Agents need somewhere to write things down

The demonstrations are all impressive and they are all the same shape. A question is asked in plain language, an agent queries the warehouse, reconciles two sources, drafts the commentary and posts it into the review pack. Eleven seconds. The room is convinced.

Then it goes into a finance function and the second month is worse than the first.

What goes wrong is not the model

In the first month the agent classifies a set of marketing recharges as operating cost. A controller disagrees, corrects it in review, and explains in a comment thread why this particular recharge sits below the line for this entity. Good. The number is right.

In the second month the same recharge arrives and the agent classifies it exactly as it did before, because nothing about that correction was written anywhere the agent can reach. The controller corrects it again, slightly differently, because she is now working from memory of her own reasoning rather than the reasoning itself.

By month four there are three inconsistent treatments in the audit trail and a controller who has decided the agent creates work rather than removing it. She is right, and the model was never the problem. The problem is that the organisation has nowhere to put a decision that both a person and a machine can read.

Retrieval is not memory

The usual response is to point the agent at more documents. More context, better retrieval, a bigger index. This helps with facts and does nothing for decisions, because the two behave differently.

A fact is true until the underlying reality changes. A decision is a choice one person made, at a point in time, with authority, for a reason, and it stays in force until somebody with equal authority changes it. A decision therefore needs four things attached to it that a document chunk does not have: who decided, when, why, and what it applies to. Without those, an agent cannot tell a live ruling from an abandoned one, and it will average them.

Most organisations have their decisions distributed across meeting notes, mail threads, a comment on a journal, and the heads of four people. None of that survives retrieval in a usable form.

What we build instead

Every agentic workflow we put into production has a decision record behind it, and the record is written as a side effect of the work rather than in a separate documentation exercise. When the controller corrects the classification, the correction is captured with her name against it, the rule it establishes, the entities it covers and the date it takes effect. The agent reads that record before it classifies anything.

The interesting result is not that the agent gets better, although it does. It is that the record becomes the most valuable object in the system. It answers the audit question, it onboards the next analyst, and it survives the agent entirely. If the model is replaced next year, the reasoning is still there.

The test before you deploy

Ask one question of any agentic workflow before it goes live. When a person overrules it, where does the overrule go? If the answer is a chat log, a ticket or a person’s memory, the workflow will decay, and it will decay fastest in exactly the areas where human judgement mattered most.