When AI Enters the Workflow, Counterexamples Become Part of the Brief

Once AI enters a workflow, an instruction such as “assign each expense to the correct category” may look complete yet still produce repeated errors. The problem is often not that the model failed to understand the words. The rule describes the normal case but leaves its boundaries unstated. A standard hotel bill may belong under travel, but a cancellation caused by a client’s last-minute change, accommodation mixed with a private trip, or an urgent expense without a receipt may need to be classified, returned or escalated. That choice has to be explicit.

In the past, experienced staff often supplied these exceptions from memory. AI does not naturally possess an organisation’s unwritten conventions. It can only infer them from the material available, and it may present an uncertain inference as a complete result. A work instruction is therefore changing from “tell someone what to do” into “define what counts as complete, where the rule does not apply, and when the process must stop”.

A good AI instruction consequently needs more than a rule. It should include representative examples, counterexamples and escalation conditions. Examples show how to proceed, counterexamples mark the boundary, and escalation conditions preserve a point of entry for human judgement. AI does not merely accelerate execution. It also forces organisations to write down the knowledge of exceptions that previously remained in people’s heads.