Your AI Remembers the Decision. Does It Know When to Forget It?

Imagine a sales director approving a discount to rescue a relationship after a disastrous delivery. Six months later, an AI agent finds the approval in the company archive and offers the same discount to a perfectly happy customer.

It has found a real decision, made by the right person. It can produce the date, the email and a beautifully concise explanation.

It has also mistaken an apology for a pricing strategy.

This is a hypothetical example. But as agents move from drafting answers to making decisions, it captures a problem worth taking seriously. An organisation needs to remember why it did something, and recognise when that reason no longer applies.

Otherwise, we could give corporate folklore its own execution engine.

In an essay published on 29 October, Spawn founder Jacob Sansbury describes using recorded conversations to build AI judges. These apply principles drawn from previous human decisions, refer unsuitable questions back, and hold unfamiliar judgements for people. Corrections become part of the record the judges consult.

He reports more than 11,000 verdicts in an eight-day window, with 1,863 held for a human. These are the founder’s figures, rather than independently audited results.

The interesting development is the machinery: recorded judgement becomes something other agents can act upon.

The AI Daily Brief covered the account on 30 October. Simon Willison’s response added a useful boundary: having tried the game, he found the experiment compelling but questioned how readily its development approach would transfer to software where failures carry greater consequences.

That distinction matters. A promising operating model in one company does not arrive with universal planning permission.

Still, the underlying idea has obvious appeal. Anyone who has answered the same operational question repeatedly can see the attraction of answering it once, properly, and making the answer useful elsewhere.

The question is what “properly” needs to include.

Consider that fictional discount again.

The approval depended on a delivery failure, a particular relationship and an attempt to repair it. Strip those conditions away and the resulting instruction changes meaning.

A searchable archive can establish that something was said. It cannot, merely by retrieving the sentence, establish that the sentence still governs the present case.

My concern is that reliable recall could make this failure unusually persuasive. The agent would not need to invent anything. It could cite the wrong precedent impeccably.

The same problem could affect product design. A company rejects a feature because its infrastructure cannot support it. Later, the infrastructure improves. An agent trained to respect previous decisions keeps rejecting the feature.

Nobody has breached policy. The opportunity simply never reaches the meeting.

That is how a capable system could preserve yesterday’s constraints more efficiently than yesterday’s staff. “We have always done it this way” finally gets a performance upgrade.

This is an inference about a risk, not a finding that Spawn’s judges behave this way. Indeed, its described correction mechanism addresses part of the problem. The harder question is how a business notices that a previously sensible answer now deserves correction.

The scale of agent work makes this more than a filing problem.

In its September disclosure, Anthropic reported approximately 30,000 agents doing research and engineering work simultaneously on its most-used internal platform, as of August 2026. Its own index classified 26% of AI R&D work as AI-led under human supervision, while finding no measured category fully autonomous.

Those are company-produced measurements, with methodological limits. They describe substantial delegation alongside continuing human direction; they do not prove that management has been automated away.

My expectation is that expanding delegation will force organisations to make more of their judgement explicit. A manager may understand which exceptions are temporary and which decisions reflect settled intent. An agent needs a dependable way to distinguish them.

For a small business, I would start with one recurring decision: refunds, delivery promises or routine purchasing.

Alongside the answer, record the conditions that make it valid, who can revise it and what change should trigger reconsideration. A temporary concession needs an expiry. A decision based on a capacity limit needs a review when capacity changes.

This need not become an encyclopaedia of everything the managing director has ever muttered. Some remarks deserve the mercy of remaining lunch.

Humans already apply stale policies. They misremember exceptions, avoid awkward conversations and tell new colleagues that an arbitrary restriction is apparently mandatory.

A well-designed agent could be better. It could retrieve the original reasoning, notice contradictions and compare present conditions with the assumptions behind a decision. A clear record could make authority easier to challenge.

I think that is the strongest case for building these systems.

But the benefit depends on what the organisation rewards. If every challenge counts as delay, and every resolved request counts as success, a system may be encouraged to find an applicable answer rather than discover that the available answers need changing.

The remedy is not to have a person approve everything. That would recreate the queue with more expensive stationery.

Give agents room to act within clear boundaries. Also give them a recognised route to question those boundaries. Resolving a routine case and proposing a policy revision are different jobs, requiring different authority.

Then examine some decisions that passed without escalation. A business that reviews only difficult cases could miss an obsolete rule being applied with complete confidence.

A useful pilot would test a policy change as deliberately as it tests the original policy.

Return to the hypothetical discount. Give the agent the original approval, including the delivery failure. Then present a new customer with no such history. Can it explain why the precedent does not apply?

Next, change an actual policy in the test environment. Check whether all relevant agents use the revision, whether work already under way needs reconsideration, and whether old advice remains marked as historical.

Measure how quickly the organisation adopts the changed decision. Count cases where the old rule persists. Look for people having to correct the same assumption repeatedly.

These are tests of organisational adaptability. They matter even if the model’s answers are fluent and its memory is excellent.

There is a real opportunity here: a business can preserve hard-won judgement and make it available without putting every question through one exhausted person.

The advantage will depend on how well that judgement can evolve.

Before expanding an agent’s authority, show that it can apply a decision, explain its limits and stop applying it when the authorised rule changes. Build a way to challenge precedent alongside the machinery that retrieves it.

A company should be able to remember its founder without being governed forever by something the founder said on a difficult Tuesday.