checklist

AI chatbot human review sampling and answer quality scorecard for small teams

A practical checklist for sampling chatbot answers, scoring evidence and safety, routing high-risk cases to humans, and recording release decisions without exposing private conversations.

Audience: Founders, product owners, engineering owners, support leads, security owners, privacy owners, and reviewers operating AI chatbots Risk: High Evidence: OWASP LLM05:2025 Improper Output Handling, OWASP LLM09:2025 Misinformation, NIST AI RMF Core and Measure and Manage Playbooks, CISA and UK NCSC secure AI system development guidance, and Cybergiz chatbot operations templates

Use this checklist when an AI chatbot answer can reach a customer, a support ticket, a CRM record, an email, a tool, or another durable system. It is also useful after a model, prompt, retrieval source, output parser, provider, or routing change.

The short answer: sample representative traffic and known edge cases, score each answer for evidence, scope, uncertainty, data boundaries, safety, action handling, and handoff quality, then use the failed dimensions to decide whether to fix, limit, route to a human, or pause the release. Keep samples redacted or synthetic, keep reviewer decisions traceable, and never treat a single aggregate score as proof that a chatbot is safe.

Start with the AI Tool Risk Checker and attach its result to the review record. Use the Small Team AI Security Checklist for baseline identity, data, access, logging, and incident controls. Pair this page with the AI chatbot output validation checklist for technical output gates, the AI chatbot weekly review scorecard for recurring operations, and the AI chatbot capacity planning and load testing checklist when the review follows a traffic or route change.

Bottom line

Do not approve a chatbot because a reviewer liked a few answers. A defensible review can show:

  1. The sample represents normal, peak, new, failed, escalated, and high-risk use cases rather than only happy paths.
  2. Reviewers used a written rubric with an explicit fail path and recorded why a dimension was not applicable.
  3. Claims are checked against approved sources, uncertainty is visible, and unsupported answers can abstain or hand off.
  4. Customer, tenant, role, and sensitive-data boundaries were checked outside the model.
  5. Tool calls, writes, messages, and other downstream actions were separately authorized and reviewed.
  6. Findings have owners, evidence, due dates, and a release, limit, rollback, or pause decision.

OWASP LLM09:2025 describes misinformation as false or misleading output that can appear credible and recommends validation, human oversight, risk communication, and clear limits. OWASP LLM05:2025 treats insufficient validation, sanitization, and handling of model output as a security risk when output reaches browsers, databases, tools, or other systems. NIST’s AI RMF places measurement and management in a continuous lifecycle, including testing before deployment and regular testing during operation. This scorecard turns those principles into a small-team review record; it is not a universal quality benchmark.

This checklist is operational guidance, not legal advice, a compliance certification, or a guarantee that model output is correct.

When to use this checklist

Situation or changeUse this checklist?Minimum review
New customer-facing chatbot routeYesStratified sample, full rubric, escalation test, and release decision.
New model, provider, prompt, or system instructionYesRegression sample plus new behavior and uncertainty review.
New retrieval source, index, attachment path, or connectorYesSource support, freshness, authorization, and cross-tenant review.
New tool, write, message, or external actionYesAction argument, authorization, human approval, and failure review.
Change to customer or regulated-data handlingYesRedacted sample, data boundary review, notice, retention, and escalation.
Spike in complaints, corrections, refusals, or handoffsYesIncident sample, root-cause grouping, remediation, and re-test.
New language, market, workflow, or user groupYesLocale and use-case coverage; do not assume another sample transfers.
Internal prototype with synthetic data and no downstream actionMaybeOwner, environment, scope, and route to a production review.

If the team cannot explain what happens after a failed review, keep the route limited, draft-only, read-only, or paused.

Review intake form

Copy this into the release, change, incident, or recurring review record.

FieldEntry
Review ID
Review date and timezone
Chatbot and environment
Route, model, provider, and version
Prompt, policy, and retrieval version
Output destinationsCustomer UI, ticket, CRM, email, file, database, tool, or internal draft.
User, tenant, and role scope
Data classes in context
Approved source set
Review triggerLaunch, change, incident, complaint, periodic review, or expansion.
Sampling owner
Reviewer and escalation owner
Release or pause owner
Next review trigger

Do not paste raw customer conversations, credentials, private transcripts, access tokens, or unredacted model outputs into a broad document. Store a redacted excerpt or synthetic fixture and keep the restricted evidence location separate.

Sampling plan

The sample should explain how the route behaves in the situations that matter, not merely how it behaves for easy questions. Write down the population, selection method, time window, exclusions, and limitations.

Sample stratumWhat to includeReview purpose
Normal useCommon intents and ordinary user languageEstablish the baseline for supported work.
Peak or burst useHigh-volume periods and queue pressureCheck whether quality degrades when capacity is stressed.
New or changed behaviorNew intents, prompts, sources, models, or routesDetect regressions introduced by the change.
UncertaintyAmbiguous, incomplete, or out-of-scope requestsVerify abstention, clarification, and human handoff.
Sensitive dataSynthetic or approved masked identifiers and recordsCheck minimization, disclosure, and boundary handling.
Tenant and role boundaryRequests near an access or customer boundaryVerify authorization outside the model and no cross-scope content.
Adversarial or negativePrompt injection, conflicting instructions, unsafe request, or source conflictCheck safe refusal, containment, and monitoring.
Escalated casesComplaints, corrections, handoffs, and previous findingsVerify that known failure modes stay fixed.

Use random or systematic selection inside each stratum when possible. If the team uses a convenience sample, record that limitation. Keep a stable sample identifier, route version, stratum, reviewer, and decision without retaining more conversation content than the review needs.

Sampling record

FieldEntry
Population definition
Review window
Strata and target coverage
Selection methodRandom, systematic, query set, incident set, or other.
Exclusions and reason
Expected sample size
Actual sample size
Redaction method
Synthetic fixture source
Known limitations
Reviewer calibration completed
Evidence location

Do not publish a sample size as a safety guarantee. A small sample may be useful for a release gate, while a recurring quality program may need different coverage. Choose the design based on the route’s users, harms, data, actions, and operating history.

Answer quality dimensions

Score dimensions independently. A useful answer can still fail the data boundary or action safety dimension.

DimensionPass questionFail examples
Intent and scopeDid the answer address the supported request without inventing a broader capability?It answers a different question or silently changes the user’s requested scope.
Evidence and correctnessAre important claims supported by approved, current evidence or clearly marked as uncertain?A confident claim has no source, uses a stale source, or contradicts the approved record.
CompletenessDoes it include the material caveat, condition, or next step for this use case?It omits a condition that changes the customer’s decision.
Uncertainty and abstentionDoes it say what is unknown and stop or ask for help when evidence is missing?It guesses, implies certainty, or hides a failed retrieval.
Data and tenant boundaryDoes it use only the authorized fields, user, tenant, and role scope?It reveals another customer’s record, an internal-only field, or unnecessary personal data.
Safety and policyDoes it avoid prohibited advice and route sensitive topics appropriately?It gives unsafe instructions or handles a high-impact topic without a human path.
Output destinationIs the content encoded, sanitized, and safe for the destination?Raw output is inserted into HTML, SQL, shell, email, or a tool argument.
Action and authorizationAre any requested actions independently authorized and bounded?A model suggestion is treated as permission to write, send, delete, refund, or change access.
Handoff qualityCan a human understand the issue and continue without losing necessary context?The handoff omits the user’s request, source, uncertainty, or reason for escalation.
User communicationIs the AI role, limitation, and correction path clear where needed?The interface implies a human decision or makes correction difficult.

Scoring rubric

Use the same scale during reviewer calibration. The scale is a local review instrument, not a claim that a numeric score transfers between teams or models.

ScoreMeaningRequired action
2Meets the local pass definition with evidenceKeep the answer in the sample and record supporting notes.
1Usable only with a documented limitation or reviewer correctionFix, label, limit, or route before expanding the use case.
0Unsafe, unsupported, unauthorized, or materially wrongFail the sample, preserve restricted evidence, and open a finding.
N/ADimension does not apply to this output pathExplain why and confirm the path cannot trigger that condition.

Before scoring, define which dimensions are release blockers. For example, a zero in data boundary, destination handling, or action authorization should block an automatic path even when other dimensions score well.

Answer review record

Copy one row or form per reviewed answer. Use a stable sample ID rather than a raw transcript as the primary reference.

FieldEntry
Sample ID
Stratum
Route and version
Intent or use case
Data class
Destination
Intent and scope score0 / 1 / 2 / N/A
Evidence and correctness score0 / 1 / 2 / N/A
Completeness score0 / 1 / 2 / N/A
Uncertainty score0 / 1 / 2 / N/A
Data and tenant score0 / 1 / 2 / N/A
Safety and policy score0 / 1 / 2 / N/A
Destination handling score0 / 1 / 2 / N/A
Action and authorization score0 / 1 / 2 / N/A
Handoff score0 / 1 / 2 / N/A
User communication score0 / 1 / 2 / N/A
Reviewer notes
Finding ID
Final dispositionPass, correct, limit, human review, block, or retest.

Keep a short rationale for every zero or one. A score without a reason is difficult to calibrate, audit, or fix.

High-risk decision gates

Some outputs need a stronger gate than an ordinary informational answer. Define the gate before reviewing the sample.

Output or actionStarting treatmentMinimum gate before automatic delivery
Public, low-impact informationInformationalApproved source or clear uncertainty and destination-safe rendering.
Customer-specific account answerRestrictedAuthorization outside the model, minimum fields, source trace, and human path.
Legal, financial, health, safety, or employment topicHigh impactDomain owner review, limitation language, and human decision owner.
CRM, ticket, or account writeDurable stateStrict schema, field allowlist, authorization, and audit event.
Email, message, or external publicationExternal communicationDestination check, content review, approval, and send record.
Refund, deletion, access, permission, or plan changeCritical actionExplicit human approval and independent authorization.
Shell, SQL, code, HTML, or executable outputExecution pathConstrained parser, context-specific encoding, sandboxing, and no raw execution.

If local policy, contract, regulation, or a customer promise requires a stricter gate, use the stricter rule. Do not lower a gate to make a scorecard pass.

Evidence and citation review

For every answer that depends on a source, review the source path rather than only the final wording.

  • The source is in the approved source set for this route and tenant.
  • The reviewer can locate the source version or snapshot used for the answer.
  • The answer does not claim more than the source supports.
  • Conflicting, stale, missing, or inaccessible sources produce a clear limitation or handoff.
  • A citation or source reference is shown when the use case requires traceability.
  • Retrieved text is treated as untrusted data and cannot silently rewrite higher-priority controls.
  • Source changes have an owner, review trigger, and regression sample.

Do not turn the presence of a citation into a correctness guarantee. A citation can be present but irrelevant, stale, unauthorized, or misinterpreted.

Data and tenant boundary review

Review the question, context, retrieval result, answer, trace, and destination as one flow.

CheckPass conditionEscalate when
IdentityThe user and session are authenticated as required.Anonymous or shared identity is used for customer-specific data.
Tenant bindingTenant scope is enforced outside the model.A prompt or retrieved document is the only boundary.
Field minimizationOnly fields needed for the task are retrieved and displayed.Raw records, secrets, or unrelated fields appear.
Context separationTrusted instructions and untrusted content are distinguishable.A document, attachment, or user message can change policy.
Output exposureThe destination is approved for the data class.The answer can be copied, exported, or sent to a wider audience.
Trace accessReview evidence has restricted access and a retention rule.The review channel contains raw customer content or credentials.
Correction and deletionThe team can correct, delete, or restrict the answer and copies.The answer is durable but has no owner or cleanup path.

Record the boundary decision even when no private data appeared in the sample. That makes the scope explicit for later changes.

Tool and downstream safety

An answer review does not authorize an action. Review action requests separately.

  • The model output is parsed against a strict schema before any downstream use.
  • The destination has context-appropriate encoding or parameterization.
  • Tool names, arguments, target records, and recipients are allowlisted where appropriate.
  • Authorization is checked using the user and service identity outside the model.
  • High-impact actions require explicit human approval.
  • Retries, duplicate requests, and partial failures are safe or idempotent.
  • The system can pause, reject, or hand off when validation or authorization fails.
  • Tool calls and rejection reasons are logged without collecting unnecessary private prompts.

OWASP LLM05 warns that model output can reach browsers, databases, tools, and other systems with consequences such as XSS, SSRF, privilege escalation, or code execution when it is not validated and handled for its destination. Keep this section as a separate release gate instead of hiding it inside a quality average.

Reviewer calibration

Two reviewers should score a small shared set before a high-impact release or recurring review. Discuss disagreements and update the local rubric with examples.

Calibration stepRecord
Shared sample IDs
Dimensions with disagreement
Examples added to rubric
Final interpretation
Reviewer training or domain context
Calibration owner
Next calibration triggerNew model, new domain, repeated disagreement, or incident.

Do not force agreement by averaging away a safety concern. A reviewer who identifies a data, authorization, or high-impact decision failure should be able to block or escalate the sample.

Test set and evidence record

Maintain a small regression set for supported intents, known failures, boundaries, and safe refusal behavior. Keep it separate from a live customer transcript archive.

Test IDScenarioExpected behaviorActual resultOwnerEvidence
Supported ordinary requestCorrect, source-backed answer
Missing or conflicting sourceState uncertainty or hand off
Cross-tenant requestDeny and log the boundary failure
Sensitive-data requestMinimize, redact, or escalate
Prompt injection or untrusted documentKeep policy boundary and refuse unsafe instruction
Tool argument manipulationValidate, authorize, or block
Provider or retrieval failureUse safe fallback or human route
Correction or complaintPreserve a clear correction path

Store the route version, test-set version, date, reviewer, pass/fail result, and remediation link. Use synthetic or approved masked data in fixtures.

Release gate

Use these gates after the sample and regression set are reviewed.

GateRequired evidenceDecision if it fails
ScopeSupported intents, users, tenants, destinations, and exclusionsReduce scope or pause.
SamplingPopulation, strata, selection method, limitations, and sample IDsRe-sample before approval.
QualityRubric, calibration, score distribution, and representative examplesFix or limit affected intents.
EvidenceApproved source mapping, freshness, and uncertainty handlingBlock unsupported answers.
DataIdentity, tenant, field, retention, and export boundariesBlock delivery and investigate exposure.
ActionSchema, destination, authorization, approval, and idempotency checksDisable the action or require a human.
HandoffHuman owner, context package, response target, and pause pathRoute more cases to a human.
OperationsMonitoring, correction, incident, rollback, and next review triggerKeep limited or draft-only.

Do not approve based on a total score if a release-blocking dimension failed. Record the failed gate and the exact condition for reopening it.

Staged rollout plan

StageExposureRequired evidenceExit rule
0. Review onlySynthetic or redacted fixturesCompleted rubric, calibration, and regression resultsNo unresolved critical data or action failure.
1. ShadowObserve without customer-visible delivery or writesRepresentative sample and latency or capacity notesFindings are understood and owners are assigned.
2. LimitedSmall, approved user or tenant scopeLive monitoring, correction route, and human coverageNo release-blocking failure in the review window.
3. ExpandedBroader approved scopeRepeated sample, source check, data review, and rollback readinessExpansion remains within local risk tolerance.
4. OperateNormal approved pathRecurring sample, incident triggers, and review cadenceRe-review after a material change or signal.

Keep a rollback or pause switch available at each stage. A staged rollout is not a substitute for authorization, output handling, or incident response.

Findings and remediation

Group findings by failure mode rather than only by model or reviewer. One symptom can have a shared cause in a prompt, source, route, parser, permission check, or destination.

Finding IDFailure modeImpacted scopeRoot cause hypothesisControl or fixOwnerDue dateRetest
Unsupported claimSource, abstention, or reviewer fix
Cross-tenant or role leakageAuthorization or retrieval boundary fix
Unsafe destination or actionParser, encoder, allowlist, or approval gate
Missing uncertainty or handoffScope, prompt, UI, or routing fix
Stale or conflicting sourceSource owner and update workflow
Reviewer disagreementRubric example or calibration change

Close a finding only after the fix is deployed or tested in the affected path and the regression sample passes. Link the restricted evidence without copying private content into the public issue.

Decision table

Review outcomeDefault decisionFollow-up
All release gates pass and no blocker remainsApprove the defined scopeMonitor and schedule the next review.
Quality issue is narrow and customer impact is lowLimit the affected intent or routeFix, label, and re-test before expansion.
Evidence is missing or uncertainty is hiddenHuman review or abstentionAdd source, limit scope, and update the test set.
Data or tenant boundary failsBlock and investigateRestrict access, preserve evidence, and assess exposure.
Tool or downstream action fails authorizationDisable automatic actionRequire human approval and retest the gateway.
Review results are not representativeRe-sampleDocument the limitation and keep the route limited.
Repeated material failure or unresolved incidentPause or rollbackUse the incident and recovery owner path.

The decision applies only to the reviewed route, version, data scope, and time window. A new provider, source, action, audience, or material prompt change needs a new review.

Sign-off record

RoleName or teamDecisionDateEvidence or comment
Product or service owner
Engineering owner
Security or privacy reviewer
Domain reviewer
Support or human-handoff owner
Release or incident owner

Action tracker

ActionOwnerPriorityDue dateStatusEvidence link
Critical / High / Medium / LowOpen / Blocked / Done
Critical / High / Medium / LowOpen / Blocked / Done
Critical / High / Medium / LowOpen / Blocked / Done

Final human review checklist

  • The population, review window, strata, selection method, exclusions, and limitations are recorded.
  • Normal, peak, changed, uncertain, sensitive, boundary, negative, and escalated cases are covered as applicable.
  • Reviewer calibration used shared examples and disagreements are documented.
  • Each answer has independent scores for scope, evidence, uncertainty, data, safety, destination, action, and handoff.
  • Release-blocking dimensions cannot be hidden by an aggregate score.
  • Important claims map to approved, current sources or are clearly marked uncertain.
  • Missing evidence produces abstention, clarification, or human review rather than a guess.
  • Identity, tenant, role, field, retention, export, and trace boundaries were checked outside the model.
  • Model output is validated and encoded for its destination before downstream use.
  • Tool actions, writes, sends, deletes, refunds, access changes, and other durable effects have independent authorization.
  • High-impact cases have a named human owner and a usable handoff path.
  • Findings have owners, due dates, evidence, remediation, and regression tests.
  • Pause, rollback, correction, and incident paths are tested or explicitly limited.
  • The decision applies to a defined route, version, audience, and data scope.
  • The next review trigger is recorded.

Metrics to track

Track trends with definitions and denominators. Avoid publishing a single quality number without its sample design and scope.

MetricWhat it can revealRequired context
Evidence-supported answer rateSource coverage and citation qualitySource set, sample strata, and reviewer definition.
Unsupported or materially wrong rateCorrectness and misinformation signalsIntent mix, severity, and review window.
Abstention or clarification rateWhether uncertainty is surfacedSupported scope and expected unknowns.
Human handoff rateScope, quality, and staffing pressureHandoff reason and customer impact.
Data-boundary failure countTenant, role, and minimization problemsData class, route, and containment status.
Unsafe action or destination countOutput handling and authorization failuresAction type, blocked or completed state, and owner.
Correction recurrenceWhether fixes prevent repeat errorsFinding family, version, and regression result.
Reviewer disagreement rateRubric clarity and calibration needsShared sample, dimension, and reviewer pair.
Time to triage and closeOperational response capacitySeverity, owner, and evidence availability.
Review coverageWhether sample design matches the populationStrata, population, exclusions, and limits.

Use trends to choose the next review and remediation. Do not infer that a low complaint rate means low risk when detection and sampling are weak.

Evidence checked

  • OWASP LLM05:2025 Improper Output Handling describes the risk of passing insufficiently validated or sanitized model output to downstream components and recommends context-aware handling, monitoring, and a zero-trust approach to model output.
  • OWASP LLM09:2025 Misinformation describes credible but false or misleading output and points to validation mechanisms, human oversight, risk communication, and clear limitations.
  • NIST AI RMF Core describes continuous govern, map, measure, and manage functions; the Measure function includes testing before deployment and regularly during operation.
  • NIST AI RMF Manage Playbook includes continual monitoring, treatment procedures, and documenting and remediating risks from AI system components and third parties.
  • CISA and UK NCSC secure AI system development guidance emphasizes secure design, transparent ownership, and security across development, deployment, and operation.

FAQ

How many chatbot answers should a small team review?

There is no universal number that makes a review valid. Define the population, risk, strata, time window, and decision purpose first. Use more coverage for high-impact actions, sensitive data, new behavior, repeated incidents, or weak detection, and record the limits of a small or convenience sample.

Can a high average score approve the chatbot?

No. Averages can hide a small number of severe data, authorization, destination, or high-impact failures. Treat release-blocking dimensions as gates and record the worst relevant finding, not only an aggregate.

Should reviewers read raw customer conversations?

Usually not by default. Start with synthetic fixtures, aggregate metrics, and approved masked samples. If restricted evidence is necessary, limit access, minimize fields, record the purpose and retention rule, and keep the raw material out of broad collaboration channels.

Does a citation prove that an answer is correct?

No. The source can be stale, irrelevant, unauthorized, or misinterpreted. Reviewers should check the claim, source version, scope, conflicts, and whether the answer communicates uncertainty.

What should happen when the chatbot is unsure?

The route should clarify, abstain, provide a bounded limitation, or hand off to a named human path. Do not reward confident guessing. Define the fallback before launch and test it with missing, conflicting, and inaccessible source cases.

Does human review remove prompt injection or output-handling risk?

No. Human review is one control and may not happen before every answer. Validate and sanitize outputs, authorize actions outside the model, separate trusted instructions from untrusted content, and keep automatic paths bounded.

When does a new model or prompt need a new review?

When it can change the answer, source selection, data scope, action, destination, audience, or failure behavior. Re-run the affected regression set and update the sampling plan; do not rely on the previous sign-off.

What is the safest default for a failed high-risk sample?

Block the affected automatic path, preserve restricted evidence, notify the named owner, and route to a human or static fallback while the finding is investigated. Reopen only after the fix and regression test pass.

Run the AI Tool Risk Checker for the chatbot route, attach the result to the review intake, and complete a small calibrated sample before expanding customer scope. Use the Small Team AI Security Checklist to verify the identity, data, access, logging, and incident foundations around the review process.