Rule or agent: a decision test for where AI actually belongs

There is a strong pull, right now, toward putting a language model in front of everything. It is understandable. The capability is genuinely remarkable, budgets have appeared, and "we are using AI for that" is a satisfying sentence to say in a board meeting.

It is also how organizations end up with systems that are slower, more expensive and less reliable than what they replaced.

What a rule gives you that a model does not

Deterministic automation has four properties that are easy to undervalue until you lose them.

It is repeatable. The same input produces the same output, today and in eight months. You can reason about its behaviour without running it.

It is auditable. When something goes wrong you can trace exactly why. This matters enormously in any regulated context, and it matters more than people expect everywhere else — usually the first time a customer disputes what happened.

It is cheap. No per-token cost on work that never required thought. At volume this stops being a rounding error.

It is safe. A rule cannot improvise. It will not get creative about a refund policy at three in the morning.

The test

Before any step in a process is assigned, we ask one question:

Could a competent new employee execute this correctly from a written instruction, without judgement?

If yes, it is a rule. Write the rule. Do not put a model in the path.

If the honest answer is that they would need to read the situation — weigh things, pick between reasonable options, interpret something ambiguous, decide what matters most — then reasoning is genuinely required, and an agent is the right tool.

And if the answer is that they would need to be trusted, or be accountable, or be able to say something is not right, that is the third category, and it belongs to a person.

Worked examples

"Send a reminder the day before an appointment." Rule. There is nothing to interpret. A model here adds latency and cost and introduces the small but non-zero possibility of a reminder that says something unexpected.

"Decide which of last night's forty enquiries the sales team should call first." Agent. This requires weighing signals against each other, and the ordering genuinely changes depending on context.

"Route the enquiry to the right team." It depends, and this is the interesting case. If routing is determined by a field the customer already selected, it is a rule. If it depends on understanding a free-text description of a problem, it is an agent. The same business requirement lands in different categories depending on what information is available at the moment the decision has to be made.

"Tell a customer their claim has been declined." Person. A model could generate the words. That is not what the moment requires.

The hybrid pattern

The best-performing designs we build are rarely purely one thing. The common shape is an agent doing the interpretation, then handing a structured result to deterministic automation for execution.

The agent reads the enquiry and produces a classification, an urgency, and a recommended route. Everything after that — the assignment, the notification, the record update, the follow-up scheduling — is rules operating on that structured output.

This gets you reasoning where reasoning is needed and determinism everywhere else. It also means that when something goes wrong, you can tell immediately whether the interpretation was wrong or the execution was, which is worth more than it sounds.

Why this is a commercial question

Every step you assign to a model instead of a rule adds recurring cost, latency, and a small amount of unpredictability. Individually, none of it matters. Across a full operation running continuously, it compounds into a system that is measurably worse than the one a more disciplined design would have produced — and considerably more expensive to run.

The discipline is not anti-AI. It is what makes the AI you do deploy worth defending.