An approval card says “Update customer record,” and a busy employee clicks approve. The AI then changes the address, contact preference, and account owner. The employee approved a vague label rather than the actual operation.
A useful AI approval shows the proposed change, its target, its evidence, and what happens when the action is applied.
Make the proposed change inspectable
For a fictional address update, the approval should identify the customer record, existing value, proposed value, source interaction, and any uncertainty. Avoid hiding important fields behind an optimistic summary.
| Approval detail | Why the reviewer needs it |
|---|---|
| Target record | Confirm the right customer or account |
| Before and after values | Understand the actual change |
| Source and request time | Verify the evidence is relevant |
| Side effects | See notifications or dependent updates |
| Expiry or version | Avoid approving stale work |
Include an explicit reject or edit path where the workflow needs it. A binary approve button is not enough when reviewers regularly discover incomplete information.
Bind approval to the exact request
Store the request version or equivalent immutable representation being approved. If the AI revises the operation afterward, the previous approval should not silently authorise different values.
Recheck current state before applying the action. Another employee may have already corrected the record. An approval based on an old value needs conflict handling rather than an unconditional overwrite.
Also verify the reviewer's permission at execution time. A person who has lost access should not retain authority through an old approval link.
Keep approval and execution separate in the status model
“Approved” means an authorised person agreed to the proposed action. It does not mean the external system accepted it.
Track pending review, approved, executing, completed, rejected, expired, and unresolved outcomes as appropriate to the workflow. Give staff a reference they can use to investigate a failed execution without approving it again.
For repeated clicks or retries, preserve the original operation identity. The database constraint guide explains one way to enforce uniqueness at a trusted data boundary.
Test what happens after the reviewer looks away
Use fictional cases where the record changes after approval, a second reviewer acts, the link expires, permission is revoked, and the external write times out. The interface should explain the resulting state without pretending approval completed the operation.
This adds more design work than sending a generic yes-or-no message. It also makes human review meaningful. A checkpoint that hides the actual changes can create the appearance of control while leaving the decision with the AI.
For public claims rather than record updates, the support-feature guide shows another case where the reviewer needs authoritative evidence.
AI approval questions
Should staff see every model thought?
No. Show the evidence and proposed effects needed for the decision, not an unverified internal narrative.
Can one approval cover later changes?
Only when the authorised scope explicitly covers them. Otherwise require a new review.
What should an acceptance test demonstrate?
Approve one exact change, alter its target state, and verify safe conflict handling. Use that case in a workflow brief.