What Is Tacit Knowledge?
Ask the best person on your team how they make their hardest call. The one where the data says one thing, your best person chooses a different path, and it turns out to be the right one.
You'll get the textbook answer. Press harder and you'll get "gut feel", "I just knew", "something was off". Twenty years of judgment, and the best explanation it can produce is a shrug.
In 1966, the philosopher Michael Polanyi gave that shrug a name. "We can know more than we can tell," he wrote. The part we can tell, he called explicit knowledge. The part we can't, he called tacit knowledge.
A thousand chicks an hour
The cleanest demonstration of tacit knowledge at work is a chicken hatchery.
In the 1930s, Japanese hatcheries solved a problem the rest of the world couldn't: telling male from female day-old chicks. The difference matters - hens lay eggs, roosters eat feed - but day-old chicks look identical.
Graduates of the Zen-Nippon school, set up to train chick sorters, could sort a thousand chicks an hour at roughly 98 percent accuracy. Here is the strange part: ask a master how she told a male from a female and she couldn't tell you. She had learned where the line was, but she could not say where she had put it. The skill was real and bankable, and the person who had it could not put it into words.
So how do you teach a skill nobody can put into words? The school's method was brutally simple. The student picks up a chick and guesses. The master says yes or no. Next chick. Thousands of guesses, thousands of corrections. With enough reps, students sorted nearly as well as their masters - and could explain it just as poorly.
No theory. No lectures. Trial and error, with an expert closing the loop. Reps, not reading.
Your organisation runs on it
Tacit knowledge sounds like a curiosity until you notice where it lives in your own organisation. Three places to look:
The gap between your average performer and your best one. They sat the same onboarding and read the same wiki. The explicit knowledge is identical. The difference is judgment, built through years of trial and error, and it appears on no balance sheet.
The door. When your thirty-year engineer retires, her successor inherits her files, her dashboards, and her documentation - everything she could write down. Which is to say: everything except the thing you actually paid her for.
Your AI strategy. A frontier model is trained on the indexed internet: books, code, Wikipedia. That is humanity's explicit knowledge, and every one of your competitors can rent the same copy. Knowledge was the first AI bottleneck, and it is solved. The bottleneck now is judgment, and your organisation's judgment is the one training corpus nobody else can buy.
Why you can't just ask
The obvious move is to interview your experts and write it all down. Organisations have been running knowledge-capture programmes for decades, and they keep producing the same artifact: a wiki nobody's best work ever passed through. The reason is the same reason the chick sorters had no manual. Judgment is reactive, not declarative. Ask an expert what she looks for and you get the textbook answer, or the highlight reel. Put her in front of a live case and something different shows up - the thing she actually does.
Your records won't save you either. The better your experts, the less evidence they leave. The nurse's 3am call shows up as stable vitals. The coach pulling a player at minute 78 shows up as a clean season. Your best people spend their careers preventing the events that would have been your training data. Absence is not training signal.
The only method that has ever worked
The chick-sorting school stumbled onto the only reliable way tacit knowledge moves from one head to another: put the learner in the live situation, make them decide, correct them, repeat. Learn through experience. Reps, not reading.
The same recipe applies to AI, for the same reason. A frontier model has already done all the reading there is to do. What it has never done is the reps - your reps, on your hard calls, against your standard of correct.
This is what Tacit builds. We construct simulated scenarios where the hard call is live: the claim is marginal, the patient is deteriorating, the customer is about to walk. Your expert acts, and we capture two things while they are still hot: what she did, and why she did it over the three other moves she could have made. One scenario is a sample. A hundred scenarios across a hundred shapes of trouble is a training corpus, generated from the one place it existed all along - inside your experts' heads, under pressure, in the moment the decision was theirs to make.
Then an AI learns from it the way your experts learned in the first place: through experience. Trial and error, with an expert closing the loop. Rep after rep, until the judgment transfers.
Your best people didn't get their judgment from a manual. Your AI won't either.
If you want to see what your organisation's judgment looks like as training signal, and bring your hardest call. We'll build the scenario.