Inconsistent handoffs between compliance, risk, and front-line staff members often led to delayed responses, duplicated effort, or missed follow-up, undermining the system’s effectiveness despite technically sound outputs. For example, alerts were sometimes routed to the wrong team, duplicated across units, or left unresolved because no group clearly owned the next step. Customer service employees occasionally contacted clients based on alerts that fraud teams had not yet validated, while high-risk cases were delayed because escalation criteria were unclear.

To address those problems, the bank mapped the alert workflow step-by-step and reassigned responsibilities at each decision point. Fraud analysts were given clearer authority to close low-confidence alerts, fraud operations focused on rapid escalation of confirmed cases, and customer service teams were engaged after a fraud review determined that outreach was necessary. Decision rules were standardized — for instance, when an alert should be suppressed, investigated further, or escalated — reducing delays, unnecessary escalations, and alert overload.

Ex post alignment focuses on evaluating AI decisions after they have been made, by testing outcomes against ethical, regulatory, and domain-specific expectations rather than reconstructing internal decision logic. Ultimately, successful ex post alignment does not eliminate risk; it sustains legitimacy by ensuring that high-agency AI outcomes remain contestable, correctable, and aligned with the standards that matter over time. Unlike traditional risk management, ex post alignment accepts that some misalignment is inevitable — and focuses governance on detection, contestability, and correction rather than prevention alone.

Propagation-Risk Controls

Rules-based controls and ex post alignment mechanisms share an important limitation: They tend to treat risk as largely confined, focusing on discrete errors or individual outputs. This approach can be effective when AI systems operate in relative isolation, but it produces incomplete outcomes when systems are interconnected through real-time data flows, APIs, and automated decision-making. The rise of agentic AI is a case in point. As AI systems increasingly initiate actions autonomously, coordinate with other systems, and pursue objectives across multiple domains, errors or misalignments originating in one system can propagate across others. The relevant concern, therefore, is interdependence and propagation risks that can have downstream effects that traditional, output-focused controls may overlook.

Regulators are increasingly recognizing the importance of propagation risks and the need for robust testing and oversight. The Bank of England, for example, has highlighted the risks posed by “deep trading agents” — AI-driven strategies that could amplify external shocks or coordinate in ways that evade human detection. In health care, biased diagnostic models can spread flawed heuristics across hospitals and insurers. In supply chains, algorithmic procurement platforms can amplify pricing errors across entire supplier networks. Similar dynamics can arise in any digitally interconnected system.

Propagation-risk controls represent a third layer of governance and are designed to surface second- and higher-order effects before they overwhelm downstream functions. In our framework, rules-based controls safeguard narrow and relatively static processes, alignment mechanisms address complex systems whose decisions are opaque, and propagation controls focus on interconnected systems. These controls are concerned with not only what happens within a system but what occurs when systems interact. Their central challenge is invisibility: Failures travel laterally, exploiting hidden interdependencies that often become apparent only when a disruption occurs. A minor logistics API error, for example, may be harmless in isolation, but when it is combined with a cyber incident affecting a payment gateway, it can contribute to systemic breakdown.

A governance framework built around a company-​centric view of risk is poorly suited to track such cross-boundary dynamics. Because propagation risks unfold across interconnected systems, often beyond the visibility or control of any single organization, managing them requires a shift from a company-centric perspective to an ecosystem-aware perspective.

This shift involves three complementary activities: mapping interdependencies, monitoring shared infrastructures, and institutionalizing anticipatory oversight. Together, these practices help surface risks that remain invisible when controls focus only on isolated systems or individual outputs. The European Central Bank’s sectorwide cyber-resilience stress tests show how ecosystem-level propagation-risk controls can be enacted. These exercises map interdependencies across clearinghouses, payment systems, and financial institutions; monitor shared infrastructures for cross-organization vulnerabilities; and simulate how localized disruptions could cascade through the financial system. These practices generalize beyond regulation to any highly interconnected environment.

Organizations can enact propagation-risk controls by redistributing visibility, accountability, and decision rights across ecosystems rather than relying solely on organization-level rules or ex post interventions. Because propagation risks are inherently cross-boundary, effective governance depends as much on coordination across organizations as it does on internal control. Some organizations must shift their cultural norms to encourage data sharing, coordination on standards, and co-investment in oversight infrastructures with partners, competitors, regulators, and, in some cases, open-source communities.3 Reducing propagation risks requires the understanding that resilience is no longer something a company can achieve on its own but instead is a property of the broader system it depends on.

As ecosystems become more densely interconnected, these risks are likely to intensify. The rise of agentic AI — capable of autonomously initiating transactions, negotiating contracts, or reallocating resources across networks — extends this logic, increasing both the speed and reach of failure propagation. In finance, logistics, and health care alike, errors may not simply spread; they may increasingly do so with limited human oversight.

Implementing Adaptive AI Governance

Once leaders have identified the AI risks that are salient to their organizations, and the corresponding controls that they need to have in place, the challenge is to integrate those controls into processes and systems, working within them and continuously adapting them. Doing so involves three key practices: embedding controls into workflows and incentives, building cross-domain fluency, and institutionalizing governance as a living learning system. Here is how to do that.

1. Embed risk-control protocols into operations. Risk protocols must be designed and hardwired into workflows, accountability structures, and incentives. Oversight should flow directly into planning, audits, and leadership reviews rather than sitting on a separate compliance layer. Only when governance becomes part of the operating fabric can AI be scaled with confidence. This is a necessary condition.

A global bank whose leaders we interviewed embedded AI controls into its standard lending workflow rather than treating them as a separate compliance step. For each approved AI use case, the bank’s AI use‑case committee documented (1) the risk tier (high, medium, or low) based on customer impact, regulatory impact, data sensitivity, and model type; (2) the mandatory controls associated with that tier (such as independent model validation, sample testing by relationship managers, or frequency of post‑deployment reviews); and (3) the decision rights (who could approve model changes and under what conditions). These requirements were then encoded directly into the credit‑approval process and systems. Relationship managers could not bypass model‑validation steps or deployment reviews; exceptions required explicit sign‑off from both the business and risk management teams. Oversight surfaced in regular decision-making cycles, not through ad hoc committees or audits.

2. Enable conclusive judgment across heterogeneous expertise and risk profiles. Adaptive AI governance does not require consensus or shared judgment. Quite the opposite: It requires mechanisms that enable conclusive judgment across heterogeneous expertise, methods, and risk profiles. This is often the hardest — and most decisive — task to accomplish. As AI risks shift across categories and cut through organizational silos, accountability cannot reside within any single function. Differences across domains are not a flaw but a feature: They reflect distinct expertise, evaluative methods, and risk tolerances. The governance challenge is therefore not to homogenize these perspectives but to create the conditions under which organizations can translate them into conclusive decisions at scale — while avoiding both judgment homogenization and uncritical rubber-stamping of AI outputs.

Among the central challenges to institutionalizing a durable capacity for conclusive judgment are that rules-based controls are often undermined by siloed knowledge when various domain experts do not share a common frame. To overcome those barriers, share knowledge across domains via joint model reviews and documentation (such as the model cards described earlier), and hold routine cross-functional validation sessions that make decision logic, assumptions, and thresholds explicit and contestable. In ex post alignment controls, the challenge involves not only knowledge silos but also misaligned risk tolerance and methodological approaches. Alignment can break down when different teams operate with different implicit risk thresholds — stopping judgment too early on the one hand or falling into analysis paralysis on the other. Relying on divergent methods to reconcile expected outcomes with observed results (such as analytical validation, controlled experiments, or case-based judgment) can also cause misalignment. In such scenarios, disagreement is not simply about what the model recommends but about how much risk is acceptable and what constitutes sufficient evidence that the model is performing as intended.

A critical response, therefore, is not merely to “build trust” in AI recommendations but to establish shared evaluative routines that surface and reconcile differences in both risk tolerances and methodological approaches. Systematic post-deployment evaluations anchor discussions in observed system behavior rather than abstract beliefs about model quality.

Organizations can do this through structured review routines that combine incident and near-miss analysis, performance-drift monitoring, and explicit comparisons between intended use cases and actual decision outcomes. Crucially, these routines create common reference points — agreed-upon risk thresholds, shared evidentiary standards, and comparable metrics — through which analytically driven teams, experimentation-oriented groups, and use-case owners can jointly assess whether the model is functioning as intended. Over time, this enables assumptions, thresholds, and controls to be recalibrated, reducing both premature shutdowns driven by excessive caution and analysis paralysis driven by methodological disagreement.

Propagation-risk control depends on a fundamental shift in mindset: from treating risk as a company-centric problem to governing it as an ecosystem-level phenomenon. As with digital business ecosystems, risks in AI systems propagate unevenly across actors that have different roles, incentives, and degrees of interdependence.4 Mapping these interdependencies beyond company boundaries is a necessary first step — and often a wake-up call — but it is insufficient on its own.

As research on ecosystem strategy has shown, coordination breaks down when accountability is diffuse, incentives remain locally optimized, and no actor is explicitly responsible for orchestrating cross-boundary trade-offs.5 Similar dynamics undermine AI propagation-risk controls. Teams remain incentivized to focus narrowly on their own systems; risk ownership is fragmented across organizational units and external partners; and downstream or reputational risks are treated as someone else’s responsibility.

Without leadership support for ecosystem-level accountability — and governance mechanisms that differentiate risk ownership by type of interdependence — interdependency mapping risks becoming a one-off analytical exercise rather than a sustained governance capability. An ecosystem mindset requires not only visibility into connections but also shared rules of engagement, escalation rights, and decision authority to manage how risks propagate across organizational and technological boundaries over time.

Overcoming these barriers is essential to creating the conditions for conclusive judgment that respects differences in expertise, methods, and risk tolerance rather than collapsing them into a single acritical evaluative frame.

3. Institutionalize governance as a learning system. AI governance cannot be static: Risks mutate, so controls must evolve.6 Effective governance therefore requires organizations to establish learning loops with clear roles, for capturing lessons from incidents and near-misses and for translating those lessons into updated standards, thresholds, and controls.

Rather than relying solely on controls, systems, or large-scale governance platforms, effective adaptive AI governance depends on building the right mindset and embedding practical learning loops into everyday oversight. This involves assigning explicit responsibility for reviewing incidents and near misses; systematically documenting what went wrong; and ensuring that insights are translated into revised policies, recalibrated thresholds, or strengthened controls. Over time, governance shifts from protocols and systems toward institutionalized continuous improvement, ensuring that AI systems remain aligned with organizational intent as models evolve, contexts shift, and new risks emerge.

Taken together, these steps mark a fundamental shift in governance of AI. Adaptive AI governance is not about multiplying controls, committees, or checklists. It is about identifying fit-for-purpose controls and hardwiring them into how the organization works, decides, and learns — into workflows and incentives, shared frames of judgment, and living systems that continuously absorb and act on experience. Organizations that treat governance as static will inevitably fall behind systems that learn, adapt, and propagate risk in real time. In contrast, organizations that institutionalize governance as a learning capability — one that connects strategy, execution, and oversight — can turn AI governance from a constraint into an enabler of scale. In the age of intelligent systems, advantage will come not from adopting AI faster but from governing it better — by embedding oversight where decisions are made, risks propagate, and value is created.

Generative AI presents an organizational puzzle. Businesses have collectively invested billions of dollars to give employees access to general-purpose large language models (LLMs) to enhance personal productivity while in most cases struggling to develop and adopt more strategic applications of the technology. Meaningful return on investment, much less competitive advantage, is likely to remain elusive unless companies can use GenAI to make innovative process improvements that scale across functions and business units.1

Our interviews with 87 practitioners in 23 large organizations revealed that leaders who scale value creation with generative AI cultivate three key practices. First, they expand the scope of use cases across processes rather than remaining focused on a specific task. Second, they treat each use case as a work in progress to be continually improved. And third, they quickly identify and abandon use cases that fail to bring measurable value to the organization.

However, most traditional companies are not structured to institutionalize these three practices. Many operate as multidivisional organizations characterized by multiple profit and loss units with duplicated functions, limited cross-functional information flow, and internal competition for resources.2 This setup makes it difficult to scale generative AI use cases across processes and units.

In our research, we found that the few leaders who are overcoming these challenges are moving beyond the classical hub-and-spoke models that many organizations have used to connect centralized AI technical expertise to each unit. They are developing a new kind of internal resource that we call the AI spine. It provides a flexible core structure for implementing, evolving, and abandoning LLM use cases at scale, keeping the generative AI portfolio both focused and current. Notably, rather than deploying technologists out into business units, as is commonly done, this structure pulls individuals with domain knowledge of business processes into the core and makes them part of the team.

A retail bank that we studied demonstrates the kind of scaling and value creation that an AI spine supports. Initially, the bank’s AI spine spearheaded the implementation of an email assistant for customer service employees. In an early, limited rollout, those using the assistant collectively saved about 700 hours. Once the tool was put into wider use, it reduced email handling time by 15%, allowing employees to dedicate more time to managing complex cases. Encouraged by that success, the spine oversaw the implementation of LLM-powered email thread summaries, call transcriptions, and analyses. That yielded data that provided new insights for customer service employees, leading the bank to start reengineering its approach to customer relationship management. The data was also used to develop the next iterations of the email assistant.

We observed another example at a medical coding company that applies standardized alphanumeric codes to documentation related to diagnoses, treatments, and procedures for the purposes of insurance and health care management. There, the AI spine provided a structure that allowed it to turn its first LLM application for automated coding into a new line of business. An internal LLM application had reduced coding time from 25 minutes per case to 2 seconds, cutting the cost of coding by 60% compared with having humans doing the work. (Affected staff members were able to take on other responsibilities.) The AI spine was able to build out the application into a product for insurance companies that need to verify whether bills have been correctly coded, increasing the company’s reach into the medical insurance market and creating a new revenue stream.

Connecting AI Efforts Across the Enterprise

The AI spine is a cross-functional backbone that is dedicated to diffusing and scaling LLM use cases across the organization, focusing on reducing duplicative efforts and achieving economies of scale as solutions expand across business units and processes. Because it is a central point for collaboration between technologists and those holding business domain knowledge, the spine holds the expertise required for rewiring and continually improving processes end to end and across functions, as in the bank example above. As we found at the medical coding company, this structure can be implemented not only in large organizations but also in small and medium-sized enterprises.

In the cases we observed, funding for the AI spine was allocated by top management, and the spine also got a cut of increased revenue or costs savings resulting from applications deployed. That mechanism creates the right incentives: It forces the organization to measure ROI, stay focused, and avoid disproportionate spending on “convenience” use cases that seem useful but don’t materially affect costs or revenue. By being independently funded, the spine maintains decision-making autonomy vis-à-vis other divisions so it can identify and encourage use cases that have the potential to improve processes cutting across divisions. A more traditional AI center of excellence with a hub-and-spoke structure is more likely to focus on cases within individual business units rather than across them.

The spine is overseen by a C-suite leader, who keeps its efforts aligned with overall strategic objectives; this may be a chief technology officer or chief digital officer, or one of their direct subordinates. Sitting within the structure are AI developers, risk and compliance personnel, and a technology owner. (See “The AI Spine.”)

The technology owner is responsible for preventing the fragmentation of data flows and tools. They typically oversee the creation of centralized data platforms, prompt libraries, models, and evaluation and technical performance metrics (including token consumption and financial costs). This oversight reduces rework and lowers marginal costs as applications are diffused across the organization.

To ensure that risk and compliance are managed throughout each iteration of a use case, the person managing those issues is a permanent fixture in the spine. Making that a permanent role expedites compliance issues and centralizes organizational learning and memory so they can be rapidly applied to other use cases. This means that governance is less ad hoc than it is within other structures for organizing AI work, where a general risk and compliance group sits outside and is often overwhelmed by issues specific to AI.

Sitting within the business units but working closely with the AI spine are business owners, knowledge owners, and designated end users. Business owners — that is, the heads of the business units working with the spine on use cases — are the primary bridge between their unit and the spine. They are responsible for identifying processes where they can make a good use case for GenAI and determining nontechnical baselines and targets (for example, star ratings assigned by end users, where a minimum 3 out of 5 stars would be the goal). They are also accountable for implementing proven use cases for the identified processes and making sure that use cases that underdeliver are dropped. This ensures that proofs of concept don’t stall out or remain confined to a specific business unit. The business owner also enlists representatives from each unit to serve as knowledge owners, and end users who can deliver feedback, to ensure that new applications create real value for the business.

Knowledge owners work with the technology owner to ensure that both explicit and tacit knowledge critical to the use case are understood by technologists and captured by the application. They curate ground truth and provide important feedback on the generative AI components that will become de facto repositories for organizational knowledge. Prompt libraries, for example, need to be current with company policies and safety guardrails to generate appropriate responses. Knowledge owners are able to quickly identify issues within their domain, reducing rework and costly escalations as GenAI use cases are implemented and evolve. They also help to rein in GenAI work when a tool’s performance is strong enough from a business perspective, regardless of technical performance. In other words, a solution may not be technically perfect but still good enough that employees find that it makes their work easier.

Select end users of generative AI solutions work with the spine to validate that a use case will be helpful to its intended users. This is important because underperforming use cases lead to low adoption rates, workarounds, and, in some cases, shadow AI. End users also provide critical feedback on post-implementation performance, such as identifying edge cases. Their input drives the next iteration of application improvements.

The AI spine serves as a central point for short standup meetings that are held regularly, perhaps every other week. At the bank, spine members use the meeting to share lessons learned, discuss ongoing and future priorities, and communicate changes that have been made to knowledge repositories, such as prompt libraries. They hold a summit event every month and invite end users to keep them updated with progress, show them demos, and gather their feedback. These events capture needed adjustments during the development process so that applications don’t have to be reworked after they go live. They also help manage users’ expectations and apprehensions.

Pulling the five roles above and the AI developers into a self-contained, highly connected structure helps to effectively align stakeholder interests, ensure that salient knowledge in the organization is contributed where it will have an impact, facilitate collaboration across functions, and more effectively diffuse knowledge and value across business units. Those activities support the key GenAI scaling practices we introduced at the beginning of this article: selecting use cases for implementation; gradually expanding their scope across processes; continuously improving each use case; and taking rapid action on underperforming use cases. The examples below show how the AI spine works.

Respecting Domain Experience and Expertise to Tune Performance

When the bank initially built its email assistant, the effort focused on technical issues, such as integrating information relevant to customer service from the company website and internal documents. While the assistant functioned as designed from a technical perspective, customer service agents were underwhelmed. They reported that responses to customer queries that the application generated were plausible but often incomplete or subtly misaligned with the bank’s standards, and that the tone of the messages was inconsistent. Agents frequently had to modify messages before they were fit to be sent back to customers. Those results underscored the importance of making better use of the tacit knowledge the agents possessed, which wasn’t captured in available data sets. The customer service agents’ feedback gave developers in the spine a foundation for subsequent iterations. Because those end users were represented at regular project review meetings convened at the spine level, and the spine had the authority to veto underperforming releases (while pushing forward when all criteria were met), the company had the opportunity to get the application right.

The AI spine also used mechanisms to surface more tacit knowledge. Customer service employees were compensated for rating each generated response on a scale of 1-5, and their manual edits to those responses were logged for review. This provided critical data to evaluate the performance of the email assistant, creating a feedback loop that surfaced the organization’s unwritten norms regarding tone, intent, escalation thresholds, and edge cases, which can take a significant amount of time to answer. Since then, the bank’s AI spine has routinized the use of ratings and the logging of edits to AI-generated contents as key metrics for other use cases, complementing technical metrics. That practice allows them to identify cases where efforts dedicated to addressing model or prompt shortcomings are likely to generate value, thus focusing human effort where it will be most valuable. For example, the email assistant’s performance was significantly improved when customer service employees shared their knowledge of how customers typically phrase requests.

Compensating employees for their efforts and directing them to higher-value tasks reduced attrition and alleviated their concerns about job security. Retaining employees with deep domain knowledge is critical for generative AI because they help curate data, refine and update prompts, and provide feedback on outputs. When GenAI is deployed in a business unit with high employee churn, over time there are fewer people who can confidently validate outputs without having to refer to documentation — a slow process that undermines productivity gains.

The spine structure brings discussion of both technical issues and business issues to a single forum and fosters a common language so that everyone can contribute to a multifaceted discussion of application performance. This played out at the medical coding company, where the coding application initially performed poorly and AI developers saw the problem as a technical issue that could potentially be resolved through additional data collection. However, nontechnical participants in the AI spine reviewed the results and challenged the engineers. On closer inspection, they found that ground-truth labels used to train the model were inconsistent and that the system was, in fact, outperforming human coders. The labeling inconsistencies arose because different groups of coders had differing levels of medical education. Discovering those inconsistencies allowed developers to implement a timely fix in the next iteration of the tool and avoid wasting time and resources training the model on an improperly curated data set.

Accelerating AI Adoption and Aligning Stakeholders Across Processes

Both the bank and the medical coding company set strategic goals for generative AI adoption, but business units were reluctant to directly fund GenAI use cases because of the high degree of uncertainty associated with such initiatives, given the unpredictability and unreliability that foundation LLMs are known for. Ideas would seem promising and sometimes work well on a small scale or as proofs of concept but fail to move further. Establishing the AI spine as a separate entity with a clear connection to executives helped maintain strategic alignment.

At the bank, the AI spine funds use cases across multiple phases through microgrants tied to meeting both technical and nontechnical performance criteria; a similar pattern was observed in the medical coding company. This approach helps build solutions and manage risk incrementally for long-term initiatives. In both companies, the spine acts as a catalyst for achievements and lessons that can benefit all business units. While moving funding decisions for generative AI initiatives away from business units was initially perceived as problematic, over time it resulted in bolder innovation.

Members of the AI spine in each company also realized that process knowledge was highly fragmented. Each business unit had a fairly clear idea as to how its part of a given process worked, but a clear vision of the entire process from end to end across units was missing. The spine championed the mapping of key processes where GenAI use cases had been proposed, involving stakeholders from each business unit until these processes could be drawn accurately. With these representations, they could clearly communicate the scope of use cases and their potential expansion, as well as see GenAI’s applicability across multiple use cases, as illustrated by the creation of a new revenue stream for the medical coding company. Unlike AI centers of excellence, which are primarily geared toward the centralization of technological know-how, the AI spine centralizes business process knowledge to fuel innovation.

Maintain Momentum for Continuous Improvement and Diffusion

The AI spine is set up not to deliver technology solutions to internal customers as in the typical hub-and-spoke arrangement but to codevelop them with business unit representatives and continually refine them based on user feedback and other performance metrics. That means business units are less likely to wind up shelving underperforming use cases that have been delivered by a team that’s moved on to something else.

The bank capitalized on its AI spine’s facility for iterative experimentation and treated generative AI as a constant work in progress. While the email assistant eventually led to significant time savings for customer service agents, getting there took multiple incremental rollouts that continually raised the bar. Clear technical and business performance targets defined an apparent goal at each iteration that determined whether the solution would be rolled out to end users. It also gave teams time to learn progressively and to assimilate new technological developments in a field that evolves quickly. Embedding varying degrees of work automation led to a more granular view of performance. One version of the email assistant was good at extracting information from customer messages, but the voice, tone, and style of its written responses were off. Messages that included a lot of bullet points and overly familiar phrasing felt more culturally attuned to the U.S. market than to the more reserved Swiss culture.

Keeping the momentum to continue innovating with GenAI requires a careful balance between the exploitation of existing use cases and the exploration of new ones. The AI spine’s orientation toward users, and the priority that its structure places on better understanding how employees interact with GenAI across processes, has helped the bank to develop new use cases. (See “Scaling GenAI at a Swiss Bank.”) By studying those user interactions through employee surveys and shadowing sessions, the bank was able to identify eight new potential use cases, two of which have since been deployed alongside the email assistant: the AskHR chatbot, which answers HR-related questions on topics such as employee benefits; and an employee handbook chatbot that responds to employee questions about what is permitted and forbidden in the execution of their jobs.

The Drawbacks of Simpler GenAI Structures

Among the 23 organizations that participated in our research, only two were able to meaningfully scale generative AI to a strategic capability, and both of them had created structures recognizable as an AI spine. While that is an admittedly small evidence base, companies achieving such results are currently in the minority, and we believe that their practices are worthy of close attention.

Building an AI spine requires significant top-level commitment, time, effort, and funding. The other organizations that were part of our research adopted one of two alternative structures — what we’ve termed GenAI units and GenAI squads — that we see as potential building blocks to a more robust capability. They were still able to exploit generative AI to achieve some economies of scale, but the scope of their efforts was narrower, and they faced challenges expanding their use of GenAI.

Thirteen of the organizations we studied set up a standalone, central GenAI unit. These are typically technology-focused and charged with providing solutions to the business units. While often the most feasible approach when resources are constrained, the GenAI unit can quickly reveal its limits regarding business (rather than technical) performance and user adoption. A Swiss health insurer illustrates this limitation. Its generative AI unit tried to roll out an email assistant for its customer service center agents. The model could produce fluent replies, but compliance specialists, consulted only after the pilot, pointed out that every customer message had to follow tightly regulated text templates. Fearing noncompliance, customer service center agents reverted back to approved texts, and so adoption plateaued and the project stalled. Without a mechanism to coordinate work across multiple functions (in this case, risk, legal, and customer service), the bank was left with a zombie GenAI use case with no clear path forward. This early failure increased organizational skepticism toward generative AI, making it increasingly difficult to garner support for future use cases.

A stand-alone GenAI unit can be useful for the early, exploratory stages of the technology: It concentrates talent and lets the organization explore GenAI’s potential before committing major resources. However, it cannot properly assess the specific needs of each unit or coordinate work across multiple use cases, and it lacks an end-to-end view of processes that could be improved with GenAI.

Another approach, taken by eight organizations in our study, is to embed small, cross-functional GenAI squads within each business unit. The companies using GenAI squads were able to take ideas to working pilots much more quickly than those using GenAI units, because they were better able to harness relevant business know-how and faced fewer coordination challenges. However, the squads struggled to manage use cases at various levels of maturity. A Swiss insurer’s GenAI squad launched a customer-facing chatbot for product information in record time. Just a few weeks later, curious users prompted the chatbot to recommend pizza recipes. That incident led to multiple iterations aimed at hardening the system — and revealed how rolling out the initial implementation of a generative AI solution is often easier than maintaining and improving it. As the portfolio of use cases expanded, the mounting maintenance burden of updating data sources and prompts stretched the GenAI squad to its limits, leaving little capacity for pursuing new use cases.

Because GenAI squads are funded and staffed by the business units, they can lead to inequities and inconsistent technology adoption at the enterprise level. Some units have the means to fund multiple use cases and hire their own staff. Others must lower their expectations, regardless of the potential value of a use case. This structure perpetuates silos and duplicative efforts; the lack of an end-to-end view of business processes limits value creation at the organizational level. GenAI squads can help to quickly spread generative AI adoption within the organization, but leaders need a different approach if their goal is to coordinate GenAI use cases across the organization.

Creating value at scale with generative AI does not come from giving everyone access to an LLM and hoping that something magical will happen.

Instead, we advise leaders to begin by creating and visibly supporting a small, cross-functional AI spine that can coordinate across business units. The spine should be charged with standardizing and centralizing core building blocks (processes, data, evaluations, prompts, and models).

That mandate must be matched by financial resources. Central funding is essential because it enables the spine to pursue cross-business-unit opportunities that no single business unit alone would sponsor. In addition, executives must define early on what “value” means, to help identify underperforming use cases and normalize dropping them.

Finally, while creating value at scale with GenAI initially depends on a top-down decision by business leaders, sustaining and expanding that value over time depends on continued contributions from business units. Executives should empower and expect the AI spine to convene the technology owner, risk and compliance, business owners, knowledge owners, and end users in continuous collaboration so that their respective expertise can be combined to surface and capture tacit knowledge, align performance with real operating standards, and ensure that improvements compound rather than stall after the pilot phase.

How should companies effectively use or respond to an unwieldy array of customer opinions? While consumer feedback can be invaluable, three recent research articles suggest that it may also be influenced by gender, niche preferences, or sky-high expectations, complicating whether and how companies should respond.

1. Not all users post critical reviews. A study of 1.2 billion online reviews across five major platforms identified a consistent and meaningful gender gap. Women’s ratings are, on average, more favorable than men’s, despite little difference in both groups’ “real” attitudes. The researchers found that women are less likely to share negative reviews when dissatisfied — likely due to societal gender expectations. Women are concerned about possible backlash: being negatively evaluated themselves after posting an unfavorable review.

Notably, however, when users were first prompted to report their opinions anonymously and only then asked to submit an online review, the gender rating gap disappeared. The researchers suggest that businesses could encourage women to share their opinions more openly by introducing a similar process.

2. Adopting user community feedback can backfire with mainstream users. Product developers can gain valuable insight by consulting user communities, but recent research finds that acting on such community feedback can sometimes undermine commercial success. An analysis of video games whose development was influenced by feedback from users who received early access found that the preferences voluntarily shared by self-selected community members can differ greatly from those of more mainstream consumers.

These unrepresentative preferences are particularly salient in lower-priced games or those with niche appeal, suggesting that companies should ignore community feedback for such products until and unless they attract more broadly representative users.

3. Expert ratings can inflate consumer expectations — and deflate reviews. In the prestigious Michelin Guide, experts award star ratings to denote restaurant quality. Favorable expert opinions — more stars — do benefit sales, but a recent study suggests a mixed effect on diner perceptions. Researchers compared reviews by diners of Michelin-starred eateries versus a control group of fine-dining restaurants on the TripAdvisor platform and found a marked “expectation effect.”

Restaurants with multiple stars can find it harder to meet the inflated expectations of customers, who in turn lower their ratings when those expectations aren’t met. While an increase in stars showed no impact on consumer ratings, a decrease in stars improved ratings as consumers became less demanding.

Together, these research findings demonstrate that interpreting and acting on customer feedback can be a tricky and uneven endeavor. The customer, it seems, is not always right.

Of the potential weaknesses of any security system, the human layer has always posed a key risk. The arrival of AI tools has made human cognition even more of a vulnerability.

Companies face three overlapping security threats from AI’s effects on human cognition. First, weaponized persuasion lets attackers manipulate employees’ judgment through personalized, adaptive deception. Second, plausible hallucinations deliver confidently false information that survives casual review. And third, as employees offload cognitive work to AI systems, they engage in less independent reasoning. Together, these dynamics make humans both the primary vulnerability and the key defense.

A simple new protocol dubbed “Think First, Verify Always” (TFVA) addresses these threats by urging employees to take two steps. “Think First” requires employees to form their own judgment before consulting AI. “Verify Always” requires that they cross-check critical AI-generated information against independent sources before acting. The protocol aims to bolster independent judgment and verification and reduce risk, even when manipulation goes unrecognized.

This simple, structured critical-thinking habit can effectively reduce AI risk. In a randomized controlled trial with 151 participants, a three-minute micro-lesson on TFVA improved decision quality by 7.87 percentage points, with a 44% relative improvement in ethical judgment and 25% improvement in information verification. After the micro-lesson, participants were tested on 18 scenario-based tasks (like spotting AI-generated phishing and suspicious executive requests) and scored 65.3%, compared with 57.4% for a control group.

At RSM France, the audit, accounting, and consulting firm where I work, we deployed the protocol in onboarding and training programs for our 1,600 employees. Early feedback suggests that it reduces risk and builds organizational trust.

Managers can embed TFVA in onboarding, security awareness training, and generative AI access policies and reinforce the habit in each instance with a simple three-minute training session. The return is a workforce that treats AI as a powerful tool requiring judgment, not a trusted authority requiring obedience.

Every year, multinational corporations invest billions in global mobility programs. The standard playbook includes training in the customs, values, and communication styles of the host country. However, our meta-analysis of research on migrants, including relocated workers, suggests that cultural knowledge plays a minimal role in expats’ successful adjustment.

In a study recently published in Nature Communications, we and our colleagues conducted the largest-ever meta-analysis of migrant adaptation. We synthesized data from 1,114 primary studies covering 571,260 people, including business expatriates, international students, and economic migrants, to identify the factors that actually correlate with a successful move.

We found that cultural differences are not the primary driver of migrants’ difficulties; stressors like discrimination and navigating new systems have far bigger negative impacts. Social resources like connection and support contribute most to fitting in, navigating daily life, and functioning effectively in the new environment.

The most striking finding for business leaders emerged when we looked at who provides that support: A supervisor’s support was one of the strongest and most consistent predictors of success. It was substantially more pivotal than support from the employee’s own community or compatriots — or even their spouse or family, in many cases.

Why does the local boss play such an important role? First, because feeling connected to others and not lonely are the absolute strongest predictors of adaptation; supervisor support signals legitimacy to the rest of the team and confers belonging. Second, supervisors can clarify role ambiguity (a major stressor for expats), help an employee navigate bureaucratic hurdles, and explain the unwritten rules of the workplace. Third, a supportive supervisor fosters inclusion, mitigating perceived discrimination — the factor most likely to hurt adaptation.

The bottom line: The success of your global talent strategy doesn’t depend on how well expat employees learn the culture. It depends on how well your managers support them.

In this episode of the Me, Myself, and AI podcast, Philips’s chief medical officer Carla Goulart Peron shares how artificial intelligence is reshaping health care — not by replacing clinicians but by expanding access, improving diagnostics, and freeing doctors to focus more time on patients. Drawing on her experience practicing medicine in Brazil’s strained public health system, she explains how technologies like AI-assisted imaging and remote collaboration can bridge critical gaps in care. Carla also explores the challenges of trust, bias, interoperability, and women’s health data in the next era of AI-enabled medicine. She offers a grounded, global perspective on how technology can make health care more human.

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

Transcript

Allison Ryder: How is one clinician thinking about applying AI to health care in additive ways that improve access to care, clinician confidence, and patient experience? Find out on today’s episode.

Carla Goulart Peron: I’m Dr. Carla Goulart Peron from Philips, and you are listening to Me, Myself, and AI.

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

Our guest today is Dr. Carla Goulart Peron, chief medical officer at Philips. Philips is a health care technology company behind imaging systems, patient monitoring, and a growing suite of AI-based clinical tools. What I find fascinating about Carla’s perspective is that she started her career as a physician in Sao Paulo’s public health system, where demand far outstrips resources. She’s carried that lens into the C-suite. She’s now leading medical strategy for a company that’s betting heavily on AI to close gaps in care. Carla, welcome to the show.

Carla Goulart Peron: Thank you very much, Sam.

Sam Ransbotham: Many listeners might still associate Philips more with consumer electronics, but can you tell us about the company in terms of health care and the kinds of things you’re doing?

Carla Goulart Peron: It’s a company that has [been around for] 130 years and has been in many areas, but the last few decades, Philips shifted into health care fully. We started with imaging — so the diagnostic area, X-ray, CT, MRI, ultrasound — and then we [went] into the interventional therapy, into the cath labs. We also are very much present in the ICU and any area of the hospital where you are monitoring signals coming from those patients. [We’re] also heavily invested in monitoring things outside of the hospital. And then the last piece, but not less important, [is] definitely AI that is supporting all those areas of care and then keeps growing [as] a very hot topic right now.

Sam Ransbotham: As you were listing these technologies, I was thinking, “Those are classic applications where AI has made huge advances with imaging and these sorts of things.” I mentioned earlier, you trained as a physician and you worked in both the public and the private sector, sometimes in the same day, I think, switching back and forth. What did that experience teach you about where health care breaks down in a way that maybe technology can help?

Carla Goulart Peron: You are right. In the morning I was working in the public sector in Brazil, where scarcity of health care workers, technology, information sometimes is very much present, so [you] need to work within the best of your capacity to offer care for those patients. But it’s universal health care, which means everybody has access. So there is a benefit there too.

Sometimes in the afternoon or in the night, I was working in the best of the best hospitals in the private sector, with everything available. I think that teaches you a lot of resilience, personally, as an individual, as a physician, but also gives you the chance to try different things and learn from those experiences. But [it] also makes so much clearer to an individual like me how much technology can actually build a bridge and help support more patients [who] are expecting to get access to health care overall … because it’s expediting the way we’re seeing those patients, because it’s connecting the data points of information, or even allowing collaboration across the specialties that may not be present in the public sector.

Sam Ransbotham: Was there a specific moment that you said, “This is something that technology could really help with or could help fix?” Was there anything that made you think technology might be the answer?

Carla Goulart Peron: Many times. I love sharing an example of ultrasounds [from] when I [was] coming out of residency, actually, and starting to see patients on my own. [The] ultrasound is one of the biggest diagnostic tools that we use in the OB-GYN practice. But we [did] not always have access to those machines in the hospital setup. Sometimes we had access to those machines, but we are not qualified to use them.

Technology [is coming] into reality today — I’m very jealous about the people [who] are learning today [in] their own clinical practice — in a way that you have clear collaboration. So you can really open the technology, open the ultrasound machine, get access to an expert [who] can be anywhere [on] the planet, let’s say in the same city, just to make it easier from the clinical practice perspective, to guide you, to see what you are seeing [in] that same imaging, help you to capture the right imaging, and expedite the technology. In some other places, like the ones that were practicing in the public sector, I would need to transfer that patient, sometimes to another facility, which means call an ambulance, be stuck in traffic, just to get the image captured and then [take the] patient back for you to be able to take a final diagnostic and initiate therapy. I mean that was like wine and water, unfortunately, between those two worlds that I was living in.

Sam Ransbotham: You mentioned the traffic. Not too long ago I was in Sao Paulo, and that was a big thing I remember there — just how long it took to get from one place to another.

But actually, ultrasounds [are] expensive machines that can’t move, but then there [are] also the other parts that information can move. I think you were sort of making a distinction between those. You’ve got some aspects of what you’re doing that seem to rely very heavily on sophisticated equipment. On the other hand, you also have information [that] can flow without that bottleneck of Sao Paulo traffic.

Carla Goulart Peron: I mentioned the ultrasound, which is highly dependent on the imaging that you are actually seeing on time, right? So you use that imaging on time to make the diagnostic. But when we think about a CT or MRI, they produce hundreds of images that the radiologist can see from anywhere, and that is also facilitating drastically the way that we are actually reporting imaging. Also, you are highly dependent on the users [who] are actually placing the patient into the machine, making sure that the patient is well positioned, [who] is holding [their] breath if needed, or [who] is kind of moving accordingly. And now with AI coming on board, you also can get that done very quickly, precisely, with not much support or training from the technical perspective. So that has also been a game changer.

I would say the way we are leveraging technology, AI, automation, to position those patients, to make sure that the exam can be as fast as possible, but also how we are processing the imaging that is coming out of those big machines, has been very different.

Sam Ransbotham: I think I was too simple. I was talking about the machine and then the results of the machine, but actually, you’re bringing up an important point, which is there’s also a knowledge transfer and an information transfer about how to get the best image in the first place. I think I glossed over that.

You started [with] bedside [care], but now you’re running medical strategy for Philips. How does that inform what you think about the health care strategy and how that works, your deep background in actually doing medicine?

Carla Goulart Peron: It’s interesting because I think when you go into med school, most of the people [who] decide to take this pathway don’t think about anything differently than just seeing patients on a one-on-one basis. At least when I started med school, that was the reality — I think today is very different.

But in that journey of understanding the health care system, how companies that are developing drugs, medical devices, or other types of equipment work, I [learned] that there is a role to play [on the] industry side, [on] the corporate side that can be as rewarding and as interesting as seeing patients on a one-on-one basis, probably [on] a much bigger scale. So I think I needed to convince myself that by moving from the bedside [to] corporate, I was not changing my background as a physician or I was not leaving my professional [training] behind. I was actually just applying that knowledge in a different way.

I’m fascinated by innovation. I’m fascinated [with] ensuring that whatever innovation we are investing in, in the corporate environment, can actually reach the patients [who] are going to benefit most. And that’s really the biggest part of my job: ensuring that the ideas that our engineers are developing in partnership with hospitals and physicians are going to meet the requirements of the regulation, because we need to prove that it’s safe and it’s effective, but also [ensure] that we will have a good plan in how this technology can be incorporated by the health care system in the way that can actually reach the patients [who] are going to benefit most.

It’s very different, but it’s fascinating, because I keep learning every day because it’s new technologies, new areas of care, new types of health care systems. If you think about a company like Philips that has a global presence, it’s very different to think about commercializing something in Brazil than it is in the U.S. or in Europe or in Africa or in Asia, so you need to have that globalized thought in mind when you are thinking about developing technology.

Sam Ransbotham: Let’s talk about some of these specifics. I think Philips just got [Food and Drug Administration] clearance for SmartHeart, which is an automated cardiac MR [magnetic resonance] planning tool. First, explain to people like me what that actually means, and then how does that actually change a radiologist’s day? What’s different?

Carla Goulart Peron: SmartHeart is a great example. As you think about an MR machine, it is a technology that can capture imaging from your entire body. In order for you to do that, you need to have a technician [who] understands exactly why you are actually being requested as a patient to do that exam.

In this case, [the] physician — a cardiologist most likely — wants to actually see how your heart is functioning. So imagine that a technician needs to know exactly how he or she should be positioning you on the MR table, at which angle, if you are tall or short, if you are someone [who] is big or small, if it’s a kid or if it’s a female or a male — there are so many different data points that a technician needs to understand in order to capture the right level of imaging with the right quality [so] that a radiologist can actually do a diagnostic out of it.

SmartHeart is an AI-driven, one-click automation that plans all setups that drive how the cardiac imaging needs to be captured. This happens in 30 seconds. So that sounds simple, but for the operator [who] is actually doing multiple exams in different parts of the body with different indications, that can be from 15 minutes to 30 seconds. It makes the machine a lot more accessible. It makes the technician capable of doing a lot more exams. It also reduces the dependency of having someone [who’s] highly trained. The burden on the technician is also something that today is a big issue.

Sam Ransbotham: I like the idea that we have a very expensive machine that we can increase the throughput for. You said 15 minutes — that’s four exams an hour, even without any setup and putting on the gown. But you talked about 30 seconds, and suddenly I feel like, “Hey, we have an expensive machine that we can use a lot more.”

Carla Goulart Peron: It is definitely [about the] right speed. There is a burden on health care providers and nurses, physicians overall. But [it] also [involves] “first time right.” So sometimes if you don’t have something like this, you’ll go through the exam, send the images to the radiologist, and they’ll say, “You need to call this patient back because we are missing one or two views.” With something like that, this doesn’t happen.

Sam Ransbotham: Nobody likes that. Nobody likes to go back and forth.

Carla Goulart Peron: Especially on a machine — you don’t want to be there, it’s small, and it may not be that comfortable for the patients overall.

Sam Ransbotham: Radiology is one of the areas that I think a decade ago people were saying, “Oh gosh, we’re never going to have radiologists again. The machines are going to do everything.” You know, that narrative has really not played out at all. But I think it’s a great example of, in general, how artificial intelligence might affect the future of work and what we do. But is there a risk that if this process works too well, hospitals are going to start thinking they need fewer radiologists? How’s that going to play out, do you see?

Carla Goulart Peron: I personally believe that AI is here to add, not to take over. Maybe this conversation will be very different five years from now because I think we’re learning that environment.

Sam Ransbotham: Predicting is so tough in this world.

Carla Goulart Peron: But I think [in] the radiology space for now, radiologists are wasting their time on things that are not valuable at all, reviewing images that were not captured precisely, or doing reports, or reassessing a lot of normal images. Interesting enough, a few weeks ago, I was with one of the radiologists’ medical societies, and they were talking about what if we could have AI defining all normal images, and then radiologists will be looking at only abnormal [images]? What’s fascinating about AI and [its] potential, someone in the audience raised their hand and said, “Well, how are we going to get the radiologists trained in what is abnormal if they are not going to be seeing normal?”

So I think the answer for your question is still TBD [for] what the future will look like, to your point, but I don’t see AI taking over. I think it’s actually helping us to see more patients because there is a big gap out there [and] also to be dedicating our time to things that are unique that we can do as clinicians.

Sam Ransbotham: One way I think about this is we would have a very different story if every possible patient in the world was completely satisfied with a perfect supply of radiology. If everybody that needed this treatment was currently getting it, then I think we maybe could be talking about this replacement type of thing. But you know far better than I do that’s just not the case. There’s a massive undersupply of these types of technologies, and a lot of it is driven by the certain return on investment and the cost structure now, which I believe this can change. How do you see that aspect changing, in terms of serving more?

Carla Goulart Peron: I personally believe that there is such a big gap out there in access to care that as we incorporate technology, we’ll be able to do more with the same, not with less. Because I was born and raised in Brazil and practiced in Brazil, I feel very comfortable in going [along] that pathway [in] low-income countries.

In those areas, the gap is really huge. You can be waiting in line for months in order to have, as an example, an MRI or CT scan done, right? Sometimes too late in the game, that actually can change the patient outcome. Well, we cannot fool ourselves. That reality also exists in the U.S. and in Europe and in Asia. We have deserts, right? We have areas where people don’t have any access to care. So I don’t personally see a timeline where I’ll say, “Well, this will quickly fix that gap in a way that is going to be reduced.”

Sam Ransbotham: I love the idea that what this leads to is better matching of need with supply.

I’ve got a beef to raise with you. You had this Future Health Index that you put together. Normally I like to skim these things before a call, but it was actually quite interesting. So it cost me a lot more time looking at it than I wanted. But I think part of your finding in the Future Health Index was that 79% of the health care professionals are optimistic about AI, but half of the patients are worried that it’s going to reduce their face-to-face time. How are we going to reconcile these two different perspectives from a market [where] both parties are important?

Carla Goulart Peron: I think the perspective of the patients is very important in this one, right? If the patients start to feel too uncomfortable and reject AI, it may become a challenge.

I also think that this is slowly changing. We’re going to be launching a new version of the assessment a little later this year. But for me, the ability to reconcile those two things from the health care professional perspective is ensuring that we are validating, we are getting good access to data that is actually being used to train the machines. So physicians are on board but really a little bit cautious about how much evidential bias do [they] get in the AI. Can I fully trust AI? How much do I need to review what’s happening? I think [from] the physician’s perspective or the health care worker’s perspective, it’s more towards data-driven.

While [from] the patient’s perspective, I think it’s more about the experience. And I think AI became a big buzzword, and so people don’t know exactly what to expect. There is this misperception that physicians are going to be substituted by machines. The reality is, the physician’s time is actually being freed up to be actually dedicated more to the empathy piece, to the touch, to that one-on-one, eye-to-eye, which I think is going to make a big difference.

But to the second point, where I think it’s very interesting, there are some studies already out there that show if you are talking to a real physician or to an AI version of that physician, sometimes the AI can learn how to be more empathetic than the physician. So I think this is going to be a journey of us as individuals actually learning how to incorporate AI into our lives and trusting a little bit that help that I think we’re going to start getting [in the] future.

Sam Ransbotham: In our research, we found that individuals who trust AI are twice as likely to use it regularly. That trust is an important part, but when we were studying it, we were sort of thinking in general [about the] use of AI. It strikes me that’s different about trusting Netflix’s recommendation versus health care, but perhaps there is something transferable between the consumer levels of artificial intelligence and the greater understanding there, and health care. Is there something unique about building trust in this clinical setting that is different than my Netflix example?

Carla Goulart Peron: I think people are more concerned about their health than eventually getting advice of which movie I should be watching or which series. But one of the areas that I think AI is actually already doing for the patients is enabling. I think even the name we gave to that individual [who] is in the center of everything we do as clinicians — patient — it’s almost like, “You stay there, be patient, and wait until somebody tells you what to do.”

While now with, first of all, AI enabling interoperability data points, you are giving more visibility to the overall health. Also, what are the options that those patients may have in front of them? And I think this is, in my view, going to change drastically the way that we, physicians and patients, will be embracing health care for the future, because I think [when] I went to med school many years ago, patients didn’t have a say.

It was really, that’s the protocol, and that’s what we’re going to do, and you just follow it. I think more and more we’re starting to talk about precise medicine, where patients will be able to be offered one, two, three potential treatment pathways with pros and cons and the ability to choose. And I think AI may enable those patients to make more informed decisions at least.

Sam Ransbotham: Actually, I really like the framing of the word patient because I feel like I often am not. We had Josh Weiner from CVS Health on a couple of episodes ago, and one of the things we got into was, “Hey … forget all this AI stuff; I’d just be happy if I didn’t have to put my name in over and over again and wait for a long time.” You can talk about all the cool AI stuff you want to, but let’s get some of those simple things done.

I think you’re very focused on women’s cardiac health. Maybe orient us. What are some of the specific gaps in care for women, and how do you see AI perhaps connecting and either exacerbating or helping with that?

Carla Goulart Peron: I’m very passionate about the topic because I always say we need to remind people that we’re not mini men; we are different as women, and we do have a heart, and it needs to function like any other heart in the man’s body.

There is a big gap. Cardiac cases [have] the highest mortality rate present in females. Despite that, females have a much longer waiting time until [they] get the diagnostic, because we experience symptoms differently sometimes, because most of the protocols that have been designed have been designed based on studies that included only males. So there is a big gap out there that needs to be covered.

What technology definitely can help in covering that gap is ensuring that the nuances and the differences in the physiology in the type of response that females used to present is actually incorporated into the way we are designing the diagnostic tools.

We talked about MR. The position of the heart is slightly different in [a woman’s] chest compared with the men’s chest. It’s a very detailed, minor thing, but it can impact the way you are capturing your imaging. If you are capturing cardiac rhythm, for example, the algorithms need to understand that the female heart has a pattern that’s slightly different from the male heart, and so I think AI will quickly get that information into those algorithms because of the speed, and be able to equalize that.

Sam Ransbotham: That sort of presumes the fact that we’ve got these perhaps underrepresented or undertreated populations within our data sets. We had Ziad Obermeyer of University of California, Berkeley, on the show a couple of seasons ago, and he was talking about how these algorithms can actually build up this equity. But it was depending on having that raw data to start with. What can we do to try to get better sampling on those underrepresented populations?

Carla Goulart Peron: The first thing is what we are doing right now — talk about it, right? So there is an opportunity for us to do better now. You are spot-on. Most of the drugs are developed based on a very limited number of females or even other diverse types of population. You name it.

The same thing for medical devices. And most of the protocols and guidelines that are created are also created based on trials that were developed a long time ago, which didn’t necessarily include the right level of variety. That’s the first point.

The second point is using technology and AI to capture that information pretty quickly and reiterate rather than starting from scratch. I learned a big lesson two years ago when I attended the WEF [World Economic Forum] for the first time, and we were talking about women’s health. I’m an OB-GYN, I have more than 20 years of clinical practice. Someone was telling me the story that when a woman gives birth, there is a standard that defines that 500 milliliters of blood loss is normal. So if you have that, you don’t need to take any actions.

Then I heard a question from the audience during the WEF discussion. “How was that established?” I paused because I had never asked that question. I never had the curiosity. That was in the book. I just kind of assumed that a very good methodology was put in place. And that standard was developed based on nine females in Germany, and exploited and used [on] the entire planet. Can you imagine how that translates to India or China, where there are much smaller bodies? So they quickly were able to iterate that with data points and create the correlation, and established that in India, that number should be 300 milliliters. So that makes a huge difference in how you’re going to be treating your patients.

I think technology, that interconnectivity, that not only the fact that we’re going into automation, but that now AI can analyze such a big data set so quickly, can really improve the way we are practicing medicine.

Sam Ransbotham: Actually, that’s an interesting example because it strikes me as a slightly different approach, which is, in your example, you had to go to the World Economic Forum to have that question get raised. But somehow I can also imagine a very simple job for agents would be, “Hey, go through all of our clinical practices in every area, and find the root study for that, and assess how that plays out.” I would feel like your nine people in Germany sample should rise pretty quickly to the top of that list. That seems pretty exciting.

Carla Goulart Peron: Pretty exciting, exactly. It’s fascinating what can be done.

Sam Ransbotham: Let’s look forward for a minute, though. You started in public health in Brazil. There’s a lot of resource constraints there about technology. If you could pick one AI capability to deploy globally that would make the biggest difference — you get to choose right now, wish list — what would you push out to the world? What do you think [is] the greatest application of the use of artificial intelligence in health care that we could push globally?

Carla Goulart Peron: Interoperability. That is going to change completely the way we practice medicine. Because today we’re very much closed or restricted to the health care system that you are operating. So the ability to see the patients longitudinally without those barriers, I personally believe, is going to change outcomes significantly. If I need to pick one, that will be my choice.

Sam Ransbotham: That’s not at all where I thought you’d go with that. That’s pretty fascinating. It’s so cross-cutting, and it affects everything. All right, so I’ll bite. What’s the biggest barrier to that actually happening?

Carla Goulart Peron: That’s a big question. That’s why it’s a dream. You said you can pick anything, don’t be restricted. I think there are many. The first one is making sure that we have access to good quality data, but also that we start thinking about that from the get-go. If you don’t have some level of standardization, it’s very difficult to think about interoperability. I think that’s the first piece, which is science. It’s how we drive this for [the] future.

The second is how we think about incorporating this new era. How do we think about incorporating reimbursement and access to technology into the discussion about AI? We are still restricted by the reimbursement systems. What kind of code coverage do we have? What is the incentive? We may have something that can actually pretty quickly take the patients out of the hospital, reduce lengths of stay. But if that’s not the incentive from the health care system perspective, this is not going to happen. So for me, that’s another big thing that we need to think about.

The other piece, which I think is how we need to partner, is regulation. Regulation will need to evolve with this new environment that we are getting into. The type of regulation that brought us here is not going to take us to the future because the future is very different than the one we are playing today.

Sam Ransbotham: That seems really hard because, yes, we built this regulatory system and we built these reimbursement processes, and so many processes we built off of the way things used to work.

Thank you for joining us. I think what really comes through so clearly is that this isn’t abstract for you. This is something that you know deeply, and you’ve seen what happens when care isn’t accessible. And I liked some of the ideas that you mentioned about how do we make the use of technology, not just AI, but technology in general, part of the solution and not just sort of a headline. Thanks for joining us.

Carla Goulart Peron: My pleasure, Sam. [I] really appreciate the conversation.

Sam Ransbotham: Thanks for tuning in today. On the last episode of Season 13, I’ll be joined by Bernard Hampton, a corporate learning leader at Bank of America. Speak to you then.

Allison Ryder: Thanks for listening to Me, Myself, and AI. Our show is able to continue, in large part, due to listener support. Your streams and downloads make a big difference. If you have a moment, please consider leaving us an Apple Podcasts review or a rating on Spotify. And share our show with others you think might find it interesting and helpful.

In 2021-2022, CareRx was handling an ambitious expansion. In a span of 20 months, the Canadian pharmacy services company tripled its business through a series of acquisitions. Each acquired company brought its own processes, systems, and cultural norms. 

Employees barely had time to adjust before the next change arrived. “We were growing so fast that the organization could not keep up,” said Adrianne Sullivan-Campeau, chief employee and customer experience officer at CareRx. “We had teams under the same roof not speaking the same language. It was an us-versus-them situation.”

In late 2022, compounded by the pressures of the COVID-19 pandemic, turnover spiked and customer complaints piled up. At one location, Sullivan-Campeau recalled, a leader from the head office arrived to roll out yet another change, and employees turned them away. “‘Leave us alone,’ they said. ‘We’re done.’”

“People were overwhelmed,” she said. “They were asking us as leaders, ‘Why are you doing this? Where are we going?’”

I see this pattern repeatedly in my work advising C-suite leaders through organizational transitions: too much change, too soon, with too little attention to the people living through it. The leaders who navigate this well focus on one thing: managing how their people experience it.

Right now, multiple forces are driving change, including mergers and restructurings, economic volatility, geopolitical instability, and AI technology. Employees and leaders are feeling all of it. Eighty-three percent of business leaders are experiencing more major change than ever before, according to research conducted by my company, The Grossman Group, with The Harris Poll. Employees can realistically absorb only one or two major changes per year, yet leaders are planning on three or four by 2027, our research found. Meanwhile, though nearly all business leaders believe that they communicate change well, we found that 1 in 4 employees say they disagree or are not sure.

Three leadership disciplines make the difference in helping teams analyze and handle change.

1. Make Dialogue Nonnegotiable

Dialogue is one thing many leaders cut back on during periods of change. It takes time and effort, both of which feel scarce in the middle of transformation. Leaders default to what feels efficient. They craft the message, send it out, and move on.

But change is emotional, not just operational. The people closest to the work have perspectives on what’s going well and what isn’t. When they don’t feel consulted, they go quiet, and leaders lose the very information they need most. 

Samantha Stark understood that. She was hired to turn around a several-hundred-person business unit at a global marketing agency that had just lost a major account. “I had tons of ideas,” she recalled. “But I needed to have conversations first and understand the team.”

She met with employees one-on-one, asking about their passions, their dream projects, and barriers they faced. She explained that their input would shape what came next.

Employees told her that the same people kept getting tapped to manage new accounts regardless of their interest, so she built a passion survey that matched people to work they cared about. She heard that different disciplines worked in isolation: The creative team and the public relations team barely spoke to each other, for example. So she redesigned the pitch process to bring them together.

She also created “culture captain” roles and restructured her senior leadership team. She made a point of walking through the office every day and regularly visiting other locations. “When a senior leader puts themselves in a position where people don’t tell them the truth, you’ve lost,” she said. Within months, the team went from negative year-over-year growth to double-digit increases and began winning industry awards.

Making dialogue nonnegotiable starts with listening before taking action and then treating feedback as strategic input. When leaders respond to what they hear and adjust course visibly, people follow.

But leaders also need to close the loop. They won’t be able to use every idea. But employees deserve a response, whether that means moving forward with an idea, modifying it, or saying no and explaining why. Go silent, and they’ll draw their own conclusions, which often leads to quiet resistance that shows up later as disengagement, burnout, or failure.

2. Align on a Change Narrative

Dialogue with employees works only if leaders are all telling the same story. Many aren’t.

Leadership teams I’ve worked with routinely underestimate how much time they need to spend getting aligned before communicating with the rest of the organization. Instead, each leader tells the story their own way, and employees are left to make sense of a fragmented message on their own.

Aaron Radelet, who served as chief communications officer at Hilton, was one of the first hires Chris Nassetta made when he became CEO following the company’s $26 billion acquisition by Blackstone in 2007. They inherited a company in serious trouble: one with more than $20 billion in debt, a slow-moving hierarchical culture, and brands and regions that were operating in silos. There was no shared story or common purpose.

First, the executive team had to agree on what the story was. The vision came from founder Conrad Hilton’s own words: to be the “light and warmth of hospitality,” with a mission to be the world’s preeminent hotel company. That message was woven into talking points, videos, posters, and team meetings. “Nine out of 10 team members knew our vision, mission, and values,” Radelet said. “That doesn’t happen by accident.”

The change was visible externally, too. Nassetta moved the headquarters from Beverly Hills to Washington, D.C., closer to the new owners and one of the country’s richest hospitality talent markets. “The goal was to send a message,” Radelet said. “We were literally and figuratively moving in a new direction.”

Nassetta also launched an executive immersion program in which every leader did stints in front-line roles, such as housekeeping and valeting. Global surveys led to improvements, including a hotel discount program for employees, new education benefits, and a better parental leave policy. Employees felt heard, and the stiff, formal culture that had defined the company gave way to something warmer, Radelet said. Employees who once called the CEO “Mr. Nassetta” were now greeting him with high-fives and hugs. The narrative around warmth helped Hilton improve its culture and go public in a successful IPO in 2013. 

Before communicating change to the organization, leaders need to answer four questions as a team: Where have we been? Where are we today? Where are we going? And what does it take to win? When leaders skip this step of developing a shared change narrative, employees fill in the gaps — and they rarely fill them in favorably, often defaulting to confusion, skepticism, and ultimately disengagement.

3. Sequence Change With People’s Capacity in Mind

Most organizations don’t fail at change because they’re doing too much. They fail because they’re doing too much at the same time, without discipline. Leaders evaluate changes individually and then tell themselves that each one can’t wait. But employees don’t have that luxury. They absorb the changes all at once.

When leaders don’t manage the pace and volume of change, the effect on employees is relentless. One of my colleagues, management psychologist Gail Golden, likened it to standing beneath Niagara Falls. “Inspiring, powerful, exhilarating,” she said. “Until you feel like you’re drowning.”

Golden worked with a leader who experienced this firsthand while running a large national retail company through a sweeping transformation: store redesigns, a new customer strategy, and an overhaul of its shopping experience. New ideas kept coming, and once an idea had run its course, it just disappeared. This leader’s people were left guessing what was still alive and what had been abandoned. Most employees were struggling.

Eventually, someone spoke up: “We love you. But we’re exhausted.”

The leader began holding what they called liberation parties. Rather than letting initiatives die quietly, endings were celebrated, Golden said: “‘We’re putting an end to that initiative,’ they’d say. ‘Have a piece of cake.’” 

The message was simple: Trying something that doesn’t work isn’t failure. That’s part of innovation. Also, the leader made it clear that they were going to kill some changes before rolling out new ones. As a result, people stopped dreading the next idea, and the team became more productive.

At CareRx, leaders faced the same risk: too many changes hitting employees too fast, with no system to manage the load. When Puneet Khanna was promoted to CEO in 2023, having served in senior roles throughout the company’s expansion, he led the executive team in a deliberate reset. They established one core rule: No new projects could collide with major efforts already underway. 

Sullivan-Campeau said the guiding principle from Khanna was straightforward: “We want every employee to get home in time to have dinner with their family.” It became the test for decisions about pace and timing. 

The company launched 15-minute daily huddles, where teams reviewed targets and could raise concerns. It also introduced structural and leadership changes to support the new operating management systems. “We needed the right people in the right roles to support the changes and ensure we paid attention to employee sentiment,” Sullivan-Campeau said.

The combination of those approaches worked. Turnover dropped, the company’s stock price rose, and employee engagement scores improved.

The lesson isn’t that leaders should do less. It’s that protecting your people’s capacity is as important as protecting the business. When initiatives overlap without sequencing, change stops feeling strategic and starts feeling chaotic. 

Here’s an analogy I often share with leaders: When change overwhelms people, it’s like losing your Wi-Fi connection. You’re standing there with all the capability in the world, but nothing connects. The job of leaders is to reset the router — to restore the conditions under which people can think clearly, engage, and do their best work.

That happens when leaders stay in conversation with their people before they act, align their teams on a shared change story, and sequence change with their people’s capacity in mind.

The pace of change will not slow down. But leaders who do those three things will find that employees stop feeling like change is being done to them and start driving it themselves.

MIT Sloan School of Management professors Deborah Ancona and Katherine W. Isaacs have thought deeply about the limits of AI’s capabilities — in their research, and in their classroom work with students and executives. We invited them to share what they’ve learned.

In this video conversation, they reflect on their own experiences and the dangers leaders face when using AI in their work and personal lives. Ancona and Isaacs also share concrete recommendations for delivering value as a leader in the age of AI.

What You’ll Learn From This Conversation

• Why purpose and presence can’t be automated.
• How AI quietly erodes users’ judgment and authenticity.
• What interoception brings to human leadership.
• Why it’s critical to retain your voice in an AI-driven world.

Video Credits

Deborah Ancona is the Seley Distinguished Professor of Management and founder of the MIT Leadership Center at the MIT Sloan School of Management.

Katherine W. Isaacs is a senior lecturer at the MIT Sloan School of Management.

M. Shawn Read is the multimedia editor at MIT Sloan Management Review.

I’ve been trying to move up to the next level and want to be considered a candidate for senior leadership roles. I think I’m well liked, and I understand the company and our business deeply and care about the future of the organization. But recently I received feedback that I’m “not enough of a visionary leader.” What am I missing?

In most cases, leaders who receive this feedback do have a vision. When someone has decades of industry experience and extensive organizational knowledge, the problem is rarely a lack of strategic thinking. Rather, it is typically a lack of visible strategic signaling. You may be thinking like a visionary, but you are not showing it.

Boards and promotion committees are not mind readers. They might not have worked closely enough with you to understand the way you think, and they can’t evaluate your future potential based on what’s in your head. They can only evaluate you based on what you consistently project in meetings, interviews, and conversations.

Many high performers excel at answering interviewers’ questions with precision. They skillfully showcase their operational competence and deep knowledge — often the strengths on which they’ve built their success and that have won them promotions. But being seen as a visionary requires demonstrating a different set of strengths. Your priority in these interactions should be to show that you have a hypothesis about the future of your industry and your company’s place within that future, and that you are actively digging into data and trends to test that hypothesis.

A straightforward question often contains much that is unsaid, and your response can be framed to include strategic insights. A board member who asks about market share in the Midwest is also implicitly asking, “Why does this region matter? How does it fit into our long-term strategy? Who are our competitors, and what can we do to win?”

A visionary response might be something like this: “I think you’re asking because the Midwest has become a growth engine for us, especially in Kansas and Missouri. Our market share is currently 20%, but what is more important is that our top competitors are currently at 28% and 31%. They’re winning through faster distribution and their e-commerce presence. If we want to close that gap, we’ll need to rethink our channel strategy over the next two years.”

In the example above, you’re still answering the question but also showing the thinking behind it, and that there are larger questions you’d like to move the discussion toward. And you don’t need to have ready answers for those larger questions: At senior levels, reflection and synthesis matter more than responding with the speed and precision that got you where you are. Visionary leaders are comfortable sitting in ambiguity and looking for ways to test their hypotheses.

Your hypotheses may continually evolve to match the pace of change in the operating environment. Your task is to make your future-oriented thinking visible to any board or committee. Help them choose you by articulating your evolving hypotheses clearly, connecting the present to the future, and inviting them into your sensemaking process. You become known as a visionary for consistently demonstrating that you are interpreting the future with humility and rigor. Do that well, and people will see you as visionary.

Caterpillar’s CEO had a problem. Jim Umpleby had stepped into the chief executive role in 2017 with a vision of achieving more profitable growth by selling more services and parts to the company’s heavy-equipment customers. Because offering real-time fleet management information services and selling parts online would depend on digital technologies, he set up a new division called Cat Digital. But the division’s head soon had unwelcome news for Umpleby: The company didn’t know its customers well enough to deliver on its goal. Customer data was siloed, fragmented, and, in the case of secondhand equipment, often entirely lacking.

That problem is one shared by countless leaders who see how digital can enable a growth strategy but are stymied by a legacy of fragmented, incomplete, and inconsistent data assets. They are frequently told that getting their data in order is a prerequisite for taking full advantage of new tools like artificial intelligence but too often see data transformation as an IT modernization project. They delegate the work to IT leaders and evaluate success based on cost, speed, and tool adoption. But when data is treated as IT infrastructure rather than as an enterprise asset, its impact is predictably limited. Companies that gain real value from their data actively involve the top management team in data transformation.

At Caterpillar, Umpleby and his team did more than simply sign off on the proposal by Ogi Redzic, chief digital officer of Cat Digital, to build a new enterprise digital data platform. In 2019, they demonstrated long-term commitment by giving Redzic’s team a generous three years to build the platform, dubbed Helios, that would allow the company, its customers, and its dealers to see consistent and complete fleet information across all applications and interactions. Executives went on to redesign data governance, elevating data ownership to senior leaders. By 2025, the Helios platform was supporting e-commerce, fleet management, and predictive and preventive maintenance — and Caterpillar had grown its services revenue from $14 billion in 2016 to $24 billion in 2024.

What Really Drives Data Transformation

Caterpillar’s experience underscores a critical lesson: Data transformation is not a purely technical exercise. Top company leaders must set a goal for the transformation in terms of business outcomes; give executives responsibility for data; commit resources to building an enterprise data platform; give all stakeholders a voice in the transformation; and direct strategic investments that take advantage of new data capabilities. Here, we’ll explore each of these areas and the lessons learned from Caterpillar’s experience.

1. Set (and monitor) an aggressive target for new revenue from the company’s use of data. One particularly consequential action that Umpleby took helped to focus data transformation efforts early on: In 2019, he set the bold goal of reaching $28 billion in services revenues by 2026. That narrowed the company’s attention on a small set of priority data initiatives and on closely tracking how data transformation led to increased services revenues.

Leaders initially prioritized developing a new tool, VisionLink, to help customers manage their fleets, and enabling customers to order parts via new e-commerce capabilities. Setting these priorities helped the team that was building the Helios platform focus on cleaning up the customer, customer contact, and equipment data needed for those solutions. Those three data types were known as “the trifecta” because they were key to answering important questions like “Which contact working for which customer should receive a sales or service offer?” and “Which contact working for which customer is responsible for replacing which machines?”

Top leaders regularly monitored Cat Digital’s progress on those priority initiatives using three measures of value. Value enablement was measured by the number of accurate trifecta records on the platform. Value created for customers and dealers was assessed based on the number of VisionLink users, how often the application was used, and how quickly that usage was growing. Value realization was tracked as revenue attributable to the new solutions, such as parts sales. This value-based reporting approach helped business leaders connect progress made in data initiatives to impact on the corporate income statement.

By establishing clear business goals, leaders keep decision-making aligned with strategic objectives. Because foundational work can take years, tracking and reporting progress that has been made is important to sustain stakeholders’ patience.

2. Give senior business executives ownership of data. Data becomes fragmented and inconsistent in many organizations because no one is accountable for its life cycle, quality, or reusability. At Caterpillar, establishing data as a corporate asset for which senior leaders were accountable focused attention on data quality. It also tied executives closely to the employees responsible for maintaining data products and related solutions.

To implement the new data-ownership policy, Cat Digital and the corporate IT team identified 14 enterprise data domains (such as engineering, finance, and human resources) and recruited about a dozen vice presidents across the company to own them. Some leaders represented functions such as finance or procurement, while others represented business units that had a strong stake in a particular data domain. Redzic took on customer, contact, and equipment data because those domains were key to services revenue. The vice presidents received information monthly about the quality of their data and were kept informed about the applications that used their data.

Cat Digital also established the role of data product owner — individuals who tracked value enablement and value creation metrics and were expected to drive data reuse and make data investments pay off. (See “Before and After Caterpillar’s Data Transformation.”) On the tech side, data product owners made sure that widely needed data, such as customer master lists and fleet information, were designed as reusable components. Owners of customer data came from the business side, bringing valuable domain expertise to the work of defining quality standards.

Owners of digital solutions were expected to generate customer value and service revenue. For VisionLink, that meant driving customer retention, customer satisfaction, and aftermarket parts revenue. Then owners of digital solutions collaborated with the owners of relevant data on creating data products, such as the fleet list, which tracked all the equipment each customer owned, operated, or rented. By maintaining a fleet list data product, Caterpillar ensured that its customers saw the same fleet list data everywhere.

Having vice presidents accountable for data assets, and establishing roles dedicated to managing data products and data solutions, made it clear that data was no longer an IT issue but an enterprise priority. Over time, executives understood why they had to be involved and took on the accountability and responsibility that came with data domain ownership. At the same time, teams of all kinds started paying more attention to data decisions, knowing that their actions were subject to vice presidents’ oversight. Executives did not step on each other’s toes: No executive would invest in setting up a siloed customer database when one of their peers had clear responsibility for customer data.

3. Commit resources to building an enterprise data platform. Caterpillar began with a fragmented data landscape resembling that of many big, old companies. Independent dealers operated disparate systems. Multiple divisions maintained their own applications and databases. About 200 different interfaces connected dealer and corporate systems. Customers relied on multiple Caterpillar applications and often saw inconsistent information about their fleet. Cleaning up and consolidating such a landscape is not for the faint of heart. Many organizations bend to time pressures and cut corners on architecture and then fall back, yet again, on developing application-specific solutions optimized for immediate needs, exacerbating the architectural sprawl.

Caterpillar’s leaders approached the cleanup with a realistic understanding of the time and money it would take to build a platform that could serve up the company’s most important information consistently, quickly, and cost effectively. Cat Digital’s work building the Helios platform ultimately resulted in the shutdown of eight legacy data platforms, reducing data management complexity by a factor of 30, by its calculation.

Over the three years spent building Helios, Redzic set milestones and regularly presented leadership with evidence that the group was meeting them. That sustained leadership’s confidence over a relatively long time horizon as legacy systems were sunset and business areas moved onto the platform.

By investing upfront in an architecture for reusable data components and providing time for teams to stabilize foundational components, Caterpillar established the capability to deploy data products and solutions quickly and efficiently. (See “Caterpillar Data Architecture.”) By 2023, Helios was serving as a single source of data for digital services to customers. When a customer managed its fleet using VisionLink and used Caterpillar’s e-commerce system to order parts, current data about the fleet was reflected in both systems. When a customer received preventive maintenance recommendations, those recommendations were associated with the bill of materials data that informed the e-commerce system. In 2024, the company saw record use of VisionLink, with thousands of customers newly onboarded.

4. Give internal and external stakeholders a voice in the data transformation process. Building solutions centrally and then pushing them to partners and business units often leads to resistance, fragmented adoption, and solutions that don’t fit how work actually gets done. Caterpillar’s leaders avoided this outcome by establishing channels by which independent dealers and a variety of business units could engage with the company’s data transformation efforts. The approach ultimately helped speed the adoption of its digital solutions.

Dealer engagement was particularly important. Redzic established the Dealer Digital Council, which brought together interested dealers representing Europe, Africa, and the Middle East; the Americas; and the Asia-Pacific region. The council held two-day meetings each quarter to review digital product road maps. The discussions provided Caterpillar’s leaders with direct feedback from the field and helped ensure that new digital services aligned with dealers’ needs and operational realities.

A second group, the Enterprise Dealer Integration Council, focused specifically on data integration across the dealer ecosystem. Led by Caterpillar’s vice president of dealer and customer support, the council was responsible for establishing standards for the data exchanged between Caterpillar and its dealers. All Caterpillar groups that sent data to or received data from dealers were represented on the council.

Internal business units were involved in shaping digital priorities via a new demand review board, where digital leaders met monthly with business sponsors proposing new initiatives. The board ranked project ideas based on their alignment with the Helios road map and their expected business benefits, cost, and required expertise.

The Demand Review Board created transparency about which initiatives would move forward and why. It also helped Caterpillar avoid a common pitfall of large enterprises: the reemergence of siloed applications and redundant data pipelines. By reviewing initiatives centrally, the board nudged teams to harness Helios’s reusable capabilities rather than build new, isolated solutions. The result was fewer applications overall and greater consistency in data, user experience, and platform utilization — ultimately reducing enterprise spending.

The value of the board was evident in the development of the Cat Vantage Rewards program (recently relaunched as Cat Rewards). Leaders of Cat Financial proposed expanding the company’s rewards program to allow customers to earn points from online purchases made with a Cat credit card. Implementing the idea required that data from Caterpillar’s e-commerce systems, dealer invoicing systems, and customer accounts be connected — a task that would have been extremely complex in the company’s previous, fragmented data environment.

Instead of building a solution from scratch, Cat Financial took advantage of the Helios capabilities already in place. Using Helios customer master data and existing data objects related to invoices, the teams quickly developed services that identified eligible transactions and calculated reward points. Within months, the solution had progressed to testing and integration with Caterpillar’s broader digital rewards experience. According to leaders within Cat Financial, implementing the program would likely have taken years under the company’s earlier architecture.

Caterpillar’s senior leaders learned the importance of ecosystem involvement and impact. By involving dealers, business units, and product teams in shaping the company’s data and digital initiatives, they drove viable solutions and desirable outcomes. Data transformations succeed not only because of strong architecture but also because of strong relationships. Organizations that give their ecosystems voice dramatically increase the likelihood that new capabilities will be adopted and add value across the enterprise.

5. Build on new data capabilities with strategic investments such as AI. While working on the data transformation, Caterpillar began building its AI expertise by using machine learning tools to improve data quality. Its data science team worked closely with business domain experts on models that could identify and correct anomalies in incoming data, such as incorrectly entered equipment serial numbers or inconsistencies in customer records. Besides automating previously manual work, the models allowed the company to maintain high quality standards as the volume and complexity of data increased. Just as important, the development of the models educated business collaborators and senior leaders on possible use cases for AI.

Cat Digital also engineered new AI capabilities directly into Helios as generative AI emerged. The team created vector data stores to house large volumes of unstructured information, such as equipment manuals and service records, so that the AI systems could quickly search them. That allowed employees and customers to ask questions and receive precise answers based on information included in company documents. Cat Digital created prompt libraries for generating service recommendations and summarizing equipment performance. It also built agent orchestration services to coordinate multiple AI agents and digital services so that they could work together to complete tasks — for example, analyzing machine telemetry, identifying a maintenance issue, generating a service recommendation, and notifying a dealer or customer of a problem. To ensure that AI development remained closely tied to business objectives, the AI Digital Product Council evaluated potential use cases and prioritized those most likely to create value for customers, dealers, and the enterprise.

Key Points for CEOs Leading a Data Transformation

Caterpillar’s experience makes a compelling case for why transforming data to unlock the value of those assets requires attention and leadership from the CEO suite. Executives who know that their organization’s data requires similar transformation — most urgently, so they can take advantage of AI — should keep the following in mind as they hone their strategy and leadership approach.

• Set a stretch business goal that the data transformation will serve, and use that goal to focus people’s efforts and keep the overall project on track.
• Tie data investments to revenue growth, margin expansion, risk reduction, or customer value creation rather than simply tracking infrastructure spending. Monitor a set of metrics that clarifies the causal chain from platform investment to business aspiration.
• Require that business leaders assume data ownership; don’t assign this to technology leaders. Establish product management roles — for data assets and data solutions — to manage the deployment of reusable data and track data payoffs.
• Invest in a data platform, not another data silo. Monitor progress using enterprise KPIs like data reuse and data liquidity (that is, how available data is for immediate use). Give people the time they need to establish foundations rather than add to sprawl.
• Involve partners and business units in informing and shaping data transformation. Establish mechanisms to identify and proactively manage potential conflicts.
• Use AI to automate data management tasks while establishing the company’s data foundations.

Every large organization today manages vast data assets, but few extract their full value. The difference lies in executive leadership. When CEOs talk about data in earnings calls, participate in governance discussions, and hold leaders accountable for data quality, the organization listens. In an era where competitive advantage increasingly depends on insight, integration, and intelligent automation, companies that treat data as chiefly the responsibility of the IT function will fall behind. Those that treat it as a corporate asset — and lead accordingly — will define the next decade of performance.

B2B sales is fiercely competitive. Companies selling big-ticket products and services to other businesses must design solutions that meet their customers’ specific needs with a provable value proposition. Increasingly, that means engaging in value-based sales, where the benefits to the customer are defined, quantified, and managed by the vendor. That’s a challenging practice to get right: Many industrial companies fail to move on from piloting solutions to delivering them at scale.

Our in-depth research on industrial equipment manufacturers (IEMs) that have embraced this business model has revealed what separates the successful scalers from those who remain in pilot purgatory. Those lessons are useful to any B2B companies embarking on, or struggling with, a value-based sales approach.

Many IEMs are facing a perfect storm as two major trends converge. First, they are encountering intense competition from low-cost players, as their advanced physical products alone no longer confer a competitive advantage. Second, many of their traditional product offerings, which already carry a high cost of ownership, emit high levels of greenhouse gases at a time when customers are striving to comply with stricter environmental standards. With the convergence of these two trends shrinking opportunities for growth and profitability, IEMs are increasingly looking to new business models that might reduce customers’ operating costs and emissions. However, the ability to execute such strategies at scale has been elusive.

To improve their customer value propositions and their profit margins, IEMs are creating value-based industrial solutions.1 We define them as customized and integrated combinations of products, service, and digital technology that allow companies to achieve profitability and sustainability simultaneously by providing value in use. In some sectors, such as construction and mining equipment, solutions like fleet optimization and equipment-as-a-service already account for the majority of total revenues.2

Offering value-based industrial solutions requires manufacturers to shift from selling products to fulfilling complex customer needs — that is, to a value-in-use business logic, where the value created is shared between manufacturer and customer.3 Such solutions have three distinct but interrelated characteristics. First, they are customized to the unique operational needs of a specific customer rather than implemented off the shelf.4 Second, they comprise integrated combinations of products, services, and digital technology that enable greater value collectively than each does in isolation.5 Third, these solutions aim to deliver value in two key ways: by improving both profitability and sustainability outcomes. Profitability may manifest as productivity improvements and reduced operating costs, while sustainability can be achieved by lowering the CO2 footprint of operations and enhancing product utilization.6 Scania’s fleet management system, Volvo Construction Equipment’s site optimization solutions, ABB’s motor-as-a-service offering, Caterpillar’s autonomous solutions, GE Aerospace’s TrueChoice Flight Hour maintenance program, and John Deere’s TimberCare service offering are all examples of such solutions.

The Scaling Challenge

Our research shows that it is relatively easy for an IEM to create an initial value-based industrial solution. Doing so on a one-off basis for a single, key customer limits the scope of financial commitments. The value created for the manufacturer in this case is not primarily about the revenue derived from the customer; rather, it’s about proving technical and operational feasibility, as well as projecting a brand image of proactive, innovative thinking. However, scaling these solutions is much more challenging than delivering pilot initiatives: Offering these solutions across a large and diverse customer base requires a repeatable, structured process and strong, entrenched capabilities.

IEMs that seek to scale these solutions expect higher profits, but that often fails to materialize. Instead, many companies find that the investments needed to deliver value-based industrial solutions more broadly increase both their costs and organizational complexity, without improving profitability.7 In other words, IEMs often struggle with managing scaling in practice, and their customers frequently complain that value-based industrial solutions are not financially viable.8

We studied 19 companies across eight industry sectors to gain insight into where the scaling process breaks down and found that many IEMs overestimate the appeal of their value proposition and their customers’ willingness to pay. Across the cases, the IEMs also found it challenging to properly configure solutions and align with partners. Many were also struggling to implement new revenue models and build sufficient capabilities to deliver promised value in practice, especially when doing so required bringing in third-party services.

There are numerous reasons why these challenges prevail. First, like many manufacturers, IEMs have a deeply rooted culture that prizes developing and improving physical products. This limits their ability to deliver new forms of customer value beyond the products themselves, such as value-based solutions.9 Second, they are often organized for transaction-based customer interaction and lack the organizational readiness for solution delivery and implementation.10 Third, creating a profitable revenue model may be particularly complex because it requires alignment and agreement with third-party service providers and other ecosystem partners, over whom a focal IEM has limited control — and it requires contributions from customers, which magnifies complexity further.11

What’s Required for Value-Based Industrial Solutions to Scale

Our analysis shows that organizations that have been able to scale value-based industrial solutions have established a set of prerequisite capabilities in the initial phase of designing and piloting such offerings. In the second phase, where scaling gains traction, they have gone on to develop additional capabilities required to repeatedly sell and deliver those solutions.

In the framework we developed from our research, we have defined three core scaling capabilities for each of the two phases and briefly described the 17 practices that are key to executing each phase successfully. (See “Two Phases of Capability Development for Scaling Value-Based Industrial Solutions.”) Our research also indicates that the chain is only as strong as its weakest link, and deficiencies in any single scaling practice can stall the overall scaling process. As we look more closely at the following six core capabilities, we’ll identify particular areas of challenge that arose for the companies we studied, as well as examples of what was executed well.

Phase 1: Scaling prerequisites. The groundwork for developing value-based industrial solutions comprises the following three steps.

Develop a solution strategy. The strategy work that kicks off the process of developing a value-based industrial solution is not a particular roadblock, our analysis found. Managers recognize that adopting a value-based model represents a profound shift. As one told us, “For most of our product lines, we no longer sell just products — we deliver comprehensive solution projects tailored to our customers’ operational needs.” That said, it’s important to start by articulating a clear vision with a rationale for the shift, and to establish a set of clearly defined goals. Otherwise, top management will struggle to win the commitment of the rest of the company. Second, leaders must use a structured planning process to design a road map that outlines key activities, milestones, and timelines that operationalize the vision and goals into practical, actionable steps. Finally, leaders need to make a credible resource commitment: They must dedicate specific people, a defined budget, technology, and other assets to execute the road map.

Create a dual value proposition. Where companies have typically struggled in this initial phase is in defining a value proposition that creates both business value (profit, growth, competitive advantage) and environmental value (resource efficiency, carbon reduction, waste elimination). Successful companies use environmental value to create business value. This is a must-do: Not a single company we studied was willing or able to pursue sustainability without a clear business case.

An ecosystem partner working with an IEM we studied told us, “Profitable sustainability is front and center. Adopting advanced electric machines is not just about reducing emissions — it’s about achieving tangible financial gains. These solutions must lead to measurable reductions in operating costs, ensuring that the total cost of ownership is justified from both a financial and environmental perspective.”

However, customers may hesitate to adopt such solutions due to perceived high costs. A supplier of construction solutions invested heavily in developing electric construction machines, thereby eliminating local emissions while simultaneously reducing noise. Despite the environmental superiority of the solution, customer uptake was much slower than expected due to the more costly investment. Further, in some markets, the impact of low fossil fuel prices, or policy incentives favoring solutions that use fossil fuels, meant that savings were insufficient to offset the higher cost.

Creating a dual value proposition requires both defining and validating the promised value. Some companies found it fairly straightforward to clearly articulate how their solution would deliver both business value and environmental value. A manufacturer of heavy trucks did this well. “For long-haul operations, digital services enable our customers to reduce CO₂ emissions by cutting fuel consumption through real-time route and driving optimization,” a leader there told us. However, few IEMs correctly defined the business value of their solutions and hence failed at getting perceived customer value right. All too often, the marketing, sales, and R&D functions conceived of business value differently — even years after a solution had been launched. IEMs also commonly held inflated perceptions of the business value of their solutions that customers did not find credible.

That is why value validation — confirming and measuring the dual value created via customer feedback and real data — is so critical and so challenging. Digital maturity helps with this, as exemplified by an ecosystem actor catering to the mining industry. The company’s digital mining intelligence service provides data on asset and operational performance, which enables it to produce a customized dashboard with visibility into actual operational efficiency, environmental impact, and safety outcomes.

However, many other actors were much less sophisticated, and it was not uncommon for them to attempt to validate value via word of mouth or ad hoc customer interactions. One manufacturer developed what it believed to be a groundbreaking solution for industry automation: a suite of semiautomated, electrified machines designed to improve both efficiency and sustainability. From the IEM’s perspective, the business value seemed obvious. A senior product manager told us, “With these systems, operators can cut emissions while reducing downtime. We were confident the total cost of ownership would be compelling.” However, the customer saw it quite differently and was not convinced that the economics of the solutions were competitive and realistic — nor that the IEM’s solution reflected its operational realities.

Design for modularity and customization. Another important scaling prerequisite for delivering solutions that are customized to a particular operating context is the capability to modularize products, digital assets, and services. For products, this means developing key components, such as motors, pumps, and transmissions, such that they can be used across a variety of customer solutions. Similarly, digital modularity means designing software, data analytics, and communication interfaces to work across diverse solutions. Developing service modules that work across solutions — such as proactive maintenance contracts, upgrading, and refurbishment — enables efficient and flexible delivery and is key to delivering on the promise of a value-based industrial solution.

One construction equipment manufacturer developed standardized products, services, and digital modules along with internal processes to configure them efficiently. This gave it flexibility in designing solutions and was cost effective for its customers. In contrast, another IEM that focused excessively on customizability and limited standardization ultimately found the modular approach to be operationally inefficient and not financial viable.

Our research found that the most challenging among the practices involving modularity and customization was solution configuration. Seamlessly bringing together product, service, and digital modules was a headache across the cases we studied. Each of the three types of modules has a different value logic. Physical products are judged on reliability and quality, services are assessed more on responsiveness and customization, and digital assets are evaluated on connectivity and usability. Because the value logics are orthogonal, solution configuration becomes a problem of aligning performance regimes that are partially incompatible. For example, service customization can sometimes undercut product reliability.

Another issue is that each module type typically emerges from different units or departments within a manufacturer, which, in our cases, caused siloed thinking and hence additional integration problems. Third, the effectiveness of digital modules depended on customer data, which some customers were reluctant to share with their vendor for fear of lock-in and loss of control — but without customer data, the IEM couldn’t unlock the full value of the product and service modules. Yet another problem centered on customers having different procurement routines for products, services, and digital assets, which complicated their acquisition of value-based industrial solutions.

The example of a manufacturer catering to the mining and steel industries shows how to configure solutions well. This company followed an agile and collaborative process where customers were involved in codesigning advanced solutions and supported the configuration process. It brought together delivery experts, product owners, and R&D early to define both module integration and performance outcomes. Structured guidelines enabled sales and delivery teams to rapidly assemble and adapt solutions using standard modules with minor adjustments — thereby making the solution both efficient and responsive to individual customer needs.

In this process, a particularly pressing point was procurement and financial cycles related to different modules of the solution. Industrial products typically had an estimated life span of at least 30 years and were acquired through significant capital expenditure investments. In contrast, services were budgeted on shorter cycles of three to five years, while digital features required recurring operational expenditure payments. By involving customers early in the codesign process, the manufacturer was able to align these cycles and negotiate a hybrid model: Equipment was sold on a discounted CapEx basis, while services and digital modules were bundled into a three-year OpEx contract with the option for successive three-year extensions if predefined performance outcomes were met. This alignment reduced friction in procurement and built confidence that both operational and financial goals could be achieved simultaneously.

Phase 2: Scaling execution. As detailed below, scaling execution centers on the capabilities needed to deliver value-based industrial solutions repeatedly and across a variety of customer engagements while ensuring that financial objectives are met.

Manage ecosystem contributions. Developing and orchestrating a partner ecosystem is essential; even large IEMs cannot scale value-based industrial solutions on their own. We didn’t meet a single one that was horizontally and vertically integrated to the extent that all of the capabilities in its product, service, and digital modules were in-house. Some added needed competencies via partnerships, acquisitions, or both. One filled competency gaps in its design and delivery structures via a combination of acquisitions and contracts with complementary partners.

Building out an ecosystem requires that complementary partners be mapped out and onboarded, which worked relatively well across the cases we studied. However, aligning with partners on who would do what and who would be situated where proved much more challenging. Partners’ interests and goals tended to diverge from IEMs’, particularly when a value-based industrial solution was novel and the relationship between partner and manufacturer was recent; those factors left partners hesitant to commit to uncertain future pathways. Such situations required active orchestration and engagement. Yet another complication was that alignment normally required the sharing of data and intellectual property, which raised concerns about opportunistic behavior and knowledge leakage.

A diversified global manufacturer of industrial automation equipment had an innovative approach to partner mapping, onboarding, and alignment. The company sponsored an accelerator program for over 100 startups and small and medium-sized enterprises, with a focus on codeveloping solutions that integrated directly with its existing product lines. This collaborative approach allowed the manufacturer to tap into the capabilities of a diverse partner network and subsequently onboard and align with a subset of these companies to scale innovative solutions. Another IEM partnered with a global digital platform provider and innovative digital startups to put together complete value-based industrial solutions. After the onboarding phase, the partners worked tirelessly on clarifying responsibilities and mutual expectations, and on data-sharing protocols, to prevent overlap and friction. As trust and experience accumulated, the partners successfully developed a scalable solution that spanned the full solution life cycle — from design and operation to performance monitoring and end-of-life recovery.

Ensure financial viability. Needless to say, it is essential that a manufacturer cover its costs, recoup its investments in value-based industrial solutions, and generate profits — while maintaining an attractive customer value proposition. To manage this, an IEM must have a detailed understanding of cost structure, which is fairly straightforward, as well as the right revenue model and risk management in place. These latter two practices were more challenging for the companies we studied.

Value-based industrial solutions normally involve adopting a new type of revenue model. Upfront revenue streams from product sales are largely replaced by monthly licensing fees, given that the IEM normally retains ownership of the equipment used in its solutions. This underscores the need for rigorous risk management to be front and center. The companies in our study experienced instances of adverse customer behavior, such as in situations where equipment was misused, resulting in additional repair and maintenance costs for the manufacturer. Some IEMs also took on unanticipated costs stemming from complexity in service delivery, and others faced new financial risks, such as working capital being stressed when revenues from conventional product sales diminished. All of these risks will effectively undermine the revenue model if they aren’t proactively addressed.

One manufacturer of process equipment for mining companies pursued outcome-based cost-per-ton contracts successfully. An executive of the company told us, “This is a high-risk revenue model in which we earn in proportion to customers’ operational success materializing. While this model tightly aligns incentives — monetizing value through shared productivity gains — it also places significant pressure on us to continuously deliver measurable outcomes.”

Many of the IEMs we studied spoke of challenges in meeting performance guarantees due largely to limited insights into customer operations. This was particularly prevalent in early attempts at scaling, before learning effects had accumulated. One manufacturer eventually abandoned its attempts at performance-based contracts after years of pilot tests due to risk management and cost structure concerns. Capability gaps in delivering contracts onsite could not be closed, and the risk of consistently failing to meet promised performance outcomes was deemed too high.

Expand the addressable market. Implicit in scaling is the intention and need to reach a larger market. And, given the specificity of value-based industrial solutions, this implies building the capability to expand internationally, where customer requirements can differ significantly across countries, regions, or segments. IEMs must identify additional customer segments that are good prospects for their solutions and/or penetrate existing segments more deeply. Here, understanding customer readiness, regulatory pressure, policy incentives, and infrastructure support is key. For example, in scaling its site electrification solutions, a manufacturer of construction equipment learned that not all customers or markets were equally prepared for it. In some cases, customer maturity was too low; in others, limited local sales and delivery capacity effectively hindered implementation at scale. To address this, the IEM conducted in-depth market analyses to identify high-potential segments and regions where both customer readiness and organizational reach were aligned — thereby ensuring more focused and effective market penetration.

Addressing a larger market requires IEMs to define and clarify the roles, responsibilities, and structure of the internal or partner-based delivery organization responsible for getting the solutions implemented in practice. Some manufacturers took a hybrid approach — that is, they used an internal setup in some markets and a partner-based one in others. Some internal delivery organizations experienced major challenges, but with a partner-based approach, complexity grew further and capability deficiencies were a major issue. The delivery organizations needed to make their solutions work at each customer’s site, but even with standardized modules, the customer context and specific requirements were seldom identical. This implies that a standardized process for solution delivery was often unfeasible, and thus solution performance, such as equipment uptime, varied. We also observed a few cases where partners disagreed on who was best positioned to exert influence over solution architecture in practice: an IEM from its domestic headquarters, or third-party service actors in a specific country.

A manufacturer of equipment for the forest industry created a delivery organization equipped to implement complex solutions and deliver operational performance over time. The company adopted a forward-leaning and proactive delivery model, where service delivery teams not only responded to customer issues but also actively ensured that its solutions delivered guaranteed results. Building this capability involved combining technical expertise, field support, and monitoring routines into an organizational structure and process to support scaling and long-term customer success. An executive from a construction equipment manufacturer further underscored how instrumental the delivery organization was for customer satisfaction: “Solutions require a completely different mindset. Our delivery teams now need to collaborate closely with customers over time, monitor performance, and ensure outcomes are achieved. That demands new skills, new incentives. … Unless we invest in these capabilities, the delivery organization will drift back to its transactional delivery model.”

Using the Framework in Practice

By considering the successful scaling of value-based industrial solutions as a two-phase operation comprising six core scaling capabilities broken into a total of 17 scaling practices, manufacturers and their partners can more easily and systematically identify their relevant organizational strengths and weaknesses. This sets the stage for deliberately building scaling capabilities to bridge the company-specific weaknesses we identified. These weaknesses can span strategic, technological, organizational, financial, and market issues, making scaling decisions inherently complex.

Based on our analysis, we offer the following guidance to managers pursuing a strategy of developing and marketing value-based solutions broadly.

• These solutions are complex, highly customized, and ecosystem-dependent. For this reason, take a systematic and holistic approach to scaling them. Rollout via ad hoc actions or uncoordinated initiatives will not work.
• Scaling success is contingent on reaching sufficient maturity in all scaling practices. That said, getting the dual value proposition, solution configuration, partner alignment, revenue model, and delivery organization right seems particularly challenging. Managers should pay special attention to these scaling practices.
• Misalignment or poor timing — such as engaging global partners before a clear internal strategy has been established or attempting wider replication before solution-market fit has been achieved — is likely to create bottlenecks, resource misallocations, and scaling fatigue. Therefore, make sure that scaling practices are conducted in the suggested order.

"If AI is going to destroy all the jobs, why don't we just stop?"

That was the rhetorical question my college-age son asked after we talked about the possibility of drastic changes to career paths and society thanks to AI (technically, generative AI). It was in line with what I've been worrying about myself.

Nobody really knows how disruptive AI will be. But young people and their parents would be foolish not to prepare for deep, unprecedented change in how we work. A huge portion of entry-level white-collar jobs — the kind that college graduates normally flock to and count on as career springboards — may not exist in the near future.

I'm not alone in these estimations, obviously. Dario Amodei, the CEO of Anthropic, has been brutally honest about what he believes his products will do to hiring. He has (repeatedly) said that half of entry-level jobs — especially in fields like finance, consulting, law, and tech — are likely to disappear within a few years. Interestingly, he’s changed his tune very recently, suggesting that there’s an opportunity for job growth. But either way, the facts on the ground bear out the concerns. Reductions have begun: Goldman Sachs estimates that 16,000 jobs are evaporating every month.

So, what's to be done? In a widely circulated clip from a May 2024 interview, former Google CEO Eric Schmidt put it plainly: Once AI agents develop a suite of skills that allow them to start working together on their own, away from our guidance, "we won't understand what the models are doing." His suggested solution? "Pull the plug."

It's a gut response that I feel a strong affinity toward, even as I dive deep into using AI myself. As I watch the world barrel toward a truly unknown future and the potential devastation that AI could wreak on job markets and young workers, I feel a mounting unease about how companies are starting to respond.

What makes this challenge particularly hard to solve is that the executives making decisions about AI deployment and jobs will be fine regardless of how this plays out. They have capital, seniority, and options — financial and otherwise. It's sadly uncommon for leaders to think beyond market cap and their own vesting schedules and consider whether we all can thrive. That inequality in exposure to risk is part of what makes this more than just a business question.

When society faces deep risk, companies and leaders tend to make choices that seem optimal for their short-term interests. From a pure short-term-profit perspective, bringing in fewer workers is probably the financially smart thing to do. But thinking about only the short term poses significant danger. With this latest existential challenge, if companies continue to head down a "people-light, AI-token-heavy" path, the risks aren't just to young workers but to businesses, too.

The microeconomic case for some degree of caution is this: If companies decimate entry-level roles, what happens to the pipeline for leadership? Service businesses have long had a pyramid model where lots of young, smart kids come in and get trained and tested, and then a small subset make it to partner or other senior roles.

So, what if companies just didn't eliminate as many jobs? Yes, we're about 40 years into this model of businesses announcing cuts and their stock rising — investors often love companies that fire people. But what if, this time, they just didn't? The companies that preserve human judgment, build institutional knowledge, and keep developing talent may find themselves with the advantage down the road.

Parallels to Inaction on Sustainability

Watching the march toward the job slashing unfold, I feel a sense of déjà vu. It's a collective action (or inaction) problem, much like climate change. We have watched the scale of potential environmental devastation rise fast, in real time, and have still struggled to respond with urgency. In both cases, with society facing deep risk, many companies have made choices that seem optimal for their short-term interests. The results could be catastrophic for everyone.

I'm haunted by a conversation I had a decade ago with the COO of a major corporation. I did my normal spiel about all the ways that sustainability can create value over time. His response: "Yeah, I understand there are reasons to do sustainability, but we can't go under."

For context, this company had netted $10 billion the previous year (not revenue, profit). Let's say that the company had promised Wall Street it would grow those profits at a modest 4% in the next year, that is, to $10.4 billion. Let's imagine now that the company had taken $100 million, an absolutely outrageous sum in the sustainability world, and invested in decarbonization or materials innovation or circular models for its products. If it had made real progress on decarbonization over the next 10 years, it would have less to worry about if, say, the price of oil suddenly spiked (to pick a metric of the moment). Today the company would be far more resilient, and it would be serving shareholders very well. And in that first year, its profits would have been $10.3 billion — quite a ways off from bankruptcy.

One of the main reasons we keep finding ourselves in this situation is a huge misperception about collective risk and the costs of action. Executives have long said some version of "but my shareholders" when faced with longer-term, collective challenges. The narrow focus on short-term shareholder value has resulted in the business community having a really poor record of managing systemic risks (or even just not making them worse).

Of course, with AI job displacement, the harm may accrue to society without ever landing back on the specific companies doing the displacing (unlike climate, where physical risk and regulation eventually hit the balance sheet). That asymmetry is what makes voluntary restraint so hard to enact and sustain, and why this may ultimately require policy, not just persuasion.

The positive interpretation for the selective blindness about collective risk and the undermining of shared resources is that every new direction, like AI, is exciting and impossible to forgo; a more realistic interpretation is that there's just money to be made in the current path, so collective well-being be damned.

A Call for Human-Focused Strategy

Could companies just decide that they won't trade people for AI? What if they didn't cut as many jobs?

It's possible that some companies would be less competitive, but it's unlikely that they would "go under," as my COO friend worried. There's a pretty big gap between today's record corporate profits and significantly worse results (let alone bankruptcy). At the same time, we truly and profoundly don't know what business will look like with AI acting as everybody's assistant — that is, “augmenting” their work, to use a rising phrase, instead of replacing it.

With the relentless pressure to cut costs and maximize profits, companies may feel like they're not in control. Talking about the role of business in society, shared prosperity, or everything under the banner of "sustainability" has been in retreat. Yes, there are a few signs that companies may be held accountable for more than their profits; the recent legal action against Meta for putting click and eyeball maximization ahead of children's well-being is one. But even with, for instance, significant financial benefits from transitioning to the clean economy, companies have collectively underinvested in action on climate change for decades. Can we figure out how to not make similar mistakes with AI?

In the end, every decision to invest or to not invest is a choice. I'll be honest about the tension here: I'm asking companies to accept potential (short-term) competitive disadvantages on the basis of uncertain future benefits and collective responsibility. That's a hard sell, and I don't want to pretend otherwise. But it's also exactly what we in sustainability have been asking from companies for years regarding climate change. As with climate, we need policy changes around AI to encourage collective action, but policy moves slowly, and decisions about AI displacing workers are being made now.

We know AI isn't going away. What it can already do can feel like magic. And its use will rise as companies mandate it and people discover what it does well, acting as their assistant, researcher, editor, and more. But GenAI has some serious issues and flaws, such as its tendency to hallucinate. And its footprint and effect on communities is enormous. I write this as a practitioner watching this unfold up close, and as someone who uses AI every day and is actively working on how to reduce its energy footprint.

But I have a sneaking suspicion that we will look back at early 2026 and kind of wish we had just stopped. Of course, this won't happen writ large. There is global geopolitical competition, and there are stunning amounts of money to be made.

We have the option to make wise, thoughtful choices about how we treat employees — you know, the people who actually make up a thriving economy by having jobs and disposable incomes to buy things. The leaders with the power to make the call about how people are cared for will land on their feet either way. The ones just entering the workforce — my son's generation — may not have that luxury.

On today’s episode of the Me, Myself, and AI podcast, Andrew Palmer, senior editor at The Economist, describes how organizations can experiment with generative AI while balancing speed, quality, and risk. At his own organization, Andrew and others test artificial intelligence with human oversight to develop editing and publishing efficiencies.

As the host of The Economist’s Boss Class podcast, Andrew speaks with leaders as well as early-career professionals, and highlights interesting insights from recent conversations around skills and hiring.

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

Transcript

Allison Ryder: How do we experiment with AI in ways that are productive but also safe? Today’s guest explains how he’s spurred new development projects with AI and recounts how various leaders he’s spoken with think about the technology.

Andrew Palmer: I’m Andrew Palmer from The Economist, and you’re listening to Me, Myself, and AI.

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

Today, we’re joined by Andrew Palmer. He’s a senior editor at The Economist, where he’s the author of the “Bartleby” column and the host of the Boss Class podcast. His current podcast season explores how the use of generative AI is changing management and jobs like ours. Andrew, welcome.

Andrew Palmer: Hi, Sam. [It’s] nice to be here.

Sam Ransbotham: Some of our listeners may not be familiar with The Economist or the Boss Class podcast. Can you give us a quick intro?

Andrew Palmer: The Economist, for almost all of its history, has been a weekly news magazine. Now, of course, we’re available in lots of different formats. We’re published out of London, but we’re global in our scope, and we cover economics, business, politics, science, technology, and much more. The Boss Class podcast is a serial narrative series podcast that I host on management and the workplace. We’ve had three series to date, and, as you said, the last one was specifically devoted to this thorny topic of generative AI in the workplace.

Sam Ransbotham: It is thorny, and I think you do a good job of exploring some of that thorniness. One of your colleagues, Ludwig Siegele, said on an episode — and I pulled this out — “The Economist embraces change. We think technology is good and should be used.”

I always find that pro-innovation bias a little bit interesting because I have a background in computer security, where we may have a little different bias about whether technology should be used. But in this case, I think I agree with The Economist. How would you describe the journal’s overall philosophy toward AI?

Andrew Palmer: I would say “open-minded experimentation” is probably the way to think about it. We have not rushed headlong into it. We have a variety of internal projects to see how we can use it in our journalistic processes. For example, we fact-check everything that we do. There’s a research team there, which has to pull through a ton of stuff. Is it possible to make their lives easier while still having humans do the critical work of checking?

Similarly, journalists have to conform to a style guide, a particular way of working. So can we make it easier for them to check that their copy is doing what it should before it gets to editors, who then are the humans in the loop? There’s a lot of internal stuff.

And then we have experimented with things like AI-generated transcripts of podcasts that are available to people on our site, and we have more secretive, “If I told you I’d have to kill you” kind of stuff around what we might be doing in two to three years. So there’s a whole panoply of things that we’re doing, but we’re always very, very clear that we have a particular brand associated with high-quality, human-intensive processes, and there’s a lot riding on us getting this right. We move fairly cautiously as well.

Sam Ransbotham: One of the ideas I think that came through [in] a few episodes is this idea of a “jagged frontier,” that artificial intelligence has really amazed you in some areas but [has] also been unexpectedly disastrous in others. How does that affect the way you think about experimenting?

Andrew Palmer: I think it probably comes back to that overarching mindset of being cautious so that you don’t thoughtlessly embrace the technology, let alone if it’s public-facing. So everything goes through an experimentation phase. One of the things that’s become apparent — and you’d see this in every kind of organization, I think, [that’s] grappling with this — is that you need to have really experienced people in the loop. For us, [those are] editors who’ve been in the newsroom for a very long time, working out what counts as quality, providing feedback on the experiments that we run so that over time it gets better and better and better, and asserting a pretty high bar for what counts as good enough.

That’s the way in which a mindset gets translated into actual processes for evaluating and checking. And it’s a new way of working for us. [For] most of our history, the journalists have kind of controlled absolutely everything, and now it has to be much more collaborative, especially with this technology.

Sam Ransbotham: What’s surprised you? What’s been the biggest positive surprise and maybe the biggest disappointment?

Andrew Palmer: I’ll take an example of my own, playing around from the latest season as a sort of roller coaster. So you know, almost every journalist on earth now is doing vibe coding. I’m no different there, but at the time, it seemed like we were really breaking new ground. That’s all recorded, but I’m a non-coder, absolutely no idea what I’m doing. You’re the expert in this conversation for sure.

We have a style guide, which I mentioned to you, which is basically our bible on how you should write, all the sort of grammar, hatred of Americanism you’ll be shocked to know — all sorts of rules like that, which you can thumb through a big PDF to get to. You can leave it to your editor, but ideally you would have something that could basically just check your copy against this. Ludwig Siegele, who you mentioned earlier, had been waiting for a year to get developer time to build this.

It was relatively simple, but [in a] busy organization, lots of people with different things to do, this was low down the queue. The magic of this was I went away, and in 75 minutes had built an extension, which did check copy against the style guide. Now when I say, “I built,” I was more like a sort of puppet. I had no idea really what I was doing, but by using Claude, this thing was generated. That was kind of amazing to me. I felt like I’d achieved something, which was totally beyond my purview, [and that] I just would have been utterly unable to do, and it clearly bypassed our internal bureaucracies.

Then, the kind of disappointment that you mentioned is that actually, we’re not going to be able to just push this out magically to people the next day. There’s an awful lot of governance to think around on this. The behind-the-scenes stuff around what architecture, what software, what our data processes are all had to be thought through. In practice, what I had built wouldn’t have worked scalably.

So anyway, at that point, a bunch of people who really knew what they were doing took it over. And it did result in something fast. It definitely accelerated the process, gave them ideas to work with, but it was also clear that I wasn’t going to be able to sort of magically bypass all organizational processes and change things.

I don’t know if that’s disappointing or not, actually, but there are other examples where this thing is just giving me kind of nonsense answers, hallucinating, all the stuff that you talk about in the show week in, week out. But I think that’s probably the best example of moving from this sense of euphoria, like a whole new world has opened up, to being brought back down to earth.

Sam Ransbotham: There’s a truism in software: If you ask any software engineer, “How far along are you? Is it close to done?” Oh, they’re 90% done. And that first 90% takes about half the time or a quarter of the time, and that last 10% is really hard, and it always takes a lot of time there.

And I think that’s what we’re seeing maybe with vibe coding. It’s doing that 90% pretty quickly. It’s making the initial screens or whatever, but like you say, there’s a lot more to that process. [When] I think about extrapolating with artificial intelligence, I think we have a tendency to draw those lines linearly or even exponentially, but diminishing returns may be the more normal shape, especially, as you describe there, with all the other processes.

Andrew Palmer: Can I ask you something, Sam? Are you seeing vibe-coded apps for yourself, and do you notice a difference?

Sam Ransbotham: I’m super into coding in general. That’s what I do as a hobby, and I love it. What I find is almost exactly what you’ve described, that this ability to prototype something quickly is amazing. You can just throw up something and get a sniff test of “Is this worth any further investment?” without, like you say, waiting on your bureaucracy to come along and take a year’s backlog.

But you can’t fool yourself into thinking that’s actually going to be production code. And so you learn a lot from that, but I don’t know about actually dropping anything in there into production, because by the time I get the code, I’ve rewritten every bit of it before it actually goes into something that’s production. But it gave me that sniff test pretty quickly.

Andrew Palmer: One of the people we spoke to was Anton Osika, who’s the boss of Lovable, one of these big vibe-coding platforms, Swedish based, and he had this nice phrase — “Demo, don’t memo” — as a way of thinking about this [that’s] really good for prototyping. Don’t do PowerPoints. Don’t write long documents. Just build the thing and show what it can do. But he was also very clear that you do not want to be putting stuff directly into production environments, which sounds totally consistent with what you’ve said.

Sam Ransbotham: I know you talk about jobs a lot, but one of the things you touched on there was this delay that you had waiting for the style checker to come along through the normal bureaucratic processes. And you were able to sort of get a quicker smell on that. My sense is that most organizations have giant backlogs of projects.

Andrew Palmer: I’m reminded of a conversation with Hannah Calhoon, who’s the head of AI at Indeed. You may indeed have had her on your show. [It’s a] big jobs marketplace, as everyone knows, and she described the organizational problems that come from [when] suddenly everyone can generate a lot of code, which means a ton of code has to be reviewed by people. It’s a bit like a sort of waterbed effect, right? I mean, the lumps of work are disappearing in one place, and they’re popping up somewhere else.

I guess the question is: I think you said that you would basically do the sniff test, but then you’re writing everything from scratch again. Is there a risk of going halfway down a route with vibe coding and then realizing, “Actually, we didn’t want to start from here, so we’re going to need to redo it again?” Thinking that through is part of the organizational challenge of this. There’s clearly a technical challenge, but actually [the way] we design things to avoid really bad outcomes is kind of also a blocker to fast progress.

Sam Ransbotham: But we don’t have to know that yet, though. We’re all, like you said, experimenting. Everyone’s experimenting with this. We don’t have to know exactly how these are going to fit into organizational processes yet without learning some. All these tools that we’re talking about, they’re going to improve. [With companies] like Lovable that you mentioned and others, they’re going to get better over time. I think we’re sort of in charge of our destiny here in terms of how that plays out and what we choose to invest in and what we don’t choose to invest in.

Andrew Palmer: I do agree with that in theory, and then there’s a whole load of incentives at work in the system, which in practice constrains your freedom to maneuver. So if you have a bunch of people in the C-suite who are under a lot of pressure to realize productivity gains, it’s quite possible that they’ll run. … We’ve seen it, right? Part of the story of the initial years of generative AI is running toward stuff without necessarily thinking about the organizational consequences.

Sam Ransbotham: Talk about those incentives. What incentives do you see out there within organizations that are in tension?

Andrew Palmer: Lots of boards are putting a lot of pressure on CEOs and their peers to come up with very material returns on investment. That can be cost savings in terms of letting people go, or it can genuinely be a jump in output. Neither of those things [is] necessarily great. It may be that you’re cutting people prematurely. It may be that you’re actually sacrificing quality over quantity.

I’m trying to think of an example. Johnson & Johnson was another guest on the show, and they had a very conscious, sort of “let a thousand flowers bloom” approach to generative AI in the initial years. I don’t think they had disastrous outcomes, but they also quickly found … this problem of bottlenecks emerging in various places.

It’s the waterbed again, right? That very big organization had lots and lots of people independently come up with the same kind of workflow to improve, so something like invoicing, making that faster, when all that would do is create a whole bunch of invoices that would then land on finance who weren’t expecting them, right? So there was a sort of chaos problem.

But more importantly, the lack of prioritization meant that there was a lot of work being done. Only 15% of projects were delivering 85% of the value. The rest of it, maybe there was benefit, people learning, etc. But in terms of a payoff, not so much. So they have now pivoted to a much more priority-driven, centralized approach where there’s a central AI council.

[It] sounds a bit Star Wars, [having] a central AI council, a central data council signing off on stuff. But it feels much more intentional, much more directed. You see surveys with C-level executives saying, “We haven’t seen massive productivity gains yet. In the next three years we expect to.” At some point, they’re either going to have to say, “Well, none are coming,” or they’re going to be bound into a kind of like “We’re going to make this happen.” The incentives to show results, they’re pretty powerful.

Sam Ransbotham: My more cynical take is I think we see a lot of job cuts due to AI announcements that I deeply suspect [are using] AI as a scapegoat. You have a choice of saying, “All right, I made a really poor decision and overhired,” or “I made really poor decisions or it’s this AI thing.” It’s really nice to point to the exogenous thing.

I want to come back to something that you mentioned before. We were talking about coding, but I think it’s a bigger issue in general. Maybe we could expand on that. Your waterbed is that you created an ease of creation, like with these tools that [have] the ability to generate. I mean, it’s even in the name: generative AI. We don’t have an evaluative AI. That’s not a huge massive trend out there. GenAI is the topic, not evaluating.

I think we’re seeing that certainly in science. Open repositories like [Internet] Archive have been overwhelmed with submissions, because the hard part is less about the generating and more about the figuring out what’s worth consuming, particularly in our time-based attention economy. How are we going to work around that?

Offer some hope. Do you have any thoughts about what’s going to happen about the world when we are able to generate everything so quickly?

Andrew Palmer: The example that comes to mind and the one that I just played around with just by myself was in the recruitment space, where there is this sort of uncontrolled generation of content on the candidate side, and then on the recruiter side, you are forced to use AI to cope with this bombardment. So there’s this sort of strange arms race.

My example of this was I just signed up to an auto-apply software provider, put in some sort of details, really scant details. … I could have spent more time on it. [I] sort of went off, puttered around for like an hour, came back, and found that I’d applied to a hundred jobs. I had no idea what they were, and they included being the head of operations for the city of New York and director of the Iran and Afghanistan Veterans [of America]. I mean, you can probably tell from the accent why I’m not the obvious person for this. It seemed sort of ludicrous, right, that you could fill up people’s inboxes. So I’d like to apologize to the Iran and Afghanistan Veterans association.

But, obviously, on the other side of that, then you need to have this automated response. No one is happy with this. It’s like a really bad equilibrium that’s been generated, and you can see that in other places, too.

So what’s the way out of it? I guess more humans might be one way out of it. You could intentionally insert humans into processes on the recruiter side and kind of see if that works in some way, but that doesn’t feel very scalable.

You could be transparent about your AI policy, saying, “It’s fine to use AI, but this is how we want you to use it,” or have an AI-specific kind of question that proves how you would use your AI fluency to [the] best advantage, whatever it might be. There’s some evidence — it’s anecdotal — that being transparent about AI use can actually reduce the amount of bogus applications. And then I guess the very long-term answer, which is a little too nirvana-like for me to buy totally.

Sam Ransbotham: Give us the dream.

Andrew Palmer: Well, this is what people in this world say: Eventually the AI is going to be so good that it is going to hunt out candidates. There’ll be what they call reverse apply. So you don’t, as a candidate, need to worry about applying to anything. The AI is going to sort of know from your entry on a site like Indeed exactly what you are suited for, understand your preferences and experience, and, basically, they’ll come to you — no need for all the kind of slop that’s already in the system. Maybe that will be the case over time. I’d really like to know whether you think that’s plausible, but it seems like we’re going to spend a long time getting there, and in the meantime, it’s kind of bad for everyone.

Sam Ransbotham: I think we [can] make some analogies to the deepfake detection that’s going on. Every time we improve the deepfaking, we improve the deepfake detection, and we go back and forth. So as you generate applications for a job, you’ll have the application for a job detector. We’ve seen this with, for example, search engine automation where you put the right keywords on your website, and you get “hired” in the search engines. We work through many dynamics like this.

I think actually on your episode you had a recent graduate, Kat Harrison-Gaze, who was talking about the experiences of applying [in] the job market. Can you give some grounding in an example?

Andrew Palmer: Kat was at Oxford, so in a U.K. context, [she’s at a] very, very prestigious university, clearly very smart, should be someone who employers want. There are lots of things feeding into this. We shouldn’t blame this only on AI, but she described two worries.

One is the process of applying and “How do I navigate this?” She has a suspicion of AI. She’s worried about cognitive dependence. She valued her own ability to think things through. It was like, “If you touch this thing, it’s going to infect me and change my ability to think,” was kind of part of it.

And then the second worry was just more generally, “If you think about a career in decades, where the heck do I put my chips? What is the career that makes sense going forward?” Both of those seem to me to be totally reasonable worries. It is really hard to navigate the recruitment process right now. I don’t think not touching AI is the answer, by the way.

So the advice to Kat was, “Use AI, don’t shy away from it. That’s not the way to think about it.” And the advice for employers was, “Here’s this very smart person who is determined to think independently. Take a look at her.”

Sam Ransbotham: What you’re describing, I think, is as we move toward machine-to-machine interaction … formerly, back in the old days, people would walk around from office to office and apply for a job and maybe drop off a resume and meet someone in person — that became online. What you’re sort of describing is a future where your factor talks to my factor; your agent talks to my agent. And much like the ballplayers who have agents to negotiate on their behalf, you’re kind of describing where you’d have recruiters negotiating on the behalf of the company, agents negotiating on behalf of the employer, and there’s not necessarily any need for those to be humans, and that can be a machine-to-machine interaction. Is that the future we’re headed toward?

Andrew Palmer: I kind of hope not. I guess it depends. … In the first stage, I can totally understand that. It’s almost essential, right, at this point, because it becomes hard to see how it can scale. But as long as there are humans in workforces — and, by the way, this is a process which is done exceptionally badly by humans right now. It’s a really difficult process to get right, but a lot of it can be improved by humans taking the time to test whether someone is a good cultural fit by being honest. “This job is good for these reasons but bad for these.” Those kinds of things don’t have to be done by humans, but they generally work better if you have humans talking things through. So humans have to be in the loop at some point, I think. As long as we are working with other people, [then] cultural fit, the values of an organization, all of those things are totally essential to a good hire.

You could imagine a kind of machine-led, skills-based process, testing whether someone can do the job. But, actually, what motivates them? Why are they joining? Would they be a good fit? Do I want to work with this person? All of those kinds of things. I don’t feel like an agent is the right way to answer those questions.

Sam Ransbotham: Maybe it’s just because that’s what I’m comfortable with in terms of people [who] I work with … but as you say, we do a poor job of that historically. We make biased decisions. We make decisions based off of attributes we should not make those decisions on. I find it somewhat appealing that perhaps the increased automation could help us at least see that or at least raise potential candidates [who] our biases may have kept us from.

Andrew Palmer: That’s true. [Psychologist] Danny Kahneman, obviously, was very, very pro the algorithm being in every process and would bet that was way better than any human. I think probably it’s a combination of both, right? I mean you have a rules-based algorithmic way of stripping out bias in the way that you ask questions, whatever it is. Still, fundamentally, there’s something important about getting on with someone [who] is quite an important part of a hire.

Can I ask about arms races? Sorry to kind of turn the tables, but you’re in cybersecurity, right?

Sam Ransbotham: Yeah, that was my dissertation research back in the day.

Andrew Palmer: That sort of arms race problem — you’ve got an AI sniffing out weaknesses and an AI trying to patch them. What’s the end point with that, and what’s the role of the human in that world?

Sam Ransbotham: I’d love to know the answer to what that’s going to be, but certainly it’s big. We’ve seen so much automation on things that used to be human penetration testings and these sorts of things. That’s all become largely automated, but there’s something fundamentally clever about people [who] figure out ways of both protecting and designing incentives that, I think, still [are] winning out in many ways.

Now, what happens is, as we get the incentives clear and we get a structured set of rules, the machine sort of takes it to the hyper-refined level, but then someone will come up with a different approach or a different idea. And, unfortunately, in security the common problem is that the humans are the weakest link. We could attack your password all day long, but it’s going to be much easier just to go in your office and take a look at what you’ve got written on the sticky note on your desk or trick you into revealing it. So I think we’re seeing that the mortals may be the weaker link here.

I mean machines because only machines can react that quickly. But I’ll pull back to one of your episodes on the Oura Ring. The one way of attacking a company would be a refund attack. I mean that’s not what we think of as a traditional cyberattack, but a refund attack might be complaining about something wrong with a product in order to get a refund. And you had the example of the Oura Ring. … Why don’t you describe it? The AI agent diagnosed the issue, checked the policy, and then ordered the replacement.

Andrew Palmer: Yeah, it was exactly that. We’re talking to Mike Krieger, who’s one of the people at Anthropic in their new products division, cofounder of Instagram. He was talking about what was on their road map, and obviously, inevitably, we started talking about agentic stuff. He recounted his best customer service experience at that point. The Ring [camera] seemed to have a battery problem, interacted only with an agent. And this thing basically sort of decided for itself [that he was] entitled to a new Ring. It asked for his address, and off it was packaged. And he regarded that as the single greatest customer experience he’d ever had.

To your point, though, what’s to stop the AI [from] just giving you batteries for life? That’s in the guidelines, right? It’s like if there’s a certain threshold number of or percentage rate of refunds that are given out by an agent, then you stop, and you’re kicking it to a human or another AI model.

So there’s a lot in the governance there. And on that I had a conversation with Bret Taylor, who’s the chairman of OpenAI but the [co]founder of Sierra, which is another customer service agentic startup. And again, metrics were super important in getting this right. So initially, they were thinking, “OK, so any call that doesn’t require the agent to hand off to a human is a successful call.” And then they realized, “OK, well, this is totally gameable. So we’ll just never hand off to a human.” Hundred percent success. Everyone’s happy. And obviously that’s not right. So now they have a combined metric of proportion of handoff calls but also Net Promoter Scores from customers who’ve experienced interacting with the agent. So that blended metric seems to work.

That feels totally obvious, but it is kind of the bread and butter of implementation of good management. What’s the problem you’re trying to solve? How do you measure success? It’s totally fundamental to this being got right. And that’s a simple example of how they’ve got to a decent metric.

Sam Ransbotham: Yeah, that measurement is a big thing. I think one thing that’s happening is that these tools are helping us measure things differently, so we can measure things we never could measure before. We’re collecting a lot more data, and that’s great. But it is pointing out that many of the existing measures may be … I think your phrase was gameable, that once people figure out what that metric is, then they will do something.

Let me switch to you. Our show is Me, Myself, and AI. How did you get interested in this? Tell us a bit about your background.

Andrew Palmer: I’ve been with the magazine since 2007, [and I’ve had] multiple writing and editing jobs covering a variety of things, from Latin America to finance to [our] data journalism team, Britain [editor], etc.

Most recently, I’ve been writing on management and work. And most of what we do here at The Economist is look from the outside in. It’s big impersonal forces at work, macroeconomic, geopolitical, technological. I’m the person inside the workplace looking out, and sort of looking at humans as a byproduct of that, like how we all interact.

I’ve been doing that for a while, and very obviously, AI is affecting the workplace and affecting us as employees and affecting managers, and raising all sorts of questions. So it was a very natural thing for me to start to write about it. And I think it’s one of those super-interesting intersections of you’ve got this incredibly scientific, cold sort of machine technology, and then you have this unbelievably messy soup of emotions, which is what a human is, and putting them together is really interesting to observe.

[Here’s] a little sneak preview, but it won’t be a sneak preview by the time this goes out: I’m writing this week on the one thing that everyone can agree on. It’s great that AI is going to get rid of grunt work or drudge work. I’m not totally sure about that, because, for humans, drudgery, in the right dose, is really good. It’s really good. There’s a bit of agency, because you can get stuff done. You can kind of relax a little bit because you can’t be tote on the entire time. There’s some evidence that mind wandering is really good for creativity. So there’s a problem with the idea [that] we’re all going to be maxing out on higher-order tasks the entire time. We’re just not built for that. So that’s the kind of territory that I’m in.

Sam Ransbotham: I think that’s pretty fascinating. It is true that only by sort of pausing and thinking and reflecting, and it takes a bit of not constantly being on, to have those sorts of ideas. That’s certainly counter to the “Oh, you’ll get rid of the drudgery.” Now, on the same hand, there [are] certainly lots of our jobs that are true drudgery, and so I think the trick is we try to figure all this out. How much drudgery is the Goldilocks amount?

Andrew Palmer: Again, you get back to incentives. A manager of a certain mindset might think, “OK, puttering around is not something I want anyone on my payroll to be doing ever,” right? “Let’s go out there and do sort of cognitively intense stuff all the time. It’ll be amazing.” So there’s also a kind of mindset shift there, like we couldn’t get rid of drudgery before — it was just part of life. What if it’s an option to get rid of it all? How would you think about that?

I spent some time at an air traffic control center once, where their job is to think about, “What’s the optimal performance environment that you do not want these people to be below their A game?” So there’s a really interesting balance they’re trying to strike there, between you don’t want to overload, so [you need] limited session timings then and mandatory breaks, but you also don’t want to understimulate because [with] too much boredom, basically your attention starts to veer off in ways which are not great if you were in charge of air traffic. So they have thought about it, right? It’s sort of a human factors discipline. And, weirdly, it’s probably coming to every office, just in a much less high-stakes way.

Sam Ransbotham: One of the things that you mentioned there was that before, we didn’t have the ability to automate these things, and so we didn’t have to make any choices. Now that we have these abilities, we have to make some hard decisions. What kind of skills help people make those choices? I’m in a university, help me out here. What should we be helping students learn? What kinds of skills help them make those sorts of decisions?

Andrew Palmer: Oddly enough, I use the term human factors, which is its own discipline, right? But it’s always been quite narrowly defined, sort of how do you get the best out of an air fighter pilot or whatever it is? But I do think there’s something about understanding human performance that is really important in understanding how the sort of complementarity of humans and machines works. There’s something around management discipline itself, right? How do you design a good process? How do you avoid bottlenecks, right? Regulating the flow of work in a way that you don’t have everything just bunching up somewhere else in the system requires you to think at an organizational level and to think about systems and processes.

So, basically, systems thinking, all management training, all of that is super useful, I think, in thinking this through, and a bit of introspection. Maybe the best thing about this technology is that once you start to use it, it forces you to be introspective about, “What am I good at? Where am I likely to have a sustainable advantage? What do I like to do? What do I not? Do I really like the idea of doing a hundred percent higher-order stuff? Is that credible for me?” I don’t know how you train self-awareness, but in the process of trying the thing out, you start to, inevitably, I think, have those kinds of internal conversations, and they’re useful.

Sam Ransbotham: Even to answer those questions, we’re going to need a lot more information about people and about how they work, and that’s a little bit back to your measurement thing. That in many ways comes in conflict with my desires to keep my own personal information guarded. At the same time, these systems could probably help me understand a little bit about my own self, and how I work, and [in] what situations I work best.

I think, overall, that one of the things I think you’re bringing out in many different cases and many different examples is this need for nuance, this need for not always going too far, not always doing too little. Experiment some, but don’t over-rely. I think that’s really the overall theme that I got from your podcast, which is this idea of bringing some nuance to all these decisions. And I think that nuance doesn’t always play well in our current “Do these 10 things to make you a better AI person.” I think a theme from what I’ve pulled from your work is clarity and nuance. And I think that’s really hard.

Andrew Palmer: It is really hard. Another way of framing it, but it’s like the worst way to market the podcast, is to be boring. So master the essentials, the basics, right? What is the thing that you’re aiming for? AI adoption is not a metric in its own right. So what’s the problem you’re trying to solve? Think about workflows end to end rather than a single thing, where work might just be being redistributed. All of those kinds of questions are just common sense, but as usual, that takes you quite a long way.

Sam Ransbotham: Andrew, thanks for taking the time to talk with us. If listeners want to hear your podcast or hear an experiment where you’ve created a bot to create your own voice in a podcast — I thought that was a fun example — Season 3 of Boss Class is going on right now with The Economist. Thanks for taking the time to talk with us.

Andrew Palmer: Thanks, Sam. It’s been fun.

Sam Ransbotham: Thanks for listening today. On our next episode, we’ll shift gears to health care and speak with Carla Goulart Peron, chief medical officer at Philips, about how AI is enhancing, not limiting, the human element in the medical space. Please join us.

Allison Ryder: Thanks for listening to Me, Myself, and AI. Our show is able to continue, in large part, due to listener support. Your streams and downloads make a big difference. If you have a moment, please consider leaving us an Apple Podcasts review or a rating on Spotify. And share our show with others you think might find it interesting and helpful.

Disability-related innovations are all around us. Curb cuts in sidewalks, originally designed for wheelchair users, benefit caregivers with strollers, travelers with suitcases, and delivery workers with hand trucks. Automatic doors intended for individuals with mobility impairments are convenient for all. Blurred backgrounds in video calls, standing desks and ergonomic keyboards, and speech and voice recognition tools were all designed to assist people by minimizing distractions, easing lipreading, reducing chronic pain, and supporting people with mobility impairments — and all are now widely used by the general public. Every day, people with and without disabilities use numerous innovative accommodations that have become indispensable mainstream tools — to such an extent that few people realize that the features were originally developed to address disability-related needs.

In short, what's often labeled a burden can be a source of practical innovation. But many managers still view disability at work through the negative lens of cost and compliance. Our research suggests a more positive, generative perspective.1 When a team includes someone with a disability, coworkers often view their own work with fresh eyes. They notice previously overlooked inefficiencies and barriers, question operating assumptions about how tasks "must" be done, and propose better ways to design work. Those changes typically improve work for all, making the job easier and safer for everyone, not just the person with a disability who needed an accommodation.

Functional impairments associated with disability, then, can signal suboptimal job design. Many workplaces are implicitly built for an "ideal," able-bodied worker who never tires, strains, or loses focus. Designing for a broader range of workers is not just fair; it's a necessary way to reflect reality. For example, as workforces age — a looming reality for many industrialized countries — jobs that function effectively only for the ideal worker will become harder to staff and sustain.

From Accommodations to Task Redesign

The usual organizational response to disability-related functional limitations is to grant an individual an accommodation so that person can keep doing the job as designed. While this fulfills legal requirements, it misses a larger opportunity: Treating a functional limitation as a spotlight can reveal where job design may be ineffective. For example, coworkers may step back and ask simple, practical questions: Why is this light so bright or this office so noisy? Why must this process be carried out in one continuous stretch? Why do we lift heavy objects here at all? Why is this reach overhead? Such questions can lead to changes in work design that reduce strain and errors and — importantly — improve conditions for all team members.

To illustrate, imagine a manufacturing job that regularly requires workers to lift 40 pounds unassisted. An employee with chronic back pain cannot safely do that and, therefore, requests an accommodation. The conventional response would be to treat the problem as individual and exceptional: to reassign duties, add a second person to assist with lifting, or add in breaks for recovery.2 A more sustainable response would be to redesign the work process itself so that the task no longer depends on unassisted human strength. That might involve the use of load-sharing equipment, height-adjustable fixtures, or a newer tool, such as industrial exoskeletons, to offload spinal strain.

Initially introduced to enable a single employee with an impairment to perform a job, such redesigns have broader value: coworkers experiencing less fatigue and a decline in injury risk overall. What begins as a disability accommodation becomes a more effective way of organizing work for everyone.

Innovation to Decrease Inefficiencies

When teams begin to brainstorm possible improvements, they often suggest incremental adjustments within the existing system, such as ways to make it easier to lift tools, rather than questioning current norms, such as asking why tools are stored at floor level in the first place. It's common for people to take their current job setups for granted and rely on well-worn patterns and processes.

But when employees work alongside a colleague with a disability, inefficiencies become more salient. Disability can function as a prompt for people to rethink a task entirely and make new, creative connections between formerly disparate concepts, leading them to adopt ideas from other areas of life. Perhaps the magnetic strip they use to store knives along their kitchen wall will inspire a new arrangement of work materials at waist height rather than on the floor, effectively bringing the tools to the worker rather than the worker to the tools.

Taking the perspective of a colleague with a functional limitation can be the genesis for higher idea counts and greater idea novelty in the office, just as on the production floor. Once teams stop accepting able-bodied and neurotypical defaults as inevitable, they start proposing solutions that challenge the job's inherent design and could improve outcomes for all workers, such as in these examples.

Documented workflows. Much of work relies on tacit and ambiguous knowledge, which becomes more obvious when a neurodivergent employee asks for clearer instructions and fewer unwritten rules. Structured processes designed as an accommodation — step-by-step guides, documented workflows, and simplified interfaces — can lower the cognitive load for the whole team, resulting in more effective knowledge sharing, fewer errors, and smoother collaboration.

Enhanced audio tools. Captions in video calls highlight how much workplace communication depends on audio-only information, which can be a challenge for colleagues who are deaf or hard of hearing. Originally introduced as an accessibility support and now built into most video-meeting tools, captioning creates searchable transcripts, improves attendee comprehension, and makes it easier to follow along in noisy or distracting environments. The result is easier documentation and more inclusive, efficient communication overall.

Structured brainstorming processes. "Think fast" dynamics can silence good ideas and stymie participation efforts. During fast-paced brainstorming and immediate critique, some colleagues may take the floor as others tend to withdraw, due perhaps to an anxiety disorder, fear of public speaking, or simple introversion. Structured practices — such as two-phase ideation (silent generation followed by later discussion); anonymous digital brainstorming, which separates ideas from the individual; and feedback templates that pre-structure forms and ways to provide feedback — can increase psychological safety. The result is broader participation and more generative, collaborative meetings not dominated by the fastest thinkers or the loudest voices.

What Leaders Can Do Right Now

Designing with disability in mind uncovers friction points, streamlines processes, and enhances the work experience for all. But managers may face some common concerns.

• "Will this slow us down?" Not if you keep changes small and reversible. Many tests take little time to set up and can run during normal operations. The goal is to save time by removing wasted motion and to prevent errors that may require hours of rework later.
• "Won't people feel singled out?" Keep the discussion about the job, not the person. Use generic prompts ("assume no overhead reach") to depersonalize the analysis. Participation by anyone with a functional limitation should be voluntary. The aim is safer, steadier work for everyone.
• "We tried a suggestion program and it fell flat." This is not about collecting more suggestions. It is about exploring more kinds of potential solutions, promptly trying them out, and integrating what works into existing operations.

To more successfully tap your team's creativity in rethinking job design, we offer the following suggestions.

Involve the people who live with the consequences. The best redesigns come from the people who do the work: the employee who raised the issue, two or three coworkers, their supervisor, and a safety or ergonomics partner. Ten minutes at the workstation beats an hour in a conference room. Keep the tone neutral and the focus on the task: What does the job need to look like so more of us can do it well and safely?

Run short "assumption reviews" on your highest-friction tasks. For primarily physical jobs, identify a work task with frequent near misses, rework, or strain complaints. Film 30 to 45 seconds of the work being done, and then watch the clip with the people who do the job and a few of their peers. Then ask questions about what you see. For example, you might ask: What would we change if no one could lift more than 20 pounds? If overhead reach were not an option, how would we set this up? If glare and noise were dialed down, what would change? If we had to add a short pause every 20 minutes, when should it occur?

For desk jobs, an employee might record their screen as they complete a given task, or document each step and subtask if making a recording isn't feasible. Then have team members review and investigate, asking questions like: How many different tools are being used to complete this task? Are they all needed? Is there any manual duplication or cutting and pasting that can be eliminated or automated? Is every subtask necessary? Is there any unnecessary data being entered and tracked?

Then ask for improvement ideas, aiming for breadth. Collect ideas in buckets: for example, for computer-based tasks, creating templates, introducing automation, or discontinuing the use of duplicative tracking tools; and for manual tasks, considering ergonomics, assessing tools and fixtures, or rethinking pace and scheduling. Encourage more categories of solutions, not just more versions of the same one. Then select the most promising ones and test and fine-tune them to arrive at the best solutions.

Treat accommodation requests as design leads, not paperwork. When someone requests an accommodation, walk the job together and capture the underlying friction in plain terms. Turn that into a design hypothesis: The load needs to be at shoulder height to ease strain, the door needs to be closed to accommodate hearing disabilities, or the light needs to be diffused to reduce eyestrain. Try the smallest change that could work. If the change makes the job better for the original worker, keep it. If it doesn't, debrief what you learned and try the next simplest idea. This way, you'll learn a lot more about the intricacies of the job and can identify promising directions for job redesign.

Rather than treating disability solely as an exception to be managed, try thinking of it as natural variation and a prompt for redesigning work. With this mindset, your team can find better ways to do the work that benefit everyone. Start small, with one high-friction task. Look at it through the eyes of someone who cannot do it the way it is designed today. Try a couple of small changes. Keep what works.

That simple practice — noticing, questioning, experimenting, and adopting — is how practical innovation can thrive. Thoughtful organizational and societal design with and for people with disabilities often leads to improvements that benefit everyone. Over time, improvements will compound and garner lasting advantages and ongoing innovation.

Operational efficiency is critical for both financial success and customer satisfaction. Efficient systems, characterized by minimal buffers and idle time, tight schedules, and maximum asset utilization, allow organizations to do more with less, thereby boosting revenue and appealing to time-sensitive customers. However, such systems often lack resilience, increasing an organization’s vulnerability to operational disruptions.

The tension between efficiency and resilience is especially visible in the airline industry. A resilient airline network can absorb disruptions, protect passengers from severe service failures, and recover quickly without incurring excessive costs. But airlines also face constant pressure to offer faster itineraries and maximize the use of costly resources, such as their fleet of aircraft and flight crews. Meanwhile, passengers strongly favor efficiency as well, in the form of shorter travel times with minimal layovers. Those preferences can lead to itineraries with little to no time buffer to absorb the delays or cancellations common to air travel, leaving passengers frustrated and stuck waiting in terminals. Moreover, such disruptions propagate throughout interconnected networks, affecting passengers, flight connections, crew schedules, and aircraft positioning. These ripple effects result in significant financial and reputational damage for airlines.1

This challenge is not unique to airlines, however. Supply chain managers need to balance inventory costs against the risks of stockouts. Health care systems strive to optimize patient flows and increase throughput while maintaining quality of care. Despite the operational differences across these contexts, the fundamental challenge is the same: How can organizations design operations that are both efficient and resilient?

Our analyses of millions of flights and airline passenger journeys in several academic studies reveal why managers do not need to treat efficiency and resilience as opposing goals. We identify three actionable strategies that enable organizations to achieve both objectives, whether managing flight schedules, patient flows, call center operations, or global supply chains.

Strategy 1: Measure What Matters to Customers

Traditional operational performance metrics often do not reflect customer experience, leading to perverse incentives that can weaken actual service quality and system resilience. For example, in the U.S. airline industry, the Department of Transportation (DOT) publishes its Air Travel Consumer Report each month. The report includes statistics for all major carriers on the percentage of flights that arrive within 15 minutes of the scheduled arrival time. Known as on-time performance (OTP), this metric serves as a proxy for service quality and reliability in the DOT’s and others’ rankings of airlines.

At first glance, publicizing OTP appears to benefit consumers: One might expect that being judged on this metric incentivizes airlines to reduce flight delays. However, as is often the case with KPIs, organizations give in to the temptation to game the metric: Airlines often add time to the flight durations they publish, to improve their chances of being “on time” — a practice known as schedule padding.2 This helps airlines boost their OTP statistics without meaningfully improving their reliability.

Putting the airlines’ strategic gaming behavior aside, it is questionable how useful a measure OTP is for customers. Non-stop-flight passengers are likely to experience a 20-minute-late arrival as only a marginal difference, and they’re unlikely to be concerned that their flight missed the scheduled arrival time by 14 minutes but was “on time” within the DOT’s 15-minute cutoff. However, for a connecting passenger with a tight layover, a 14-minute delay may be long enough to cause a missed connection — and hours spent in the airport waiting to get onto another flight.

Given that OTP does not accurately capture what matters to passengers, in a recent study we published, we urged the DOT to release more informative passenger-level statistics, such as the proportion of passengers reaching their final destinations within defined delay time intervals (within 15 minutes, 15 minutes to one hour, one to two hours, two to three hours, and more than three hours of the scheduled arrival time).3 In the long run, shifting the focus to passenger travel times (including total flight times, layover times, and potential missed connections and flight delays in each leg) can incentivize airlines to pay more attention to passengers’ travel experiences. Both efficiency and resilience could improve as a result, because airlines would have a stronger incentive to create more efficient routes and itineraries, with shorter total travel times that are less susceptible to missed connections and long delays.

Similar issues arise in other industries, where the operational performance metrics an organization uses can negatively affect efficiency and reliability. For example, in health care, hospitals often track operating room utilization rates as a key efficiency metric. While high utilization sounds desirable, it can incentivize hospital administrators to create extremely tight surgery schedules, with back-to-back procedures. This practice leaves little buffer for unexpected complications or overruns, causing subsequent patients to experience long delays or even cancellations and negatively affecting efficiency, care quality, and patient satisfaction.

In supply chain management, companies typically use high inventory turnover ratios as a proxy for operational efficiency. However, an excessive focus on this metric can prompt businesses to cut safety stock too aggressively, heightening the risk of shortages, delaying downstream production, and leaving customers with long wait times for the products they want to purchase. And in customer service, when call centers measure agents’ performance by the number of calls they handle, that may motivate agents to rush their interactions with customers. That, in turn, can lead to a reduction in problem-resolution quality and an increase in repeat calls, thereby prolonging the total time it takes to resolve an issue. These examples illustrate how narrowly defined metrics can distort incentives, prompting companies to optimize for the metric rather than the actual service experience.

Defining the right performance metrics can be challenging because metrics must balance an organization’s efficiency objectives with other managerial considerations (such as service reliability, customer satisfaction, and fairness). One way to attain this balance and ensure that high efficiency does not come at the expense of customers is to treat performance measurement as a paired system, with one metric that tracks operational efficiency and another that measures what matters to customers. For example, in our recent study, we developed two metrics — one to track efficiency and the other to assess resilience — to ensure that an airline’s focus on efficiency (measured by short scheduled travel times for passengers) would not lead to long travel delays due to missed connections.4 Similarly, hospitals can measure both operating room utilization and the delays surgical patients experience when extremely tight surgery schedules are disrupted.

Appropriate performance metrics should be easy for stakeholders to understand while accurately capturing an organization’s performance objectives and customer satisfaction levels. Defining such measures requires input from key stakeholders, including customers. Misaligned incentives pose another risk, especially when metrics are tied too rigidly to employee performance evaluations: Such correlations may incentivize employees (or even an entire organization, as documented in the airline industry) to game the system rather than improve service quality. Thus, designing incentive structures that reward long-term service quality based on a combination of efficiency and resilience rather than short-term metric gains is another important step. Finally, periodic reviews and stakeholder input can help ensure that metrics evolve with changing customer expectations and operational dynamics.

Strategy 2: Avoid a One-Size-Fits-All Approach and Deploy Buffers Strategically

Measuring what matters to customers can motivate an organization to be more proactive in preventing major service or supply failures and improving its resilience to disruptions. In the airline industry, major service failures often take the form of flight cancellations and long delays, which hurt airlines’ current and future financial performance.5 While some disruption triggers, such as severe weather, are beyond companies’ control, airlines can influence how disruptions propagate through their networks by strategically designing flight schedules.

Strategic scheduling involves building sufficient flight and ground time buffers to reduce the risk of delays cascading from one flight to the next. Traditionally, some airlines have relied on simple rules of thumb, such as using historical data to estimate the average time to complete a flight and then adding a fixed buffer. For example, if a flight averages one hour, adding a 15-minute buffer results in a scheduled flight time of one hour and 15 minutes. However, this one-size-fits-all approach can negatively affect both efficiency and resilience: A buffer may be unnecessarily long for some flights, reducing efficiency, but insufficient for others, increasing an airline’s vulnerability to disruptions.

A more strategic and data-driven approach considers both the likelihood and consequences of a disruption. In the airline industry, our analyses revealed that airport congestion levels, layover times, weather conditions, and the time of day predict the likelihood of a disruption.6 Moreover, the operational consequences of a disruption vary by flight: A delayed flight with many connecting passengers can create severe ripple effects, whereas a delayed aircraft on its last flight of the day, carrying no connecting passengers, will have minimal impact on the network. Allocating larger buffers where disruption risk and impact are high, and smaller buffers where they are low, can improve efficiency and resilience simultaneously.

The same principles apply to other industries. In health care, setting the duration and sequence of surgeries based on patients’ risk profiles and the consequences of potential delays to subsequent procedures can improve overall system performance.7 In call center operations, staffing schedules can be determined by considering not just the expected length of each call but also the likelihood of follow-up interactions resulting from unresolved issues. In supply chain management, rather than applying uniform inventory policies across all products, companies can tailor their safety stock levels and replenishment intervals to account for supply disruption risks and their downstream consequences, such as production delays due to material shortages. In project management, allocating contingency time to tasks with high interdependencies or critical-path activities can prevent small delays from cascading into major schedule overruns.

Deploying buffers strategically is easier said than done. It requires granular data, advanced analytics, and coordination across multiple business functions (such as marketing, network planning, and an airline’s ground operations teams). Historical averages and one-size-fits-all approaches are simple and familiar, which makes them hard to replace with dynamic, context-specific, risk-based approaches. Moreover, managers often face resistance when buffer adjustments appear to reduce efficiency in the short term, even though they can improve both efficiency and resilience in the long run. To overcome these challenges, organizations should start small by simulating what-if scenarios and piloting data-driven scheduling in high-risk areas within a division. They should also invest in predictive analytics to identify disruption patterns and communicate the long-term benefits of resilience to stakeholders. Embedding analytics into planning processes and getting buy-in from stakeholders can ensure that buffer allocation becomes a strategic lever rather than an ad hoc decision.

Strategy 3: Curate Personalized Customer Options to Maximize System Performance

In many organizations, operations teams prefer to limit the number of product or service options presented to customers to keep processes streamlined. Sales and marketing teams often advocate for more options to increase the likelihood of meeting diverse customer needs. However, offering a broad choice set significantly increases operational complexity, which can compromise reliability.

In the airline industry, itineraries with layovers highlight the risks of offering too many options to passengers. For instance, consider a passenger who planned to travel from Knoxville, Tennessee (TYS), to Greensboro, North Carolina (GSO), on Sunday, Jan. 11, 2026. Upon searching for flights on a popular travel website, we found that American Airlines offered 13 itinerary options via its hub in Charlotte, North Carolina (CLT), including one with only a 30-minute layover, where the first leg would arrive at CLT at 1:38 p.m. and the second leg would depart for GSO at 2:08 p.m. Similarly, Delta Air Lines offered six options through its hub in Atlanta (ATL), one of which allowed just 40 minutes between flights, with the first leg arriving at ATL at 5:13 p.m. and the second leg departing from ATL at 5:53 p.m. Given the sizes and congestion levels of CLT and ATL, both itineraries posed a significant risk of a missed connection. More broadly, airline reservation systems routinely display such options as long as they meet the minimum connection times published in the International Air Transport Association database, even when they leave little margin for actual delays.

In our recent study, we used a proprietary passenger-level data set provided by Southwest Airlines to simulate the impact of tight layovers on system performance.8 Our analysis found that identifying itineraries with short layovers and preventing passengers from booking them can significantly enhance resilience by reducing missed connections. Notably, curating the choice set by removing risky itineraries does not materially deteriorate efficiency, given that switching from an itinerary with a short layover to one with a slightly longer layover typically increases the total scheduled travel time only marginally (by 10 to 15 minutes, in many cases). Indeed, the actual travel time of an itinerary with a slightly longer layover may be significantly shorter due to the elimination of a missed connection.

The concept of curating customer choices to enhance resilience can be applied to many industries. In health care, to reduce the risk of cascading delays, hospitals could limit scheduling options for elective surgeries when operating room capacity is constrained. To avoid missed commitments, retailers might restrict delivery time slots during periods of high uncertainty in supply chains. By strategically shaping the choice set rather than leaving all options open, businesses can mitigate vulnerabilities without materially compromising customer experience, especially when the curated choice sets impose only minor trade-offs in efficiency.

Implementing curated choice sets is far from straightforward. It requires predictive models that accurately assess risk under varying conditions, as well as real-time systems to update available options dynamically. Internal tensions naturally arise when multiple functions, such as marketing, operations, and IT, must agree on where to curtail customer choices. In particular, marketing teams may worry about lost revenue from eliminating certain options. It is thus essential that organizations quantify both the revenue potential of risky options and the actual costs of disruptions that arise as a result of offering them. (For airlines, paying for stranded travelers’ hotel accommodations, meal vouchers, and rebooking expenses erodes profitability.) Organizations should first pilot curated options in high-risk contexts to validate their benefits. They can use the same analytical insights that helped them curate options to explain to customers that their choices have been limited to safeguard reliability.

In an era of rising customer expectations and increasingly complex service systems, overcoming the traditional efficiency-resilience trade-off is a critical operational skill that can define the trajectory of a service organization. Companies that can deliver both speed and reliability will not only meet customer demands but also differentiate themselves in highly competitive markets. The key managerial takeaway from our research studies collectively is that organizations should proactively design their operations to build resilience into their systems rather than relying on reactive, ad hoc fixes after disruptions occur. Such proactive design requires that they choose performance metrics that reflect customer experience, build systems that can absorb variability, and shape customer choices so that the organization continues to run reliably when conditions become challenging. As disruptions become the norm rather than the exception, reconciling efficiency and resilience by design rather than reaction can separate companies that cope from those that succeed.