AI can now capture an expert's hard-won judgment — but can people still become experts once the routine work is automated away?

This explores whether the hard-to-articulate judgment experts carry can survive once AI takes over the routine work, both whether it can be captured in the AI system and whether people can still develop it. The corpus says yes to the first and is much more worried about the second. Preserving what experts know today turns out to be easier than preserving the process that produces experts tomorrow.

Capturing expertise works better than you might expect. In one industrial case study, writing domain rules and design principles into an agent's scaffolding (the structured instructions and tools around the model) let non-experts produce work that experts rated at their own level, with quality up 206%. The gain came from the externalized expertise, not from a bigger model Can codified expertise let non-experts match specialist output?. Agents can also pull reusable routines out of their own past work and build on them Can agents learn reusable sub-task routines from past experience?. Source code can serve as raw material for skills that can be checked, even with no prior agent experience Can source code replace experience as skill raw material?. One project treats a specific person's distilled expertise as versioned files that can be inspected, corrected and rolled back. It also tracks what someone knows separately from how they act, so each can be audited on its own Can person-grounded skills remain auditable without hidden prompt state?.

What gets captured, though, is mostly procedure, not judgment. An analysis of more than 8,000 trials found that skills mostly work by keeping an agent's actions on track (65.7% of cases), not by supplying knowledge it was missing (4.5%). Skills fail when they're applied in the wrong context or followed too rigidly, which are exactly the situations where tacit judgment matters Do skills teach procedures or inject missing facts?. Even written-down knowledge can erode. Repeated summarizing and rewriting can quietly strip out the details that mattered, unless updates are made in small, curated steps Can context playbooks prevent knowledge loss during iteration?.

The deeper problem is a feedback loop that runs in two directions. Automating junior work removes the apprenticeship where people build the judgment needed to check AI output. Meanwhile, when experts label data for AI systems, they turn their own judgment into training data for their replacements, which narrows the chances for anyone to learn it the old way. Researchers call these the 'missing junior loop' and the 'codifier's curse' How does automating junior work destroy verification oversight?. This connects to Wei's 'verifier's rule': AI learns fastest on tasks whose answers are easy to check Does task verifiability determine what AI systems will learn to solve?. Tacit expertise is often the checking ability itself. So the more work AI absorbs, the more we depend on human verifiers, and the fewer new ones get trained.

The takeaway: tacit expertise can be stored in files and scaffolding, but the problem moves from storing knowledge to training new people. The open question in this collection is how organizations deliberately rebuild the apprenticeship that routine work used to provide. Some point to multi-agent setups with independent verification roles Do single agents always hit organizational limits?, but none of the retrieved work says how people learn to fill those roles.