Why AI does not get to decide
The interesting question about AI in operations is not what it can do. It is what it is allowed to do - and who can stop it.
4 min read
A division of labour, not a debate
Most arguments about AI at work get stuck on capability - is the model good enough? Wrong question. A system can be impressive and still be unsafe to run, because capability and authority are different things. The workable split is older than AI: separate the work of understanding, the work of deciding, and the power to intervene - and give each to the layer that does it best.
AI interprets and recommends
Reading messy input is what models are good at: an email that is half a request, a scanned delivery note, a complaint written in a hurry. The model turns unstructured reality into structured proposals - what this probably is, what it probably needs, what it might be worth. Proposals, not actions.
Deterministic code decides
Whether a request is routed, priced, flagged or held is a decision - and decisions in a business need to be explainable, repeatable and auditable. So the deciding is done by ordinary code: rules you can read, test and stand behind. If the code cannot justify a decision, the decision does not happen.
Humans override
Every flow keeps a way out. A person can see what the system interpreted, what rule fired, and take the work back. Overrides are recorded, not bypassed - because an override that happens twice is not an exception, it is a rule the system has not learned yet.
What goes wrong when the layers blur
When AI decides directly
The system works until it does not, and nobody can say why a decision was made. Confident wrongness ships at machine speed. The failure mode is not dramatic - it is quiet erosion: a wrong classification repeated four thousand times before anyone notices.
When humans do everything
Judgement gets spent on routine. The expensive people read every inbound item, copy every field, triage every queue. Work that needed no judgement consumes it anyway, while the genuinely ambiguous cases queue behind it.
When nobody can override
A flow without an override is a trap. Staff route around it with spreadsheets and side-channel emails, and now you have two systems: the official one and the real one. The shadow process is always worse than no process.
Why this matters commercially
This division is not a safety ornament - it is what makes the economics work. Interpretation is cheap and gets cheaper. Decisions stay inspectable, so they can be defended to a customer, an auditor or a regulator. And because overrides are recorded, the system gets more honest over time: every correction is evidence about where the rules are wrong.
It is also how we run our own commercial operation - the enquiry that becomes a brief, the lead that becomes a worklist, the scheduling that books a call. AI reads and recommends; deterministic code classifies and routes; a person can always take the pen. The same principle we would put inside your operation, because we are already running it inside ours.
Occasional notes on building systems people can actually audit - and overrule.
Want to see the principle applied to a real workflow?
About the Review