Responsibility map

Start by drawing who is downstream. Not the users of the tool — the people the output lands on. The applicant whose CV is screened, the customer whose complaint is categorised, the employee whose shift is assigned. Most AI harm is invisible to the person operating the system because they never meet the people affected by it.

For each of them, ask one question: if this goes wrong for them, how would they find out, and what could they do about it? A workflow where the answer is "they would not, and nothing" is a workflow that needs a human review step regardless of how accurate it tests.