How to put AI to work in your organization

GenAI can revolutionize business, but managers need to roll it out sensibly. Warning: it may get messy.

By Raffaella Sadun

Today, generative artificial intelligence (GenAI) seems to be all that executives talk about, often in terms of it replacing everyone’s jobs or the massive impact it’s going to have on productivity. I’ve been studying technology adoption for 20 years, and in some respects this kind of talk makes me feel like I’m trapped in a time warp: A new technology arrives, we feel that something quite powerful is about to happen, and yet its impacts are still very uneven across the economy. This is pretty much the same script we lived in the ’90s when we were grappling with the arrival of the first wave of information and communication technologies (ICTs).

There’s a lot of hype right now, but so far the tremendous uptake of GenAI has happened mostly at the individual level. Although most executives believe we are in the presence of a technology that will have incredible impact on economic activity, the adoption of GenAI for the production of goods and services across firms stood at around 10% at the end of 2025, according to U.S. Census Bureau estimates. This is much lower than the adoption rate we see for individuals. We also see wide differences across and within industries. For example, the uptake is at least three times higher in ICT-producing industries. And while there have been some effects on employment for exposed occupations, to date the aggregate employment and productivity effects of GenAI are not yet clear.

Why do we see such uneven uptake of GenAI in firms? Part of it is down to the fact that the productivity impact of new general purpose technologies (such as GenAI today, and ICTs and electricity in the past) typically follows a J curve: In the initial stages of adoption, productivity sees an initial dip or lull, as organizations engage in a process of exploration, before the curve rises steeply and takes off, when the ways in which the technology can be used are clearer.

This is where managers should be focusing their energies right now: not on the hype of immediate productivity gains or apocalyptic scenarios of massive job losses, but on thinking very carefully about how to integrate GenAI into the firm and, most importantly, how to reduce and ideally eliminate the likely frictions they will encounter as they adopt GenAI in their firms. This is key because the adaptation part of GenAI is messy. Yes, GenAI can do incredible things, but whether this translates into competitive advantage depends on how the technology is integrated in the organization, so that its capacity is genuinely empowering and creates real value for firms.

Don’t get too excited yet

We need to start by appreciating what GenAI can do. The most exciting applications we have seen so far are of GenAI as complements to humans, acting as copilots able to raise expertise levels. A well-known study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that a GenAI copilot in a call center not only raised the productivity of all employees, but the productivity increases were even greater among employees with the least experience.

I’ve done similar experiments to test whether AI could similarly improve innovation outcomes in a fast-moving consumer goods company, and found that individual experts in either R&D or Marketing coupled with GenAI copilots were able to come up with innovative ideas that were as good as the results given by human teams of experts in the same fields.

This notion of GenAI as a cybernetic teammate, which can help augment the quality and the productivity of experts, opens up exciting possibilities. After all, if organizations are limited by knowledge and time constraints, then having a tool that allows you to unlock those bottlenecks suggests the sky is the limit in terms of what companies can do. That’s the good part of the story.

But there’s a less rosy side, and that’s the substitution effect of GenAI for human workers. The rise of agentic GenAI makes this a real possibility, and we already have examples of companies that have gone deep into this application of GenAI, such as the fintech company Klarna, which eliminated 700 customer service jobs and replaced them with AI agents (only to later hire some human workers back).

Both augmenting and replacing applications of GenAI at the task level, however, may not immediately translate into organizational gains. Why? Because general purpose technologies like GenAI require organizational and managerial complements in order to be effective. Without them, you are not going to reap the technology’s full benefits. Discovering and adopting those complements is not an easy feat, especially because the ways in which GenAI creates value in a firm are typically idiosyncratic to the organization, meaning that they depend on the specific culture, management and skills that exist in the firm.

This is why the adoption of a general purpose technology may be followed by a decline in productivity, the dip in the J curve. It’s also why it took 40 years before the advent of electricity showed up in productivity statistics at the aggregate level, because the entire way firms were organized had to be reconceptualized first, before the benefits became tangible and visible.

The organizational mindset is critical

Given this reality, the organizational perspective is the real lens through which we can understand and perhaps predict much better what’s going to happen with GenAI. There is plenty of labor economics research on how automation technologies change the composition of tasks and the effects on workers before and after the arrival of a new technology, but not so much about the messy transition period, and that’s really what we need to explore and understand.

Still, there are lessons that we can take from the adoption and diffusion of ICTs 20 years ago that echo today’s story. In my own research with Nick Bloom and John Van Reenen, we found that U.S. multinationals achieved twice the productivity return from IT investment compared with domestic firms in the same industries and countries. The difference was not the technology itself, but management practices: structured approaches to promotions, rewards, hiring and performance management. Management was the complementary input that made IT productive.

If management practices explain differential returns to IT, we should expect the same for AI, perhaps even more so given that AI is an even more general purpose technology. Based on my collaborations with multiple organizations, I have detected some managerial interventions that may facilitate, not just GenAI adoption, but the whole learning process necessary for organizations to make the most of GenAI today.

AI experimentation: rethinking the rollout

To really understand how GenAI works and whether it generates cost benefits or better quality, you need to have people in your organization experimenting with these technologies in the actual flow of work. But not everybody is an experimenter, and not every firm has learned how best to “manage” experimentation.

Typically what happens is somebody in the firm has spent nights trying out different GenAI applications. They become GenAI champions, an important driver of experimentation. The problem is that companies then get carried away and just roll out these technologies in an untargeted fashion across the whole firm. A copilot for everybody!

However, according to a case study I recently did on this, it turns out that this massive rollout can backfire spectacularly. Sometimes this is because there is a big mismatch between the expectations and the maturity of the technology. The more people start playing around with it, the more they start to perceive that this “genius in their pocket” is actually a little bit stupid or more like a “hallucinating intern,” so they prematurely reject it, and never start the necessary work of adapting the technology to their needs.

The underlying issue is that companies often massively roll out GenAI applications that are not yet ready for prime time to employees who may not have the time, skills or, more importantly, the right attitude to engage with adaptation in the flow of work. Workers who face real time constraints, or who are already very productive without GenAI (hence, whose time is extremely valuable), face the highest experimentation cost. Besides the opportunity cost of their time, they may see GenAI as a possible threat to their expertise, and therefore resist the notion that they should invite GenAI into their work routines.

In the multinational company I studied, we followed the rollout of a sales copilot across the whole sales function. At the beginning of the rollout phase, adoption seemed to be going great: over 20% of the sales function was using it. But then came a massive drop-off, down to around 5% in a matter of weeks. What happened? The sales engineers had quotas to meet, and time spent playing around with a technology that wasn’t working for them was time they weren’t spending doing their jobs. They very quickly lost interest in the copilot and simply stopped using it. Interestingly, the loss of interest tended to be more acute among technical experts, probably because they experienced a smaller productivity uplift and a higher threat of substitution compared with nontechnical workers.

This company took a step back and reconsidered the rollout. Instead of a copilot for everybody, they reset expectations. They admitted the copilot wasn’t a genius but needed further training. They segmented the employee population into those who would be using and experimenting with the copilot, and called them champions. These became the early adopters. Then they started enrolling influencers, people who weren’t necessarily tinkering with the technology, but they had enough social capital to be able to diffuse it or communicate its value to others. And only after the most promising use cases had been solidified, they reengaged the whole organization and worked on the adoption at scale, introducing KPIs for measurement and compliance of adoption.

If you’re a manager who really wants to roll out GenAI productively, then you need to make sure that you carefully orchestrate adoption, not just massively dump it on your employees and put the burden of discovery on employees who may not have the skills or the appetite to engage with it.

Convincing people to switch tasks and occupations

Careful management of the rollout process is especially crucial in cases in which GenAI has the potential of making some jobs obsolete and/or requires people to move across tasks and maybe even occupations.

Changing tasks is not something that happens automatically, especially for workers who have accumulated task-specific human capital that depreciates once they move across different tasks. In an ongoing study, we followed a large bank that automated some routine tasks in branches. This bank had a commitment not to lay off workers. They asked their workers to use the slack time created by automation to do higher value-added tasks that required more social interaction, like loans. Two things happened: productivity increased, thanks to greater automation; yet the workforce didn’t switch to more complex tasks but continued doing the easy tasks as before. Our interpretation is that these employees, especially high-tenured ones, had accumulated specific human capital and weren’t prepared to move to doing other, more complex tasks.

Getting workers on board with the adoption of a new technology is necessary, not just to allow the automation of routine tasks and the transition to higher value-added tasks, but also because internal expertise is needed to understand how to embed the technology in the organization. For example, GenAI could capture tacit knowledge and unleash it across the organization, but this can happen only if the experts in the firm are willing to transmit their knowledge to train GenAI.

It turns out that this process of knowledge transmission is far from trivial and can lead to resistance if employees perceive this process as a way to expropriate their expertise, taking away what made them valuable. To avoid these frictions, managers need to articulate how the transmission of expertise will impact the organization, including the possible substitution of roles. A key point is the contract that will be used to regulate this process. For example, will the employees who trained the AI on their knowledge be able to bring their agents with them when they leave the company? Is there a contractual need for defining ownership?

Managers must take these matters seriously and think deeply about what motivates people, both extrinsically and intrinsically, and how best to incentivize change. If not, frictions will get in the way of adoption, and employees may rationally sabotage implementation.

Beyond changing tasks is changing occupation. Let’s face it: an improvement at the individual task level is not going to be as consequential as multiple improvements embedded within a complete change of process, marking a systemic shift. When this happens, entire occupations, not just individual tasks, will change. And that is an even more costly shift to undertake, because it goes to the heart of another topic I’m studying: occupational identity.

In a recent study focused on a large sample of unemployed people, we find that “identity fit” (meaning the extent to which you see yourself working in a specific occupation) dominates a person’s decision to reskill much more than money alone. Similar dynamics may happen also for workers inside firms who are asked to move to different roles. Managers need to take identity concerns very seriously.

Change is coming: how will you manage it?

We cannot think about GenAI adoption in a deterministic fashion. While an AI future is inevitable, how it unfolds depends on how managers shape the J curve. The adoption of GenAI will look different for firms that are GenAI native versus incumbents: Native GenAI firms will be able to organize themselves around GenAI, rather than change their legacy structures and processes. However, even large incumbents in highly regulated industries, such as banking, have made impressive strides in GenAI adoption, initially introducing simple chat interfaces, then moving to AI agents to support client-facing functions, and gradually even deploying AI orchestrators of agents, all within a very short time period.

Over the next few years, training will be crucial, but it will need to be designed differently from standard human capital investments. If we want to reach a point where the technology actually changes workflows and how work is done, training will need to be part of a broader “change management” program. Managers will need to think much more carefully about how they communicate why their employees need training and how this fits their strategic objectives; how to incentivize leaders to support the initiative; and how to articulate the future for those who choose to train, learn new things and join in the process. In this sense, the HR function will also need to change dramatically, both in terms of technological expertise as well as its connection with corporate strategy.

Part of management is cognitive, understanding what you want to achieve with the technology, which is no small feat itself. But then there is the social part of management, which involves higher-order meta-skills, such as making good use of social capital and the social intelligence that will support the direction you want to go in. It involves coaching and understanding the human side of motivation to encourage people to embrace change and move in new directions.

In my research, I find that middle managers are the linchpin of this process: They have local information about individual workers’ barriers and capabilities, and they have the authority to allocate time, redesign jobs, and make credible commitments about career outcomes. Variation in manager quality can explain a surprisingly large share of the differences we observe in training and technology adoption across firms.

Clearly, there is a tremendous amount of work to do, but also tremendous potential. At this moment, every path of possibility is open, but for GenAI to live up to its positive promise depends on the managerial choices we make. This includes taking control of the narrative around AI, which is currently being shaped by extreme views, either incredibly positive or incredibly negative, about what these technologies can do. The risk of passively accepting these narratives is that they may inadvertently create uncertainty and fear in organizations.

Part of what managers can do for their employees is to help the organization navigate this inevitable uncertainty through their communication and, most importantly, their actions. There’s urgent demand for managers to roll up their sleeves and figure out the best paths forward for their firms. We’re in an interesting moment: Are you ready to get your hands dirty?

SOURCE: Based on the keynote talk, “Management and Reskilling in the Age of GenAI,” delivered by Raffaella Sadun at the Economics of AI Conference 2026, a gathering of experts organized by IESE professors Ricard Gil, David Wehrheim and Sampsa Samila to analyze the impact of artificial intelligence on productivity, organizations and the economy.

This article is included in the annual publication, Insight for Global Leaders No. 2 (2026).

READ ALSO:

GenAI: Easy to use, harder to manage

Will AI upend the offshoring of global knowledge work?

Job maker or job taker? How artificial intelligence is defying the doomsayers