Rules first, models second

Every operator I meet wants AI. Very few want an AI that might, once in a thousand calls, invent a discount, misread a tax rule or approve a refund it should not have. In a business that settles real money, a 99.9 per cent correct model is not good enough for the parts that touch money.

Split the problem in two

Most operational decisions are not creative. Is this bill settled? Does this product map to this HS code? Is this staff member available for this shift? These have correct answers, and correct answers should come from rules: deterministic decision trees that behave the same way every time and can be audited line by line.

Language models are extraordinary at the parts that are genuinely fuzzy — understanding a guest’s request, drafting a reply, summarising a call, pulling structure out of a messy document.

Let the rules decide. Let the model explain, translate and listen.

In practice

In our systems an agent may read a request and propose an action, but the action executes only if it passes the same rule checks a human operator would face. The model never has a direct path to money. It works inside lines that rules have already drawn.

This is close to what the NIST AI Risk Management Framework describes in its Govern, Map, Measure and Manage functions: know where the risk lives, and design the system so the risky parts are measured and controlled rather than trusted.

The result is AI that is boring in exactly the right places. Operators trust boring.