awesomechecklists
llms.txtGitHub ↗
/

14 — AI Human Review and Escalation Checklist#

Ensure human involvement in AI-assisted decisions is meaningful, authorized, timely, and auditable. A review button, nominal human sign-off, or the ability to view model output does not establish independent oversight or prevent unauthorized downstream execution.

Scope: AI-specific. This file is one part of the 14_AI_Human_Review_and_Escalation.md checklist. Apply the domain-wide release gate across all applicable scope files; each checklist ID remains unique.

Review details#

  • Project / product:
  • Release / version / environment:
  • Scope and applicable jurisdictions / markets:
  • Human review use cases, affected persons, and consequential actions:
  • Reviewer roles, delegation, approval policy, and escalation owners:
  • Queue targets, safety thresholds, and review system versions:
  • Reviewer / owner:
  • Review date:
  • Decision: Pending / Approved / Approved with documented exceptions / Blocked

How to use#

  1. Copy this checklist into project-specific review records and leave all repository template boxes unchecked.
  2. Assess whether a human can genuinely understand, challenge, reject, correct, override, and where necessary halt the AI-assisted result. Determine oversight requirements based on use case, harm, autonomy, actual applicable rules, and the authority of the decision maker.
  3. Mark [x] only after verifying reviewer training and permissions, representative review cases, queue behavior, safe approval/rejection/resume tests, appeal flows, incident records, or qualified policy decisions. Keep evidence, defects, accepted exceptions, and justified N/A outside this shared template, keyed by HUMAN ID.
  4. Separate ## API Checks and ## Frontend Checks; document why either may be inapplicable to headless, advisory-only, or offline deployments.
  5. Human approval is not an authorization boundary by itself: independently validate the initiating user's permissions, reviewer's role, exact action, and execution constraints. Do not treat approval as proof an external side effect completed exactly once.
  6. Reassess oversight after changes to AI tools, policies, reviewer staffing, prompt/model versions, automated thresholds, user populations, or supported markets. Coordinate the general governance, agent, safety, and evaluation requirements in those dedicated checklist files.

Human oversight design#

  • HUMAN-001 Define which AI-assisted decisions, recommendations, content releases, and external actions require human review based on severity and user impact.
  • HUMAN-002 Assign accountable owners for oversight policy, day-to-day review operations, reviewer competence, and incident escalation.
  • HUMAN-003 Distinguish human-in-the-loop approval before an action, human-on-the-loop supervision, and post-decision review in the documented workflow.
  • HUMAN-004 Identify consequential decisions where a qualified person must retain effective authority rather than merely rubber-stamping model output.
  • HUMAN-005 Specify which decisions can be fully automated and record the rationale, permitted use, and any legally required safeguards.
  • HUMAN-006 Define independent reviewer authority to reject, amend, postpone, reverse, or stop an AI recommendation or proposed action.
  • HUMAN-007 Require reviewers to see the real decision context, input evidence, known limitations, and consequences appropriate to their permissions.
  • HUMAN-008 Present model confidence or scores only with tested calibration and explanations of what the metrics actually measure.
  • HUMAN-009 Provide a clear distinction between AI-generated suggestions, verified source facts, reviewer judgments, and completed external actions.
  • HUMAN-010 Identify the minimum domain, language, cultural, accessibility, and risk expertise needed for each review category.
  • HUMAN-011 Establish reviewer training on automation bias, erroneous AI explanations, uncertainty, prompt injection, and escalation duties.
  • HUMAN-012 Define separation of duties for creating a model output, reviewing it, and authorizing high-impact execution when justified by risk.
  • HUMAN-013 Provide sufficient review time and staffing so reviewers can exercise judgment instead of being forced into automatic approvals.
  • HUMAN-014 Specify when reviewers should consult a second specialist or independent source before making consequential decisions.
  • HUMAN-015 Document how impacted people can request explanation, correction, or meaningful human intervention where applicable.
  • HUMAN-016 Review privacy, employment, sectoral, and automated-decision requirements for the actual jurisdiction and use case with qualified counsel.
  • HUMAN-017 Define the extent of human discretion and confirm that organizational procedures do not undermine genuine ability to override the AI.
  • HUMAN-018 Identify tasks where reviewers need access to the original document, image, audio clip, transcript, or tool result rather than a model summary.
  • HUMAN-019 Require deterministic authorization checks for every action independently of a reviewer's click and the model's proposed reasoning.
  • HUMAN-020 Specify safe behavior if the reviewer lacks permission, relevant expertise, or enough evidence to decide.
  • HUMAN-021 Design accessible review surfaces with clear status, understandable controls, and explicit confirmation for irreversible decisions.
  • HUMAN-022 Reassess oversight roles and decision authority whenever models, workflows, tool scopes, user populations, or risk classifications change.

Review queues and decision thresholds#

  • HUMAN-023 Inventory all review-triggering conditions, including risk severity, low confidence, policy flags, disagreement, and requested user appeals.
  • HUMAN-024 Define separate handling for routine validation, high-impact actions, incident triage, safety escalation, and regulatory or customer disputes.
  • HUMAN-025 Document task-appropriate thresholds for automatic approval, manual review, abstention, and hard blocking before production use.
  • HUMAN-026 Confirm numerical scores used for routing are calibrated, versioned, and meaningful for the reviewed task.
  • HUMAN-027 Define mandatory human review conditions that cannot be waived by an aggregate model score or low queue volume.
  • HUMAN-028 Specify what happens when classification, routing, or risk-scoring components fail or return malformed results.
  • HUMAN-029 Prioritize review queues by potential harm, urgency, case age, and any applicable service or legal deadlines.
  • HUMAN-030 Monitor queue volume, age, backlog, processing time, escalation rate, and cases that exceed agreed service levels.
  • HUMAN-031 Forecast capacity under expected peak demand and major policy, model, or provider changes.
  • HUMAN-032 Prevent queue overload from silently approving or dropping high-impact cases; use defined safe degradation or pauses.
  • HUMAN-033 Ensure work distribution respects reviewer expertise, working hours, availability, language, and conflicts of interest.
  • HUMAN-034 Use clear ownership and atomic claim/lease behavior to prevent two reviewers from making conflicting decisions on one case.
  • HUMAN-035 Define expiration and reassignment rules for stalled, abandoned, duplicate, or unclaimed review tasks.
  • HUMAN-036 Handle requests with missing source evidence or incomplete model outputs through clarification or safe rejection rather than a forced binary verdict.
  • HUMAN-037 Distinguish recommendation confidence from probability of actual harm and avoid conflating them in risk routing.
  • HUMAN-038 Test thresholds with representative benign cases to measure false alarms and avoid unnecessary obstruction of legitimate users.
  • HUMAN-039 Test threshold sensitivity with high-impact adverse cases and measure failures that incorrectly bypass review.
  • HUMAN-040 Measure queue and approval outcomes by relevant language, task category, access level, and justified population segment where lawful.
  • HUMAN-041 Review whether reviewer workload or incentive structures create systematic pressure toward approval.
  • HUMAN-042 Record threshold owners, revisions, supporting evaluations, exception criteria, and effective dates.
  • HUMAN-043 Use named backup reviewers and escalation chains to cover absence, holidays, incident surges, and specialist unavailability.
  • HUMAN-044 Provide a way to escalate an individual case immediately regardless of its model-assigned risk score.
  • HUMAN-045 Test dynamic reassignment, queue replay, partial outages, and resumed sessions for lost or duplicated review requests.
  • HUMAN-046 Reassess automated routing and review thresholds after material evaluation regressions, incident findings, or distribution shifts.

Escalation and overrides#

  • HUMAN-047 Define severity levels and routing contacts for safety, privacy, security, operational, legal, and domain-specific escalation.
  • HUMAN-048 Document when first-line reviewers must refer a case to an independent specialist or authorized higher-level decision maker.
  • HUMAN-049 Give reviewers an explicit ability to reject an AI recommendation and record a corrected outcome or required human action.
  • HUMAN-050 Test that a rejection prevents execution of the corresponding pending tool call or action, including through nested agents.
  • HUMAN-051 Bind every approval to the authenticated reviewer, actual run, target tool, normalized parameters, and specific proposed action.
  • HUMAN-052 Expire approvals and require renewed review when material parameters, target resources, permissions, or the action itself change.
  • HUMAN-053 Reject altered, forged, stale, duplicated, or replayed decisions and do not trust client-supplied serialized approval state.
  • HUMAN-054 Recheck the reviewer's authority and the original user's current permissions immediately before execution.
  • HUMAN-055 Prohibit broad standing approvals for high-impact actions unless an independently approved and narrowly scoped exception policy explicitly permits them.
  • HUMAN-056 Assess whether resumed execution after a handoff changes the approval scope or introduces additional unreviewed operations.
  • HUMAN-057 Require additional confirmation or dual control for irreversible or high-risk actions when the task or applicable rules call for it.
  • HUMAN-058 Design a safe route when reviewers are unavailable, approval service fails, or risk classification is uncertain.
  • HUMAN-059 Allow reviewers and affected users to cancel a pending action and verify cancellation propagates to queued work.
  • HUMAN-060 Document which external side effects can be reversed, compensated, or only mitigated after approval.
  • HUMAN-061 Use operation identifiers, authoritative status checks, and duplicate safeguards when approved actions resume after crashes or retries.
  • HUMAN-062 Distinguish approval of a tool call from evidence that the tool actually completed its external side effect.
  • HUMAN-063 Provide an emergency halt or suspension capability for confirmed AI harm or unauthorized activity.
  • HUMAN-064 Test overrides, appeals, and emergency stops during an active model rollout, provider outage, or degraded monitoring.
  • HUMAN-065 Define a transparent review/appeal process for contested outcomes with accessible submission and case-status tracking where appropriate.
  • HUMAN-066 Ensure legal or contractual notification and redress obligations are evaluated for incidents and significant automated decisions.
  • HUMAN-067 Record reasoned override decisions without requiring reviewers to endorse unsupported model explanations.
  • HUMAN-068 Escalate repeated low-quality decisions, model disagreement, unsafe tool proposals, and queue anomalies to responsible model and risk owners.

Auditability and reviewer quality#

  • HUMAN-069 Log a stable case identifier, input version, model/prompt/policy version, review task, actor, and relevant event timestamps.
  • HUMAN-070 Record the proposal, decision category, approving or rejecting role, rationale, and actual execution outcome at an appropriate level of detail.
  • HUMAN-071 Keep required audit records in controlled storage protected against unauthorized editing and deletion.
  • HUMAN-072 Ensure logging and review exports minimize personal information, model secrets, and sensitive source materials.
  • HUMAN-073 Restrict review histories and appeal records by user, tenant, case, and reviewer authorization.
  • HUMAN-074 Define audit retention, access, deletion, legal hold, and geographic processing requirements by applicable policy and law.
  • HUMAN-075 Version review policies, threshold rules, reviewer instructions, and decision rubrics for reproducible incident analysis.
  • HUMAN-076 Check that logs distinguish model-generated recommendations from human judgments and confirmed tool effects.
  • HUMAN-077 Test reconstructing a disputed case from approved evidence without relying on undocumented ephemeral client state.
  • HUMAN-078 Measure reviewer error, disagreement, override frequency, false approvals, and false rejections on representative sampled cases.
  • HUMAN-079 Calibrate reviewers with qualified example cases and objective or appropriately adjudicated reference judgments.
  • HUMAN-080 Review cases with unusual decision speed, repeated approvals, conflicting outcomes, or unexpectedly low escalation rates.
  • HUMAN-081 Assess differences in outcomes across reviewer teams, language groups, and justified user segments with privacy and sample-size safeguards.
  • HUMAN-082 Provide a second review or structured adjudication process for high-severity disagreements and possible reviewer conflicts.
  • HUMAN-083 Use quality findings to improve guidelines, workload, tooling, and training rather than hiding exceptions in aggregate rates.
  • HUMAN-084 Track recurring escalations and user appeals to identify AI failure patterns that deserve model or workflow remediation.
  • HUMAN-085 Review access logs for unusual bulk case browsing, unauthorized exports, or attempts to manipulate approval records.
  • HUMAN-086 Test that monitoring alarms and incident workflows still function when a review service, queue, or audit sink is unavailable.
  • HUMAN-087 Define independent audits or spot checks proportionate to the harm and regulatory exposure of the application.
  • HUMAN-088 Record reviewer training completion, delegated authority, and requalification after significant policy or tooling changes.
  • HUMAN-089 Make approval and review reporting understandable to qualified internal oversight staff without disclosing restricted records.
  • HUMAN-090 Document corrective actions with owners, deadlines, retraining or policy changes, and subsequent retest evidence.
  • HUMAN-091 Review whether humans have enough time, context, and organizational authority to exercise meaningful independent judgment.
  • HUMAN-092 Reassess quality and fairness when the model, client interface, evidence sources, user population, or review team changes.

Release gate#

  • Documented decision categories identify which AI outcomes require human intervention, specialist review, dual control, or immediate suspension according to impact and applicable law.
  • Qualified reviewers have real access to the necessary evidence, time, training, authority, and accessible controls to approve, reject, change, escalate, or stop a decision.
  • Review queues, prioritization, assignment, decision thresholds, staffing, expirations, and safe backlog/outage behavior have been tested under representative workload.
  • Approval, denial, and override controls are bound to authenticated, authorized reviewers and the exact action parameters; stale, replayed, forged, and changed requests cannot execute.
  • Significant decisions, reviewer disagreements, appeals, overrides, completed actions, and incident escalation are traceable with appropriate privacy, integrity, and retention safeguards.
  • Human-review effectiveness and potential automation bias have been evaluated with representative benign, ambiguous, and high-impact cases, including false approvals and unjustified rejections.
  • API and frontend pathways accurately distinguish proposals from confirmed effects and handle waiting, denial, timeout, failed resume, cancellation, and correction accessibly.
  • A named risk owner has reviewed residual limitations, complaint/redress obligations where applicable, response runbooks, emergency halt mechanisms, and the exact release configuration.

Domain-specific approval rule: Block release if human review is legally or risk-required but the reviewer cannot independently challenge or stop outcomes; if unapproved, altered, or replayed actions can execute; if high-impact cases can be silently auto-approved during overload; or if mandatory records and escalation routes are unavailable. Human oversight must be proportionate and effective, not merely ceremonial.

Trusted references and scope#

The primary standards, laws, risk-management guidance, and implementation references below inform the checklist. Apply vendor features and legal requirements only where relevant to the particular application and deployment.

Applicability: The NIST AI RMF is voluntary. The EU AI Act imposes Article 14 oversight design requirements and Article 26 deployer obligations for applicable high-risk AI systems; its implementation and application dates require current, qualified legal review and are not identical for all AI products. GDPR Article 22 applies in specific circumstances involving solely automated decisions with legal or similarly significant effects, including exceptions and safeguards; not every AI suggestion triggers that article. OpenAI Agents SDK and Microsoft Agent Framework approval facilities have different behavior for reused/standing approvals, nested agents, saved state, and resumed tool calls. Their native controls must be supplemented with trusted application authorization, concurrency controls, replay prevention, and recovery of external side effects. WCAG 2.2 requirements depend on the applicable conformance target and governing accessibility law.

Source review date: 2026-11-12.

Suggested testing tools#

Choose tools appropriate to the project's stack and risks; automation supports evidence but does not approve the checklist.

  • Test focus: review queues, approval boundaries, appeals, escalations, and audit trails; include model and dataset behaviors where applicable.
  • Cross-stack: Playwright.
  • Python: pytest; for pytest-based suites, run python -m pytest -q.
  • TypeScript: Playwright + Vitest; for Vitest-based suites, run npx vitest run (or npx playwright test for E2E).
  • JavaScript: Playwright + Jest; for Jest-based suites, run npx jest --ci (or npx playwright test for E2E).
  • Evidence: Save tool versions, configuration, test environment, fixtures, results, and links in project-specific review records using the existing checklist IDs. Expert review remains necessary where relevant.

Suggested AI skills for verification#

These are optional agent review aids, not automated passes or release approval. Each linked repository had at least 1,000 GitHub stars when checked on 2026-11-12; counts are per repository, not per skill, and do not indicate security or correctness.

  • agent-governance — Review human oversight and escalation rules (github/awesome-copilot, 41,073 repo stars).
  • signed-audit-trails-recipe — Check review-event and override auditability (wshobson/agents, 42,437 repo stars).
  • audit-integrity — Assess approval and audit-history integrity (github/awesome-copilot, 41,073 repo stars).

How to check: In a project-specific copy, have the agent audit the AI-specific scope and map findings to this file's existing checklist IDs, with code/configuration references, reproducible test results, failed or inapplicable cases, and uncertainties. Independently verify findings; keep shared template boxes unchecked and retain human approval requirements.

Before use: Read the skill's SKILL.md, referenced scripts, license, and requested tool permissions. Only execute vetted skills with appropriate access in a controlled environment.

Important warning: reference only#

⚠️ Reference only — not a definitive or exhaustive checklist. This template is a starting point for review, not the ultimate authority or a guarantee of project safety, quality, regulatory compliance, accessibility, security, or launch readiness.

Use project-specific judgment and independent verification. Requirements and risks evolve and differ by project, technology, and jurisdiction. Adapt the checks, consult current authoritative sources, perform relevant tests, and obtain qualified specialist review or approval when appropriate. Checking every box does not replace these steps.

Reference, not a certification. Check applicability, verify important claims against current primary sources, and seek expert review where necessary. Suggest an improvement.