Training AI models for workplace use turns out to require something companies cannot purchase or download: tacit knowledge, the working expertise locked inside employees' heads that never makes it into a manual or a database. That gap is becoming clear to companies building internal AI tools. The people who hold that knowledge are, for a specific and probably narrow window, in a stronger position than they may realize.
What it means for knowledge to be locked inside someone
There are two kinds of workplace knowledge. The first kind lives in written policies, databases, and documented processes. The second kind lives in the people who do the work, covering the judgment calls that never made it into the manual, the shortcuts that actually work, and the context that makes an official answer wrong for a specific situation.
AI models trained to assist or replace tasks need the second kind. Without it, a model learns the documented version of a job. The documented version is often incomplete, sometimes outdated, and regularly at odds with what actually happens day to day.
Why AI training inverts the usual power relationship
Previous workplace technologies generally required workers to learn new tools. The company owned the tool; the worker adapted. AI training inverts that relationship at a specific moment. The company needs the worker to teach the tool. That is a different kind of dependency, and it runs in a different direction.
The window may be narrow. Companies that successfully extract and encode employee knowledge into a working model reduce their dependence on those workers going forward. Workers who understand that sequence have reason to think carefully about how much they share, and when.
The opening that currently exists
A worker whose expertise is needed to train an AI is in a different position than a worker whose expertise is simply used to do a job. Once training is complete, the company's dependence on that worker may fall. Before training, it may be at its highest point.
That is the window the source describes as a potential power play. The knowledge is currently locked in, and companies have not yet found a reliable way to extract it at scale. That is a structural fact, not a projection.