If an employer can't be sure a trainee will stay long enough to repay training, can AI tools pass expertise along instead?

This explores how hard-to-write-down know-how gets passed on when the economics of training break down, meaning an employer can't be sure a trainee will stay long enough to repay the investment. The collection doesn't hold the labor-economics papers on apprenticeships and incomplete contracts, so it can't answer the financing side directly. What it does have is a set of notes on a nearby question: if training people is hard to pay for, can tacit knowledge travel some other way? The answer they point to is that AI is starting to offer a second route. Expertise gets pulled out of people and built into tools, so less of it has to pass from one person to another.

The clearest case is an industrial study where domain rules and design principles were built into an LLM agent's scaffolding, the structured instructions and checks wrapped around the model. Non-experts using it reached expert-level ratings and a 206% improvement in output quality. The gain came from writing down what specialists knew, not from a bigger model Can codified expertise let non-experts match specialist output?. In economic terms, this sidesteps the training problem: the firm pays once to write the knowledge down instead of paying again for every trainee who might leave. A study of 8,135 trials suggests what this codified knowledge actually does. Skills mostly work as procedural anchors that keep action on track (65.7% of cases), and only rarely by supplying missing facts (4.5%). They also fail when they're retrieved at the wrong moment or followed too rigidly Do skills teach procedures or inject missing facts?. That's close to what apprentices learn: a feel for sequence and judgment, not a list of facts.

The hidden cost is that the apprenticeship relationship carried something along with the knowledge. One note argues AI returns knowledge to a 'flow' economy, like oral and gift cultures before print, but without the embodied carrier, the person whose presence and reputation anchored what was passed on Is AI returning knowledge to flow-based economies?. Notes on machine learning make a related point: teaching only works when there's an information gap, meaning the teacher knows something the student doesn't Why does teacher-student information asymmetry enable learning signals?. Even then, guidance from a teacher mostly steers students toward abilities they already had rather than raising their ceiling Does on-policy distillation actually expand student capability?. If that holds for people too, codified expertise may let novices perform like experts without ever growing into experts. That could leave the next generation of specialists thinner.

A last twist: whether models themselves hold tacit knowledge is still open. There's preliminary evidence from model-editing studies that they do, but it rests on a single case that hasn't been cleanly replicated Do language models possess tacit knowledge in Davies' sense?. Externalized skill libraries let agents keep adding abilities without forgetting old ones Can agents learn new skills without forgetting old ones?, which hints that organizations could build up know-how the same way. The question the collection leaves open, and that you might not have thought to ask, is this: if firms stop paying to train people because the scaffolding does the job, who will hold the expertise needed to write the next version of the scaffolding?