FIELD NOTES / adoption
Design the review step
Human review needs an interface and a decision, not just a checkbox.
When a person must review an AI result, show what they need to make the decision. Highlight the proposed action, supporting evidence and unresolved questions.
A reviewer should be able to correct, reject or return the work without losing the original context. Record the decision and keep it separate from the model's confidence. Avoid suggesting that a person approved something merely because they opened a page.
Observe a realistic review session. If checking the output requires more effort than doing the original work, revisit the scope, evidence presentation or automation boundary.