This spring, Meta told thousands of U.S. employees that software on their work computers would begin capturing mouse movements, clicks, and keystrokes. Meta needed this data, they said, to help train its AI agents on real examples of how people work.
And it’s not just Meta that needs this kind of data. Any company trying to automate workflows with artificial intelligence needs access to how humans actually perform these tasks. McKinsey, for instance, created its internal generative AI, Lilli, by drawing on the expertise that had been developed in the firm over its long history.
The director of design for Lilli made the underlying logic explicit. For most of McKinsey’s history, she said, “our knowledge was with our experts.” Lilli is the attempt to take that knowledge out of the experts and make it an organization-wide resource.
From the perspective of the organization, this makes perfect sense. But the benefit is not quite so obvious from the perspective of the expert. In May, a month after news of Meta’s click-tracking broke, the company laid off 8,000 people and gave AI as the reason. More broadly, more than 175,000 people have been laid off in tech in 2026, and the main reason given for these layoffs has been AI’s rapidly improving capabilities.
So let us put it in simple terms. Companies want their employees to give them their expertise to train AI agents. And once that expertise has been adequately transferred, companies will get rid of their people and let AI agents take over.
This is not like asking turkeys to vote for Christmas. This is like asking turkeys to buy the presents, decorate the tree, cook the meal, and put the drunk uncle safely to bed. It is rational for the turkeys to resist. Business leaders need to approach their own goal differently, and doing so begins with understanding how knowledge transfer works.
Knowledge transfer needs cooperation
In a recent article for Harvard Data Science Review, my coauthors and I argued that AI has helped solve the technical challenge of knowledge retrieval. Instead of having to manually look for needles in very disorganized and intimidatingly large haystacks, we can now use large language models to search unstructured data and answer questions in ordinary language.
This is an enormous improvement. But there is a very important limitation that no amount of AI can overcome: Better retrieval only helps with knowledge that made it into the system in the first place. And when it comes to expertise, a lot of knowledge never reaches the system at all.
Now, some expertise is easy to capture—any expertise that has been codified, that has been described, that has been captured in words or in processes is relatively simple to access. So things like templates, playbooks, and workflows can be turned into data straightforwardly and used to train AI agents. Artifacts like these can be taken and incorporated without the expert’s cooperation.
But there are parts of expertise that cannot be harvested purely from the outside. For example, you can record that an experienced salesperson abandoned a lead after the third call, but not that a particular hesitation in the client’s answer reminded her of three deals that later collapsed. You can observe a senior engineer departing from the standard troubleshooting procedure, but not that she recognized a combination of weak signals that, from experience, usually points to a deeper problem. Nor does the record show the alternatives the expert considered and rejected, the exceptions she has learned to recognize, or what she would have done had one detail been different.
Those things can become training data, but only if the expert helps make them visible: by explaining her reasoning, comparing cases, identifying exceptions, and articulating the intuitions experience has taught her. That means the most valuable targets for knowledge transfer cannot simply be extracted. They depend on the expert’s choosing to cooperate.
The limits of asking
Well, then, why not ask your experts? You can, and you should. But it is no panacea. Well before the rise of AI, organizational psychologists analyzed a phenomenon they called “knowledge hiding”; research suggests that roughly half of employees intentionally withhold or disguise what they know from their own colleagues.
Knowledge hiding is prevalent in the context of AI too. In one recent survey of 4,000 workers, 35% said they are hoarding knowledge specifically because they fear being replaced by AI, and 38% are reluctant to train colleagues in areas they consider personal strengths. Relatedly, a three-wave study of 348 knowledge workers working alongside AI found that employee-AI collaboration raises fears of job insecurity, and job insecurity drives knowledge hiding.
Companies cannot coerce their way to a solution to this problem, because compliance can be performative. This is what the research calls “evasive hiding.” Evasive hiding is cooperation-shaped: the employee attends every session and answers every question, but they nevertheless withholds the information that really matters.
Observation is not enough
If asking fails, watching looks like the natural next step, which is precisely the logic behind Meta’s keystroke capture program. But this does not offer a complete solution either.
History gives us an eerie parallel. A century ago, Frederick Taylor brought stopwatches onto machine-shop floors as part of an effort to turn workers’ practical know-how into something management could observe, measure, and standardize. The machinists understood what was at stake. They coordinated around a deliberately restricted pace, discouraged one another from revealing how quickly the work could really be done, and thereby denied management an accurate picture of their productive capacity.
Taylor called this kind of behavior “systematic soldiering” and condemned it as one of industry’s great evils even while acknowledging the reasons behind this behavior: Workers understood that once management discovered that a job could be done faster, this would lead to higher demands on their labor without equivalent increases in pay. The workers’ reaction was a rational attempt to protect the value of their knowledge and effort. The ensuing battles over time studies helped fuel years of labor conflict.
In sum, surveillance is not a clean source of data because the interests of the watcher and those being watched are not aligned. Instead, it changes the work being performed, and can also lead to conflict. And finally, the part of expertise that leaves no outer traces is in any case invisible to surveillance methods, and no amount of watching a worker will surface it.
Helping your experts help you
There is a world of difference between cooperation and compliance. And the fact that true knowledge transfer requires cooperation changes the organizational problem. Rather than asking, “How do we capture more?” organizations need to ask, “Under what conditions will our experts want us to succeed?”
Here are four things business leaders can do to create the conditions that encourage experts to cooperate:
1. Be explicit about the bargain. Tell employees what information is being collected and why. If the plan is to reduce head count, say so; if the plan is to move people into different work, explain what that means in practice. People are much more likely to withhold when they suspect the real purpose of the exercise is being concealed, because the uncertainty provides motivations to conceal a broader range of knowledge.
2. Address job insecurity directly. If employees reasonably think that successful knowledge transfer will lead to their losing their jobs, no amount of “trust building” will fix the incentive problem. Companies need to make credible commitments about what happens to the people whose expertise is transferred.
3. Keep experts attached to the system. Make them stewards who audit, correct, and update the workflows. This also helps with insecurity because experts can see a future role for themselves rather than a one-time knowledge extraction event.
4. Give them a stake in the upside. If employees are being asked to transfer expertise that makes the company less dependent on them, give them a tangible benefit for doing so. That might mean bonuses for documenting and validating expertise or creative solutions like profit-sharing when the system produces measurable savings. If the company captures the upside of the transfer while the employee absorbs most of the risk, there is little rational reason to cooperate.
Success requires partnership
Your experts are not passive sources of training data; they are participants who understand what knowledge transfer may mean for them. And if you want their best judgment, you cannot merely extract it—you have to make cooperation rational. That requires transparency, security, an ongoing role, and a share in the upside.
The companies that treat their experts as a resource to be exploited rather than as partners to collaborate with will collect plenty of data. But they will fail to collect the deep expertise they actually need.
