Agent-Ready Policies

In a keynote, I said that every policy or operating instruction used by an AI agent should have an agent-ready counterpart.

It looked quite straightforward on the slide. Afterwards over nibblies often we returned to the more difficult question: what exactly would such a counterpart contain?

Take a sentence that asks an employee to obtain ‘appropriate approval’ before processing an unusual payment. A person may know who that means, which exceptions are accepted and when Compliance needs to become involved. For an agent, several decisions are hidden inside two ordinary words.

Putting the manual into a retrieval system does not resolve them. The agent may find the right paragraph and still have no reliable way to determine whether it is current, whether another rule takes precedence or where it should stop.

The counterpart I have in mind is therefore closer to an operating instruction. It would describe what starts the task, which information may be used, what the agent is allowed to do, what evidence it must retain and where a human decision is required. The original policy would remain authoritative because it carries the purpose, context and room for professional judgement.

The annoying part is that we would now have two documents to maintain. If they drift apart, people may follow the current policy while the agent continues to execute last quarter’s interpretation. I would resist treating the agent version as a separate manual. It needs the same owner, a visible link to the source and a review whenever that source changes.

Some ambiguity should also remain with people. A policy may be vague because it is poorly written, but it may also be asking for judgement that cannot sensibly be reduced to a rule. Making an instruction agent-ready includes saying where that boundary lies.

I would start with one recurring process rather than converting an entire policy library. Use its normal cases, known exceptions and failure points, then see where the agent hesitates, escalates or behaves differently from an experienced colleague.

My suspicion is that this exercise will expose assumptions that human teams have carried informally for years: who ‘appropriate’ means, which evidence is enough, and when a rule gives way to judgement. Those are useful things to make visible, whether or not the agent eventually performs the task.

First published on LinkedIn on 18 September 2026.