Tribal knowledge: the AI unlock, not the hidden cost

Every bank board is asking how to adopt AI without losing what makes the institution trustworthy.

It is the wrong question. The one that matters is quieter: what happens to the bank when the people who actually know how it works move on?

Not the org chart. Not the policy pack in SharePoint. The working knowledge: why one flagged transaction is held for review while a near-identical one clears in seconds, or why a particular report is always checked twice before a deadline. Ask the person who has done the job for fifteen years and the answer comes back in seconds. Ask the system meant to take it on, and nothing comes back at all.

That silence is the real exposure. And it is a reason to adopt AI, not to fear it. Done right, AI does not replace the people who hold the bank together. It captures what they know while they are still here, and keeps it after they move on.

The knowledge that walks out of the door

Some knowledge is written down: policies, procedures, the manual. The rest lives only in people's heads, the judgement they apply almost without thinking, which never makes it onto paper. Banks are well stocked with the written kind and quietly dependent on the unwritten kind. The unwritten kind is the part that leaves when people do.

Most AI pitches treat it as something to migrate: extract what is in people's heads, load it into a knowledge base, point a model at it. But you cannot interview judgement out of someone in a 45-minute workshop. People hold it the way an experienced driver reads a road, built up over years of doing the job, not from reading a manual.

And that judgement is not the obstacle to AI. It is what decides whether AI is worth anything inside a bank at all. An assistant that does not know why that flagged transaction is held will clear it automatically. Once. Then the regulator calls. This is why so many bank AI programmes stall at the pilot: the technology works and the data is there, but the judgement layer is missing.

Where the knowledge actually lives

It is scattered across the organisation, and almost none of it is written down in a usable form.

• The analyst who knows which requirement always gets read the wrong way, and how to word it so it does not.
• The engineer who knows that one piece of old code must never be touched, though the reason was never written down anywhere.
• The compliance officer who can tell which new regulation actually affects the desk and which is just noise.
• The operator who knows that one counterparty's data always looks wrong on screen and is always fine.

When one of those people leaves, the bank does not lose data. It loses the reasons behind the data. And in a regulated business, the reasons are what make AI safe to use.

Capture it from the work, not from the people

So stop trying to extract this knowledge. Start capturing it from the work.

Every exception reviewed, every override approved, every decision escalated carries a moment of judgement. That moment is the knowledge, and today it vanishes: the case clears, the system logs the outcome, the reasoning is gone.

A hybrid workforce captures it as it happens, not by monitoring the person but by putting an agent in the loop alongside them: working the same queue, observing the decision, recording the why with the person's sign-off, every action on an audit trail. AI here is not a system you switch on. It is a workforce you onboard, under the same controls and audit obligations as your people, so they stay where they add the most value, on the judgement, not the repetition.

The knowledge is not waiting to be uploaded. It is waiting to be witnessed.

Take a live programme where a team works through thousands of requirements, the kind of analysis that normally takes days each cycle. Which requirement is ambiguous, which gap will cause trouble later, which user story is not quite right: that judgement sits in their heads. On one such programme at a tier-1 bank, agents working beside the team turned more than thirty hours of manual analysis into under half an hour, and surfaced gaps the people would have missed. Encode a piece of reasoning once and it can be reused everywhere the same problem appears, turning one person's hard-won judgement into something the whole organisation can draw on.

The knowledge is the moat

This is where the lasting advantage sits. Warren Buffett calls it a moat: in his 2007 letter to Berkshire Hathaway shareholders he wrote that "a truly great business must have an enduring 'moat' that protects excellent returns on invested capital." For a bank adopting AI, that moat is the encoded judgement of its own people. Each approved decision compounds into the next, so the system grows sharper over time, and the knowledge stays inside the bank. It cannot leak, and it cannot be folded into someone else's model.

None of this holds if the AI runs outside the bank's control, calling a model the bank does not manage on data it cannot account for. Scattered AI tools only make it worse, fragmenting knowledge instead of compounding it.

How Promenaut does it

Promenaut is an enterprise agentic workforce platform: it maps, designs and runs a hybrid workforce of people and agents, inside the bank's perimeter and model-agnostic, with full audit and control built in. Its agents work alongside human operators across the software lifecycle and, increasingly, the wider operation: credit, controls, reconciliation, audit. The governing principle is agentic oversight: we govern what agents do at the workflow level, not which model they run on. You cannot download what your people know, but when agents work beside them they learn it from the decisions themselves, while human review keeps judgement exactly where it is required.

The people who hold your operation together will not be in their seats forever. Their judgement does not have to leave when they do.

See what this looks like in your operation.