For as long as companies have existed, there has been one thing they could never fully own. They could buy equipment, acquire buildings and patent inventions, but never lay claim to the knowledge in their employees’ heads. Training appeared as an expense. Expertise lived in people. And every evening, that knowledge walked out the door, sometimes to never return.
What firms have always needed but never owned is judgment. Skilled workers carry not just functional knowledge but the institutional instincts, shortcuts and problem-solving techniques for how to get things done. A new generation of artificial intelligence systems born in Silicon Valley are now trying to capture exactly that.
The debate in the room, the memory that a similar fix caused a cascade failure two years ago, the instinct that the framing of a problem is wrong before anyone can explain why — none of this appears in mouse movement logs.
The Appropriability Problem
The economist Gary Becker, who won the Nobel Prize for his work on human capital in 1992, put it plainly. Firms pay for training, but workers own the knowledge it produces.
Skills are portable. When an employee leaves, their experience and intuition travels with them. This portability has always limited how much value a firm can capture from its investment in people.
Becker drew a distinction that still shapes how economists think about this problem. He separated general skills, which raise a worker’s productivity at any firm, from specific skills, which matter only inside the firm where they were learned. In a competitive labor market, he argued that firms will not pay for general training because workers can take those capabilities to another employer that offers higher wages. Workers therefore bear the cost of general training themselves, usually through lower wages during the learning period.
But the most valuable expertise, such as broad judgment, adaptive problem solving and institutional intuition are types of general skills firms cannot own. What current AI workflow capture systems appear to be attempting is something Becker didn’t consider. They aim to convert general human capital into firm-specific encoded assets by capturing transferable expertise before it walks out the door. Whether that conversion succeeds is a separate question.
In 2024, U.S. organizations spent an average of $1,254 per employee on direct learning and a total of $102.8 billion nationwide, according to the Association for Talent Development. These figures include only direct costs such as materials and instructor time. They do not include the larger indirect costs of lost productivity during onboarding or the long period before a new hire reaches full effectiveness.
When a trained worker leaves, the firm absorbs those costs again. The Society for Human Resource Management estimates that replacing an employee costs between 50 and 200 percent of their annual salary, depending on seniority. Gallup places the cost of replacing a manager or senior leader at roughly 200 percent of salary, with technical roles about 80 percent. For a software engineer earning $150,000, a single departure can cost the firm between $120,000 to $300,000, even before considering the institutional knowledge that cannot be reconstructed at any price. The table below shows how this plays out across different types of companies.
| Firm Size / Sector | Direct Training Cost (per employee/year) | Avg. Hours of Training (per year) | Estimated Replacement Cost on Exit |
|---|---|---|---|
| Small firm (100–999 employees) | $1,047–$1,091 | ~18 hrs | 40–80% of annual salary |
| Midsize firm (1,000–9,999) | $739–$782 | ~14 hrs | 80–150% of annual salary |
| Large firm (10,000+) | $398–$468 | ~12 hrs | 50–100% of annual salary |
| Technology / Advanced Software | $1,200–$2,000 | ~14 hrs | 80–200% of annual salary |
Sources: ATD State of the Industry 2025; Training Magazine 2025 Industry Report; SHRM; Gallup.
What Can Machines Know?
Meta has committed between $125 and $145 billion to artificial intelligence infrastructure, including servers, data centers and the compute capacity needed to train and operate large language models. The company carried out a structural workforce reduction in May 2026, cutting about 8,000 roles and permanently freezing another 6,000.
These cuts fell on mid-level program managers, standard software engineers, and customer support roles – the exact categories where AI systems are being trained to operate autonomously.
Whether these systems can truly reproduce the judgment they observe, or merely imitate its surface patterns, remains an open question.
The mechanism Meta is pursuing is simple in concept, although its effectiveness is not yet established. Systems like Meta’s Model Capability Initiative run continuously in the background, recording how skilled workers perform their jobs. They capture mouse movements, keystrokes, application sequences and intermittent screenshots in an attempt to extract knowledge. The mechanics of this capture system were first described in detail in reporting by Reuters on April 21, 2026., which obtained internal Meta documentation outlining the program’s data-collection design.
When an engineer resolves a complex bug or a manager works through a difficult decision, the system records the entire problem-solving pathway. That pathway is then labeled, fed into the company’s models, and integrated into automated workflows. Once encoded, the worker’s tacit knowledge becomes a permanent part of the firm’s intellectual capital. It can be replicated at scale, deployed instantly, and used indefinitely at almost no additional cost.
Meta’s leadership has been explicit regarding both the operational mechanism and the ultimate objective of this strategy. “If we’re building agents to help people complete everyday tasks using computers, our models need real examples of how people actually use them—things like mouse movements, clicking buttons, and navigating dropdown menus,” Meta spokesperson Andy Stone said in a statement to Reuters. CTO Andrew Bosworth described a future where AI agents “primarily do the work” while human capital shifts toward roles that “direct,’’ review and help them improve, according to Reuters.
The Debate in the Room
What these systems capture, and what they miss, matters. The individual keystroke sequence of an engineer solving a familiar class of problem is not the same as a team working through an novel issue. The debate in the room, the memory that a similar fix caused a cascade failure two years ago, the instinct that the framing of a problem is wrong before anyone can explain why — none of this appears in mouse movement logs. Workflow telemetry captures procedure but not the social and contextual dimension of expert judgment, which is where the hardest and most valuable problems are actually solved.
As a result, Becker’s general skills, the adaptive and portable creative capacity that makes a worker genuinely hard to replace, remain for now beyond what these systems can reach.
Artificial intelligence is attempting to close the gap between skills employees own and what firms own. Firms are attempting to convert at least the procedural layer of human capital into firm-owned fixed capital. It isn’t clear if they can capture the problem-solving intuition and collaboration that makes a skilled people difficult to replace.
