Do not approve it. Treat the unease as information: name what triggered it, check the sources the draft rests on, then approve with a recorded reason or decline and write down why. A reviewer's instinct catches what rules miss, and pausing almost always costs less than a wrong approval.
Why does "feels wrong" deserve attention?
Experienced reviewers carry patterns they cannot always state. A customer's tone that does not fit the request. A figure that is plausible but unusual for this kind of case. A sentence that is accurate and still somehow misleading. These signals come from thousands of earlier cases, and they often fire before the reviewer can say why.
Agents produce fluent, well-structured drafts. That fluency is a strength and a risk: a polished draft feels finished, which makes it easier to approve something that would have looked doubtful in rough form. The reviewer's unease is often the only counterweight. The reasoning is set out further in the reviewer is the control, not the agent.
What should the reviewer do, step by step?
- Stop and name it. Write one line: "the amount seems high for this customer" or "the reply answers the question but not the concern". Naming turns a feeling into something you can test.
- Check the sources behind the draft. Every fact carries a provenance stamp: sealed, signed, declared or external. Look for the fact your unease attaches to. Is it resting on a customer's declaration or an emailed document when the decision needs something stronger?
- Check the verdict and the bound. The draft is held for a reason, and the verdict names the bound that decided. Does that bound fit what you are seeing, or did the item reach you for an unrelated reason?
- Look at the case, not only the draft. Read the customer's original message, the history, or the document the agent summarised. Drafts compress; the detail you need may have been dropped.
- Decide, and write the reason. Approve if the check clears your concern, noting what you checked. Decline if it does not, stating why in a sentence the builder can use.
- Do not quietly fix it. Editing a draft into shape without recording what was wrong loses the lesson. If you change it, say what you changed and why.
What may the agent do, and what stays with the reviewer?
| The agent | The reviewer |
|---|---|
| Drafts the action with its sources | Decides whether to approve, edit or decline |
| States its reasoning and the facts used | Judges whether the reasoning fits the whole case |
| Holds items outside its bounds | Owns the outcome of anything they approve |
| Receives the decline reason in its next qualification round | Writes that reason clearly |
Agents that only draft, with every action sent by a person, are described in what draft-only mode is. The role itself is defined in what an AI agent reviewer is and, more briefly, in what an agent reviewer is.
What does this look like on a real case?
Take an insurer's claims team. The agent drafts a settlement of 2,500 for water damage, held because it is above the agent's limit. The draft is tidy: policy clause cited, repair estimate attached, no prior claims. The reviewer feels something is off and writes: "estimate seems precise for a first visit".
Checking sources, the reviewer finds the repair estimate is external, a PDF the claimant emailed, from a contractor with no web presence and an address that matches the claimant's own street. The reviewer declines, requests an independent assessment, and notes "estimate external, contractor unverifiable, address near claimant". The assessor later finds the damage is a fraction of the claim. Nothing in the draft was false on its face; the unease was right.
What controls support reviewers who pause?
Pausing must be safe. Reviewers should never be measured only on speed, held items should escalate rather than auto-release, and declines with reasons should be valued as much as approvals. Where decisions have legal or similarly significant effects on people, the law may require meaningful human involvement; GDPR Article 22 and agents that draft but do not decide explains why a rubber stamp does not count. More skills for reviewers by role are in AI agent skills by role. Training such as the Operator track at Pryme Intelligence Academy builds this habit through practice on real review decisions.
Questions people ask
What if I cannot find anything wrong after checking?
Approve it, and note what you checked. Unease that does not survive a check is still worth the minute it took, and the note shows the review was real.
Should I ask a colleague?
Yes, when the stakes are high or the feeling persists. A second view is cheap compared with a wrong approval.
What if the agent keeps producing drafts that feel wrong?
Tell the agent's owner. A pattern of unease often points to a source, bound or brief problem that a builder can fix.