Tacit knowledge
What experts know but can't fully describe. The reason enterprise AI rollouts stall at the 20% that wasn't in the SOP.
Tacit knowledge is what an expert knows but cannot fully articulate. The radiologist who spots the malignancy before being able to explain which features tipped them off. The senior underwriter who declines an application a junior would have approved, citing “something off about the cash-flow pattern” without enumerating what.
Michael Polanyi summarized it in 1958: we can know more than we can tell. The gap between what experts do and what they can describe is the central reason knowledge transfer is hard.
Why it matters now
Tacit knowledge is the friction in every enterprise AI rollout. The visible work (the SOPs, the forms, the system fields) gets automated quickly. The invisible work (the judgment calls, the exceptions, the moments where senior staff override the documented process) does not, because the people who do it can’t put it on paper.
This is the actual frontier of agentic AI in production. The model can already perform the documented 80% of a workflow. The 20% that requires tacit judgment is where it fails, and that 20% is exactly the part the system designers couldn’t write down to begin with.
The misuse
The classic mistake is treating tacit knowledge as a documentation problem: “we just need to write better SOPs.” Decades of attempted knowledge management have shown this doesn’t work. Tacit knowledge can’t be extracted by asking experts to describe their process; the asking activates the wrong part of the brain.
What does work: observing the expert’s actual behavior on real cases — the inputs they look at, the order they look at them in, the cases they pause on, the ones they override. The tooling that makes tacit-layer extraction tractable in 2026 is observation, not interview. The agent learns from watching, not from reading what the expert said they do.