What actually made an AI agent commit to a decision?

An AI agent is asked to review a set of records against an organisation’s rules. It returns “needs review” on nearly all of them, which is no help to anyone.

The obvious next step is to give it more of the governance context: the rules, the sign-off limits, the policy document itself.

We tested whether that works, five rounds at a time, on public UK procurement records. Adding rules did almost nothing. Something else moved its AI decision-making, in a single jump.

Five rounds, one AI, one more layer of context each time

We took the same 283 UK government purchase records from the first experiment and put the same AI in front of them five times.

Each round added one more layer of the governance context an organisation might give an AI agent. Round one was the bare records, with no rules at all. Round two added a one-paragraph description of how the buyer is supposed to operate. Round three named the actual rules — publication windows, sign-off limits, the usual checks. Round four added something different: past decisions on similar records, with the verdict each one received. Round five added the full policy text on top of everything else.

We wrote down what we expected before running any of it: as the AI got more context, it should get more decisive — fewer “needs review” verdicts, more straight answers. Every decision the AI made was bound into a signed record at the moment it made it, so the whole experiment can be replayed later.

Three rounds of added rules changed almost nothing

Through rounds one to three, the AI hedged on 97-100% of records. It returned “needs review” on nearly everything, whatever rules we showed it. More context, in the form we expected to matter, did almost nothing.

Past decisions broke the hedging in a single round

Round four was past decisions on similar records, and the AI committed at once. It denied 107 records, more than a third of the corpus — not gradually, not partially.

Round five added the full policy text on top, and the AI eased about half of those denials back to “needs review”. The shape is not a steady climb from cautious to confident. It is a single sharp jump in one round, partially reversed in the next.

The AI did not notice an inverted rule

We then ran a sanity check. We inverted one of the rules in the policy text — “must publish within thirty days” became its mirror image — and gave that version to the AI to see whether it would notice the swap.

It did not. It kept reasoning as if the original rule were still there, citing what the rule was supposed to mean rather than what the page actually said. Its confidence did not budge.

What this does not establish

This is one model reading one corpus under one policy: 283 public UK procurement records. It is a research preview with stated limitations and it has not been peer-reviewed.

It shows how this agent’s verdicts moved under controlled changes to its context. It does not establish that any of those verdicts were correct, or how a different model would behave.

What it means for an AI review workflow

Two findings land at once here.

The first is practical. If an AI agent is hedging too much, more rules will not fix it. What moved this agent was being shown real past decisions on similar cases. Anyone designing an AI review workflow has to be deliberate about which examples the agent sees, because in this experiment those examples did more of the work than the rules.

The second is harder. The inversion test showed the AI’s reasoning anchored to what it thought the rule meant rather than to what the page said. That failure mode is invisible in normal operation, because the rule and the agent’s understanding usually match. It shows up only when the agent is put in front of an inverted rule and watched.

A regulator, a board or an auditor reviewing AI-augmented decisions needs both: the past examples the agent was shown at the moment of decision, and a way to check that the agent was reading what it was actually given rather than what it expected to see.

What comes next

This experiment opened three questions we could not settle in the same run. Was it the past decisions doing the work, or the combined weight of all the layers stacked together? Was the policy text’s pullback driven by the rules themselves, or by the cautionary instruction included with them? And was the failure to notice the inverted rule a quirk of one model, or something that holds across different AIs?

The third experiment in the series was designed to disambiguate each of those, with another model run alongside the original to test how much of this was the AI we chose. That paper is now live, and the next post covers what it found.

The full paper for this experiment is below — the five rounds, the inversion check, every signed record, and the verification flow. A worked example of one of those records is on the receipts page.