The question every heating firm asks first
Before anyone signs up, the same question comes up: what happens when a caller says they can smell gas? It is the right question, and it deserves a precise answer rather than a reassuring one.
The answer: a fixed rule, not a judgement call
A report of a gas smell, a carbon monoxide alarm, or wording matching an approved list of danger signals is caught by a deterministic safety rule before any AI-generated response is considered. The caller receives the client-approved safety wording and is pointed to the National Gas Emergency Service on 0800 111 999. A human on the team is alerted at the same time. No language model decides what to say on this path.
This is a deliberate architectural choice. The governing principle is: AI interprets language. Code authorises actions. A model can understand that “there is a funny smell near the boiler” and “smells like gas in the kitchen” mean roughly the same thing. It should not decide what happens next when the answer could be a matter of safety.
Why this matters beyond gas safety
The same logic governs every other safety-relevant decision—service-area boundaries, quiet hours, bank holidays, and anything Gas Safe registration or compliance touches. These are checked against fixed, testable rules that behave the same way on the ten-thousandth call as on the first.
A language model’s job is understanding what someone said. A rule engine’s job is deciding what is allowed to happen. Blending those responsibilities is how an automated system can do something no responsible person approved—not because the AI was malicious, but because it was asked to be the judge instead of the interpreter.
What this looks like for your firm
During onboarding, your safety wording, escalation contact and trade-specific rules are written into the deterministic layer—not left to the system to improvise from a generic template. A new rule is a reviewed configuration change, approved before it goes live, not a prompt tweak.