Human Review Sampling for AI Agents: How Often to Check Their Work
A risk-based method for reviewing AI agent work: check more on high-stakes tasks, less as measured accuracy holds, and reset when behavior drifts.

Human review sampling means a person checks a chosen portion of an AI agent's finished work, then uses those results to decide how much checking comes next. The right rate is set by the cost of the worst plausible error in each task group, raised when risk or error findings rise, and lowered only after accuracy stays steady over a defined window. A single rate for all work is almost always wrong, because it either wastes reviewer time on low-stakes tasks or misses rare failures in high-stakes ones.
Key takeaways
- Set the review rate by the cost of the worst plausible error in each task, not by how routine the task feels.
- Sample by task type and risk tier, never as one blended number across all agent work.
- Reduce review only after measured accuracy stays steady across a full window of real outputs.
- Return to heavier review after a model change, a tool change, a serious error, or early signs of drift.
- Give the sampling plan one named owner and record every change to the rate with its reason.
Why one review rate fails
A common first move is to pick a percentage and apply it everywhere. The problem is that agent work is uneven. A summary of an internal meeting and a change to a supplier bank account may come out of the same agent on the same day, yet a wrong summary is a minor annoyance while a wrong bank detail is a payment error that someone has to unwind. A blended rate treats both the same, so it either overloads reviewers with trivial checks or leaves the dangerous tasks under-examined.
The second problem is that a flat rate hides trends. If reviewers only see an overall pass rate, a deterioration in one narrow task can be masked by strong results elsewhere. Sampling works best when each group of tasks has its own record, its own rate, and its own owner who can see when that group changes.
Start by grouping tasks by risk
Before choosing any rate, sort the agent's tasks into groups that share the same consequence of error. A practical starting point is three tiers.
- High risk: actions that move money, change records that other systems depend on, send messages under the company name, or touch personal or regulated data. Start with full review or route the action through an approval step before it is written.
- Medium risk: outputs that customers or staff rely on but that can be corrected cheaply, such as a draft reply, a meeting summary used internally, or a classification that feeds a queue someone checks.
- Low risk: lookups, formatting, and internal drafts where errors are visible at once and reversible without outside effects.
Within each tier, separate task types when their failure modes differ. Refund estimates and address updates may both be high risk, but they fail in different ways and deserve separate records. Keep the groups few enough that a reviewer can learn each one, and specific enough that a finding points to a cause.
Measure accuracy before you trust a rate
Sampling only produces useful decisions if reviewers score outputs the same way. Write a short rubric for each task group with a small number of checks, such as whether the correct record was changed, whether every required field is present, whether the cited source supports the claim, and whether the output stayed within the agent's permitted actions. Reviewers record a pass or fail, the error category if it failed, and whether the output was used before correction.
A review record needs enough detail to be read later by someone who was not there. Useful fields include the task group, a reference to the input, the output, the reviewer, the verdict, the error category, the time taken to review, and whether anyone acted on the output. The format can be a spreadsheet at first. What matters is that the same fields are captured every time, so trends can be seen across weeks rather than argued about in meetings.
A review rate is a decision about risk, and each change to it should be explainable to the person who carries the consequences.
Illustrative scenario: a procurement team's sampling plan
The following scenario is illustrative and does not describe a specific company. A procurement team uses an agent to draft supplier onboarding emails and to update vendor records. Vendor record updates are high risk because a wrong bank detail causes a payment error. Onboarding emails are medium risk because they are visible to suppliers but easy to correct.
During the first period, the team reviews every vendor record update before it is written, and reviews emails in a random sample. When the record updates have been clean across a full window of real volume, the team moves record updates to sampling, while keeping a small approval step for any change to bank details. Later, the supplier system changes its date format. Sampled records start showing date errors, so the group returns to full review until the parser is fixed and the new outputs are verified.
Sampling method that rises and falls with risk
The method below can be run with a spreadsheet and a shared rubric. It does not require special tooling, though it does require a stable record of outputs.
- Step 1: Group tasks by risk tier and task type, and write one owner for the group.
- Step 2: Set a starting review level for each group. For high-risk groups, begin with full review or an approval step.
- Step 3: Draw samples at random within each group. Include a fixed slice of outputs the agent rated as confident, not only the ones it flagged as uncertain, because confident errors are the ones most likely to pass unnoticed.
- Step 4: Score each reviewed output against the group rubric and record the reason for every failure.
- Step 5: Review findings per group across a rolling window long enough to include normal swings in volume and in input types.
- Step 6: If a serious error appears in any group, stop automatic release for that group and return to full review.
- Step 7: If the whole window shows steady accuracy and no serious errors, reduce the rate one level, not straight to zero.
- Step 8: Log every rate change with the date, the evidence, the reviewer, and the approver.
The key discipline is that the rate moves in small steps and in both directions. A rate that only goes down is not sampling. It is a gradual withdrawal of attention.
When to reset the rate
Some events should return a group to heavier review regardless of recent results, because past accuracy no longer predicts future accuracy. Treat these as standing triggers.
- A change of model version or a provider update to the model behind the agent.
- A new tool, a changed API schema, or a changed format in a source document the agent reads.
- Any serious error found in a sampled output, even if the rest of the sample passed.
- A sudden change in volume, input type, or customer segment that the group has not handled before.
- A change to the agent's permissions or instructions.
- A slow rise in corrections across several weeks, which is a drift signal rather than a single bad day. The drift quality decline guide explains how to separate drift from noise.
When a reset happens, record it as an event, not as a failure of the plan. The record should say what changed, which groups were affected, and how long the heavier review will last.
Sampling detects problems, approval prevents them
Sampling is a detection method. It finds errors after an output exists, which is acceptable for low-risk work and for medium-risk work that can be corrected before it reaches anyone outside the company. For actions that cannot be undone, such as an outgoing payment or a change that triggers an external notice, review must happen before the action, through an approval step. Permissions should follow the same logic: the agent should hold only the access its tasks need, and each approval boundary should be written down. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) is a useful reference for thinking about least privilege in this way. When a human is needed for a decision the agent cannot safely make, the handoff itself should be designed, as described in the guide on human escalation for AI agent failures.
Be Human designs these boundaries as part of the first engagement, then operates the sampling plan in managed operations. That work includes recurring evaluations of reviewed outputs, permission reviews, a change log that the business owner can read, and controlled adjustments to the review rate when the evidence supports them. The aim is an operating record that shows, for each task group, what the agent did, who checked it, and why the level of checking is what it is today.
Choosing the first task to review
Start with one task group, not the whole agent. Pick a group where errors are costly enough to matter but where a reviewer can read the output in a few minutes, and write its rubric before the first output arrives. The sampling guide for agent outputs describes how to size the first sample for a single group. If you are not sure which task should carry the first review plan, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
How often should a person review AI agent work? There is no universal rate, and anyone who offers one is guessing. Begin with full review or an approval step for high-risk actions, then move medium and low-risk groups to random sampling once their accuracy has held steady across a full window of real outputs.
How big should a review sample be? A sample must be large enough to reveal the errors that matter in its task group, and it should be drawn at random across the window rather than picked by hand. When volume is small, a handful of outputs cannot show a rare error, so heavier review should last longer for those groups.
Should reviewers check only the outputs the agent marks as uncertain? No. Uncertainty flags are useful, but the sample should also include a fixed slice of outputs the agent rated as confident, because confident errors are the ones most likely to pass without notice. Mixing both gives a fairer picture of real accuracy.
Who owns the sampling plan? One named operations owner should be responsible for the plan, with reviewers scoring against a shared rubric. Every change to the review rate should be recorded with its reason, the evidence behind it, and the person who approved it, so the decision can be read and questioned later.
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.


