Decoding the Judgement

I once asked a lending director how he knew a borrower was in trouble before the numbers showed it.

He thought about it for a while. Then he said, "I just know." Thirty years of files, and that was the sentence I wrote down. He was not being evasive. He genuinely could not tell me. He had read a set of management accounts the way you read a friend's face across a room, and something in the gap between the revenue line and the tone of the commentary had told him to look harder. By the time he could have explained it, he had already moved on to checking the thing his instinct flagged.

The official answer to "how do you decide" is always available. It sits in the credit policy, the lending manual, the sanctioning matrix. Every bank I work with has it written down, and the best people in those teams will tell you that the written version sets the framework, and their judgement does the rest.

The documented process is the skeleton. The judgement is the soft tissue around it.

Michael Polanyi described this sixty years ago. "We can know more than we can tell." The chess master who sees the strong move before he calculates it. The doctor who is uneasy about a patient before the chart justifies the unease. The lending director reading the room in a set of accounts. The knowledge is real, it is hard won, and it resists being written down because the person holding it never held it as words in the first place.

You cannot interview that knowledge out of someone. I learned this the slow way.

Where the judgement shows itself

When you take a team's written policy and turn it into the rules an agent will apply, the encoding needs a precision the policy was never written to provide, because a skilled reader never needed it spelled out. A policy like that is written for a skilled reader. It sets the direction and trusts the person applying it to bring the judgement, because that is what their experience is for. A machine has none to bring, so every step has to resolve to something.

So the work of encoding keeps surfacing the places where experienced people were supplying judgement the policy had always, sensibly, left to them. You do not have to hunt for the undocumented judgement. Trying to make the policy explicit enough for a machine surfaces it on its own.

I watched this with a lending operations team. They had a good policy and we were working through how cleanly it became rules the agent could apply. We kept reaching points where the policy set the direction and the person brought the judgement, the way it is meant to work. The written version had never needed to be more exact than that, because the team had been making those calls by judgement for years, the same way each time, without registering that a judgement was being made at all. It was complete. The experience the team brought was part of what complete meant.

That has shaped how I build more than almost anything else I have learned in this work.

Two kinds of knowledge

There are two kinds of knowledge that do not live in the manual, and they need different handling. The first kind is simply undocumented, knowable but never written down. One expert can explain it once you catch them in the act, and every other expert would nod along. That knowledge you capture, you encode, and the workflow is genuinely better for it. The second kind is contested. The bank holds two legitimate answers and resolves the tension case by case, inside people's heads, usually without admitting that is what is happening.

Contested knowledge needs the most care. Bury one of two legitimate answers inside the agent and the choice becomes invisible and permanent, made by no one. So you encode that judgement differently. You teach the agent to spot the contested call, to lay out both legitimate readings and the grounds for each, and to evidence its reasoning in the open. The human stays the ultimate reviewer, weighing an argument the agent has set out in full and signing the call that carries their name.

When coverage is the wrong goal

The first instinct, once you see all of this, is to go hunting for edge cases and encode them one at a time. I had that instinct. It does not survive contact with a real portfolio. The edge cases do not form a list with an end. They form a long tail with no floor. For every released guarantee still sitting in a term sheet, every February year-end hiding inside a December reporting group, every covenant quietly tightened in a footnote, there is another shape waiting in next week's file that nobody has imagined yet. Coverage is a race you cannot win.

So coverage stops being the goal. For the genuinely new shapes, the ones no policy and no expert has met yet, what makes a workflow trustworthy is the accuracy of its own "I am not sure." Eight hundred clean cases and then one badly wrong one, delivered with full confidence, is dangerous in a way that is hard to catch until it has already done harm.

The engineering that builds trust is narrower than that: a workflow that stops cleanly at the edge of what it understands, hands the rest to a human, and is reliable precisely because it knows where it runs out.

When the agent abstains well, the human inherits a smaller and sharper pile of work. What that concentration does to the shape of someone's working day is its own story, and not this one.

The real work

The people whose knowledge is most worth capturing are often the most reluctant to give it, and the reluctance is rarely stubbornness. If your judgement is the thing that makes you valuable, being asked to spell it out so that a system can hold it is a strange request to sit with, and some part of the room understands that even when no one says it aloud. The reframe that actually lands is about where that judgement gets spent. It is too expensive to keep spending on the routine eighty percent. Capturing it frees it for the contested twenty, where the bank truly needs a person in the room.

The familiar version of this argument says "the model is the commodity and the knowledge is the moat." A year of the work has taught me something narrower and more useful underneath it.

The knowledge you can write down was never the hard part. The hard part is the knowledge your best people cannot quite reach, the calls they make without noticing they are making them, the judgement a good policy leaves to them.

A machine cannot learn what the experts themselves have never put into words. So the real work, the work that takes the months and earns the trust, is sitting with them while they say it for the first time.