Reviewing an AI agent's decision means reading the record of what the agent saw and proposed, finding which bound held it for you, checking the sources behind each fact, and then approving, editing or declining it with a short written reason. Your decision is filed under your name, so the review is a judgement you own, not a click.
Why are decisions held for review at all?
A governed agent never acts on its own say-so. Every action it proposes meets five bounds at once: its role scope, its certificate for that capability, the sources behind the facts, the business's own limits, and the platform guardrails. The lowest one wins, and the verdict names which bound decided.
There are four verdicts. Allowed means the agent may act. Held means it has drafted the action and is waiting for a named person: you. Refused means the action falls outside the agent's bounds. Withheld means the platform blocks it on purpose for every agent at every depth. Reviewers mostly deal with held decisions, and the record tells you why each one reached you.
What are the steps of a review?
- Read the proposal. What does the agent want to do, for whom, and for how much?
- Find the deciding bound. Was it held because of an amount over the business's limit, a weak source, a capability the agent has not earned at this depth, or a rule that always needs a person?
- Check the sources. Each fact carries its provenance: sealed from the system of record, signed by a named party, declared by the customer or a person, or external from email or the web. Ask whether the decision rests on strong enough ground.
- Check against policy and context. Does the proposal fit the rules and what you know of this case?
- Decide. Approve if it is right; edit if it is nearly right; decline if it is wrong or unsupported.
- Write the reason. One or two sentences: what you found and what happens next.
- Report patterns. If you see the same issue repeatedly, tell the agent's owner so its sources or bounds change.
What does a review look like in practice?
Take a reviewer we will call Tariq at a lettings agency. An agent has drafted the return of a tenant's deposit of 1,200, with a deduction of 400 for cleaning, and held it because deductions always need a person.
Tariq reads the record. The agent relied on the check-out report, which was prepared and signed by the agency's independent inventory clerk, and on the landlord's email asking for "a full professional clean". He opens the check-in report, also signed, and finds the property was let "in fair condition, not professionally cleaned". The agency's own policy is that tenants return a property in the condition they received it, allowing for fair wear.
He edits the deduction to 120 for specific marks the check-out report photographs, writes "Check-in condition 'fair, not professionally cleaned'; deduction limited to items evidenced in signed check-out report," and approves. The landlord's email was a declared request, not evidence.
What makes a review good or poor?
| Good review | Poor review |
|---|---|
| Reads the sources, not just the summary | Approves because the draft looks polished |
| Names the deciding bound | Does not know why the case was held |
| Edits when the agent is nearly right | Declines everything that is not perfect |
| Writes a specific reason | Writes "not approved" or "see notes" |
| Reports repeated problems | Fixes the same error every week in silence |
Employers judge reviewers on decision quality: right outcomes, consistent with colleagues, with a clear decision rationale.
How do you get better at it?
Practise on real or simulated cases and keep a log of each decision and reason; the decision log template for practising AI agent review gives a format. How to learn to read an AI agent decision record goes deeper on step two. And if you are wondering how this translates to job adverts, see what employers mean by experience with AI agents and AI skills for job seekers.
Where the Academy fits
Pryme Intelligence Academy's Operator track is built around this work: how a governed agent decides, human review, and teaching the agent. Its practicum is two weeks of review decisions to a standard of nineteen in twenty correct and no missed red line, signed off by a Pryme assessor.
Questions people ask
How long should a review take?
It depends on the case. Routine held items may take a minute or two; complex ones much longer. Speed matters, but employers care more that decisions are right and explained.
What if I disagree with the agent but cannot prove it is wrong?
Say what you need to be sure, and hold or escalate rather than approve on doubt. A reason such as "need supplier confirmation before release" is a valid outcome.
Does my review change the agent?
Not directly. Your decisions and reported patterns feed the agent's owner, who changes its brief, sources or bounds. Agents are qualified, not retrained, from that evidence.