The hidden 20%: your best person's judgment is your most expensive undocumented asset
Kung fu masters were said to hold back the last of their craft from their disciples. Teams do the same thing when AI arrives. What that costs a small business, and a simple way to capture the judgment that never makes it into the manual.
Your best person can tell from one message whether an enquiry is serious. Ask them how, and they'll say "I just know". The philosopher Michael Polanyi named this in 1966: "We can know more than we can tell." That unexplained judgment is the part of your business no manual holds, and the part no AI can learn until someone writes it down.
There's an old saying about Chinese kung fu masters: a master never taught a disciple everything. He held back the last few moves, for fear that a student who knew it all might one day betray him. Some people at work do the same thing when AI walks in. If the machine learns my judgment, what's left for me? So they hand over the easy 80% and keep the last 20% to themselves, and the business ends up with an AI that learned nothing from the person who knows the most.
Tacit knowledge, and why AI can't read it
Polanyi called it tacit knowledge, in The Tacit Dimension (1966). It's the expertise people use every day but struggle to explain, like recognising a face or knowing when a deal is going cold. The economist David Autor calls it Polanyi's paradox and uses it to explain why some work has been so hard to automate.
Most small businesses run on it. It's the senior technician who knows which aircon faults are worth a second visit. It's the admin who knows which customers need a call rather than an email. It's the owner who can tell in one message whether an enquiry is serious.
Ask these experts how they reached a judgment and you'll often hear, "I just know." But whether they realise it or not, your experts often already run a system. The intuition and the SOPs in their heads are really flowcharts nobody has drawn.
A practical place to start with Polanyi's idea is the approval decisions, not the know-how. When is a quote good enough to send? When is a new hire ready to handle a client alone? Which complaint needs the owner, and which can the team handle? That judgment lives in context and history, not in any manual.
What it costs when it stays hidden
One client put it at 10 to 20 hours a week for their team. That's their estimate, not a measurement. In their description, it isn't salary. It's time spent doing the wrong thing, reading notes in the wrong places, and fixing mistakes that come back.
When the person who knows is in the room, they can catch the problem in five minutes. When they're on leave, busy or gone, the work can go out wrong and someone loses an afternoon undoing it.
I heard the symptom from two independent insurance advisors in a workshop in September 2026, in one sentence: "I'm not sure where it is." They had plenty of knowledge. It was scattered across notes, drives and their own heads, and nobody could retrieve it when it mattered. What happened in that workshop.
An AI model out of the box knows a lot about the world and nothing about one company's narrow workflow: your customers, your pricing exceptions, the jobs you'd never take. Anyone can buy the same model. Nobody else has your best person's judgment.
How I capture my own hidden 20%
In the systems below, the AI was the easy part. The real work was asking the right questions to pull the judgment out of a person, which running my own businesses with AI required.
A receipt catalogue. My content agents can only make claims that trace back to a catalogue of facts I've verified, with sources. It grew from under a hundred entries to more than 170. It also records the judgment around them. Its rule is "Would Tony publish this?", not "Is this true?" It names three ways a true fact still gets misused: attributed to the wrong project, stretched beyond its scope, or applied to the wrong subject. That's tacit knowledge written down, and it gives the agents a written rule to check against instead of a guess.
A voice guide. A plain list of words I use and words I never use, so drafts sound like me without my rewriting them.
The same habits I used with people. At Mindreader we grew to a team of ten across Asia. The questions I asked then are the ones I settle before any AI touches a decision now: who owns this decision, who can tell the AI it's wrong, who has the final vote, and what does "good" look like?
A simple way to start: the approval inventory
You don't need a knowledge-management project. Start with the decisions that currently wait for one person:
- List the approvals. For one week, note every time someone asks, "Can you check this?" or "Is this okay to send?"
- Write the rule behind each one. Ask the person who approves: "What would make you say no?" Their answer is the hidden 20%.
- Record three real examples. One yes, one no, and one borderline case, with the reason for each.
- Hand it to someone else, a new hire, a colleague or an AI, and see where they get stuck. The places they stall are the rules still missing.
This complements a piece I wrote about writing down one process. That one captures the steps. This one captures the judgment inside them.
Why your team should want this
The fear is the kung fu master's fear, and it's understandable. Judgment takes years to build, and handing it to a tool can feel like volunteering to become obsolete.
But holding it back rarely protects anyone. It can leave the business depending on one person never taking leave, never getting sick and never leaving. And it keeps that person stuck answering the same questions forever instead of doing the work only they can do. Captured judgment can be how your best person gets promoted rather than replaced.