The more strictly an AI follows procedure, the worse it handles the unexpected
The instinct when deploying AI is to write very detailed procedures: step one, step two, in case of X do Y. It feels safe. It works on the cases the procedure foresaw. On everything else, the more precisely the system follows the script, the worse it does.
Why precision becomes brittleness
A procedure is a record of the situations someone thought of. Real work keeps producing situations nobody thought of: a customer with two accounts, an order that is both urgent and incomplete, a regulation that changed last week. A system trained to follow the procedure exactly has nothing to fall back on when the procedure is silent. It either forces the case into the nearest rule, or stops.
People handle these cases because they know what the procedure is for. They can reason from the purpose when the steps do not fit.
Give it the purpose, not just the steps
The answer is not to loosen control. It is to give the system what people have: the reason behind the rules. What the process is meant to achieve, what must never happen, and how to weigh competing goals. A system that knows the purpose can handle a new case sensibly, or at least recognise that it is new and hand it to a person.
And watch what it does
Judgment brings variation, and variation needs watching. Keep the controls that catch drift: regular tests against known cases, a log of every decision, a person reviewing anything unusual. The combination is the point. Purpose lets the system cope with the unexpected. Monitoring makes sure it coped well.
Write fewer steps and more reasons. Then check the results.