A short guide to evidence, ownership and approval in AI workflows.
The signature still belongs to someone
We think “human in the loop” is too easy to say when nobody has named the human. A director still has to stand behind the report. A manager still has to act on it. We want the workflow to help that person understand the answer, including where it might be wrong.
Make doubt usable
An exception list is useful when it tells someone what to do next. We would want it to identify the missing source, the ambiguous mapping or the unsupported explanation. Uncertainty needs an owner and a next step. Otherwise it simply travels to the most conscientious person in the team.
Give the reviewer something to work with
We would ask what evidence the reviewer needs before deciding how to present the output. A traceable figure, the relevant source period and a visible assumption can make the difference between a quick check and rebuilding the report. That review effort belongs in the business case.
Practise disagreement
A handover should include an imperfect example. Let your team find the problem, explain it and correct it. We would learn more from that exercise than from asking whether everyone enjoyed a training session. It shows whether people can challenge the system when we are no longer beside them.
Our position
We want AI to give capable people more room to exercise judgment. That requires a process which welcomes a challenge and makes corrections possible. Before your next demonstration, ask to see what happens when the answer is incomplete. We would expect to answer that question about our own work.