How Many AI Agent Outputs to Review: A Risk-Based Sampling Method
A risk-based method for deciding how many AI agent outputs to review, which ones to always check, and how to record what the sample misses.

Review enough agent output to catch the failures that would cost you money, customers or trust, and no more than your team can read carefully. For most business agents, that means a fixed floor of full review for any output that touches an outside party or a system of record, a risk-weighted sample of everything else, and extra checks after any change. Sampling never proves that the unchecked work is correct, so the method also has to state plainly what is not being checked.
Key takeaways
- Set the review rate by what a mistake would cost, not by how many outputs the agent produces.
- Keep full review for actions with external effects until the agent has earned less scrutiny.
- Build the sample by task type, channel and recent change, so rare paths still get read.
- Trigger extra review after every model, prompt, tool or policy change, and after every incident.
- Write down what is not being checked, and review that list on the same cadence as the sample.
Start by sorting outputs by the cost of a mistake
Before choosing a number, sort the agent's outputs into tiers. The tier decides the review rate, so the sorting is the most important step. Use the question: if this output were wrong and nobody caught it, what would happen next?
- Tier A, low consequence: internal summaries and drafts that a person reads before anything leaves the team.
- Tier B, moderate consequence: customer-facing text, status updates and quotes that a person could correct after sending, with limited harm.
- Tier C, high consequence: anything that moves money, changes a record other teams rely on, commits to a date or price, or handles personal data.
Write the tier next to each task type in a short register. If a task could belong to two tiers, put it in the higher one. Revisit the register whenever a new tool or channel is connected, because the consequence of a task often changes when its permissions change.
How many outputs to review: a starting method
A reasonable starting method has four rules. Use them as a first version and adjust after the first few review cycles.
- Tier C: review every output before it takes effect. Do not sample this tier during the early period, and do not reduce it until the agent has a clear record on that exact task type.
- Tier B: review a fixed minimum each week for each task type and channel. Set the minimum as a count, not as a percentage of volume, because a percentage of a large stream is a lot of reading and a percentage of a small stream is almost nothing.
- Tier A: review a small random draw each week, plus every output the agent itself flags as uncertain or incomplete.
- Raise every rate for a new workflow, a new model version, a newly connected tool, or the first weeks after an incident. Lower a rate only after several consecutive cycles without a material miss, and lower it one step at a time.
The counts are a matter of team capacity and of how many distinct situations the agent handles. A reviewer who reads carefully can cover a limited number of items in a sitting. If the minimum is more than that, the review is not a review, so reduce the number of task types being sampled or narrow the channel rather than skimming.
Why random sampling alone misses the failures you fear
Random sampling is good at showing the typical quality of common work. It is weak at finding rare failures, and rare failures are often the ones that cause the largest harm. If a particular error appears only on a small share of cases, a random draw of a few items may never contain one, and a clean review will look reassuring when it means very little.
For that reason, add targeted samples on top of the random draw. Good targets include outputs that involved a tool error or a retry, outputs with unusually large amounts, dates near a month or quarter boundary, customer records that match a known exception, requests that cross time zones, and any output where the agent's own confidence signal was low.
A clean random sample tells you about the typical case. Targeted samples tell you about the cases where a failure would hurt.
Illustrative scenario: a billing agent drafts a few hundred routine answers each week and handles a handful of refund approvals. A random draw of routine answers would rarely include a refund problem. The team sets refund approvals to full review and samples routine answers by category, so the rare path is read every time it happens.
For a deeper treatment of how automated graders can pre-screen outputs without replacing this human layer, see LLM judges for agent evaluations.
What a reviewer should check and record
A review that only asks whether the output looks good will miss the failures that matter. Give reviewers a short checklist tied to the task, and record the result in a form that can be searched later.
- Did the agent take the right action, or produce the right output for this input?
- Did it use only the sources and tools it was permitted to use?
- Do dates, amounts, names and identifiers match the source record?
- If this were sent or applied as written, would a customer, colleague or system be harmed?
- Verdict: accept, accept with a note, reject, or escalate.
- Reason code from a short fixed list, so patterns can be counted later.
Record the reviewer, the time, the task type, the tier, and the model and prompt version that produced the output. Without the version, a miss cannot be tied to a change, and the review cannot tell you whether an update helped. Rejected outputs should feed back into the agent's rules and into the test set, so that the same miss is caught earlier next time.
Say what is not being checked
Every sampling plan leaves most of the work unread. A plan that hides this is less safe than one that states it. Write a short coverage statement each cycle with four parts.
- Which outputs were never sampled, and why: volume, channel, time of day, or a tier that was reviewed at a lower rate by decision.
- What fraction of each tier was actually read, and which items were skipped by reviewers.
- Known blind spots: failures that reviewers are not trained to recognize, such as subtle policy errors that only a specialist would catch.
- Failures that appear only downstream, such as a complaint two weeks after a message was sent, and how those signals reach the review queue.
Also check for selection bias. Reviewers often pick the outputs that look easiest to judge, which makes the sample cleaner than the work. Draw items from a list by a rule, not by hand, and record the draw. The NIST AI Risk Management Framework at [NIST](https://www.nist.gov/itl/ai-risk-management-framework) describes measurement and documentation of AI risks as an ongoing practice, which fits this kind of coverage statement.
How Be Human handles this in managed operations
In managed operations, Be Human works with the business owner during the first engagement to define the tiers, the review floors and the targeted samples for the agent's actual role. The team then runs the review queue, keeps the record of each verdict and reason code, and includes the coverage statement in recurring evaluations and permission reviews. When a sample shows a miss, the incident path and the rule change are tracked together, so the review rate and the operating rules change on the basis of recorded evidence. The business keeps the decision about how much authority each tier receives.
Tightening and loosening the rate over time
Review rates should move in both directions, but on a schedule and with evidence. Tighten immediately after a material miss, a new integration, a model change or a complaint that traces back to the agent. Loosen only after several consecutive cycles in which the sample is clean, the coverage statement shows no new blind spots, and the targeted checks found nothing new. When you loosen, change one tier or one task type at a time, so that if a problem appears you know which change caused it. Keep the Tier C floor in place for external actions unless the business has deliberately accepted the risk in writing.
To see how controlled trials can come before any authority is granted, read how to run an agent in shadow mode before giving it authority. If you are not yet sure which work in your business is suitable for an agent, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
How many AI agent outputs should I review each week? Start from the tiers rather than from a fixed percentage. Review every high-consequence output before it takes effect, set a minimum count for each task type in the moderate tier, and draw a small random sample from the low-consequence tier. Raise the rate after any change and lower it only after several clean cycles.
Is a random sample enough to trust an agent? A random sample shows the typical quality of common work, but it cannot prove that rare failures are absent. Add targeted samples for tool errors, unusual amounts, boundary dates, low-confidence outputs and other cases where a failure would do the most harm. Pair the samples with a written statement of what was not checked.
When can I stop reviewing an agent's output? You should not stop entirely while the agent takes actions with external effects or changes records that others rely on. You can reduce the rate for lower-tier tasks one step at a time after sustained clean results, and you should raise it again after any change to the model, prompts, tools or policies. Keep the floor for high-consequence work unless the business has accepted the risk in writing.
Who should do the reviews? Choose people who know the business rule behind each task, not only the people who built the agent. Time-box the reviews so reviewers can read carefully, rotate the work to limit fatigue, and record each verdict with a reason code. The reviewers' findings should feed back into the agent's rules and test set so the same miss is caught earlier.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


