Case: an impact assessment before launch
A bank was weeks from launching an AI system to flag loan applications for review. Technically it worked well. Under ISO 42001, before deployment, they ran an impact assessment — and paused.
The assessment asked plain questions the build phase hadn't: who is affected by this system's decisions, how could it be wrong, and who bears the cost when it is? The model, it emerged, had learned from historical decisions that carried historical bias. Left unexamined, it would have automated and scaled that bias, flagging some applicants disproportionately while wearing the authority of objective maths.
This is exactly what impact assessment is for: catching harm before launch, when it is cheap to fix, rather than after, when it is a scandal and a lawsuit. The team retrained on corrected data, added monitoring for disparate impact, and kept a human decision on every flag.
The launch slipped six weeks. That delay was the cheapest insurance the bank bought all year. Governance before deployment is not friction — it is the difference between a tool and a liability.