# AI Human Oversight Checklist and Review Decision Template Designing oversight that is real: a workflow design checklist, a reviewer decision template that captures reasoning on agreement as well as override, a reviewer qualification standard by decision type, and a spec for override-rate monitoring. **Who this is for:** The system owner designing or rebuilding the human review step for a high-risk AI decision. Source playbook: https://aigovernance.com/playbook/human-oversight-for-high-risk-ai-decisions --- ## Human oversight workflow design checklist _The conditions that make review meaningful rather than a rubber stamp._ ### Template | Condition | In place? | Note | |---|---|---| | Reviewer sees the AI recommendation AND the key factors behind it | Y / N | | | Reviewer sees at least one alternative or the counterfactual | Y / N | | | Reviewer has enough time allocated per case for genuine judgment | Y / N | target seconds/minutes | | Reviewer can access the underlying record, not just the AI summary | Y / N | | | Reviewer has authority to override without escalation for routine cases | Y / N | | | The default action is not "accept" on timeout | Y / N | | | Reviewer is shielded from productivity metrics that punish overrides | Y / N | | | Reviewer documents reasoning for every decision, agree or override | Y / N | | ### Worked example | Condition | In place? | Note | |---|---|---| | Sees recommendation + key factors | Y | top-5 features shown | | Sees an alternative / counterfactual | N | design gap: add "what would change this score" | | Adequate time per case | N | 45s average; target 3 min for High tier | | Access to underlying record | Y | one click to the full application | | Authority to override routine cases | Y | | | No accept-on-timeout | N | current UI auto-advances after 60s | | Shielded from override-punishing metrics | N | handle-time target discourages scrutiny | | Reasoning documented every time | N | only overrides are logged today | ### Acceptance criteria - Every condition is assessed against the live workflow, with gaps listed. - Time-per-case is a real number measured from logs, compared to a target. - The workflow does not treat inaction or timeout as approval. --- ## Reviewer decision documentation template _Captured on every case. Reasoning on agreement matters as much as on override._ ### Template | Field | Entry | |---|---| | Case / decision ID | | | AI recommendation and score | | | Key factors shown to the reviewer | | | Reviewer decision (accept / override / modify) | | | Reviewer rationale (required on every decision) | | | Information the reviewer used beyond the AI summary | | | Time spent | | | Reviewer ID and qualification level | | ### Worked example | Field | Entry | |---|---| | Case ID | dec-2026-0091823 | | AI recommendation | score 2/5, "advance: no" | | Key factors shown | low years_experience; medium skill_match | | Reviewer decision | Override to "advance" | | Reviewer rationale | Candidate's portfolio shows shipped work the model cannot see; skill match understated | | Info used beyond summary | opened full application and portfolio link | | Time spent | 4 min | | Reviewer ID / level | u-4471, certified recruiter (level 2) | ### Acceptance criteria - A rationale is required and recorded for accept decisions, not only overrides. - The record captures what the reviewer looked at beyond the AI output. - Records are retained for the period the underlying decision type requires. --- ## Reviewer qualification requirements _What a person must know and hold to review each decision type._ ### Template | Decision type | Minimum role / certification | Domain knowledge required | Training on this AI system | Recertification | |---|---|---|---|---| | | | | | annual / biennial | ### Worked example | Decision type | Minimum role | Domain knowledge | System training | Recertification | |---|---|---|---|---| | Hiring screen override | Certified recruiter L2 | Role requirements; adverse-impact basics | 2h module on the screener's factors and limits | annual | | Fraud hold release | Fraud analyst L3 | Payment fraud patterns; false-positive cost | 3h module + shadowing | annual | | Benefits eligibility | Caseworker + supervisor sign-off for denials | Program rules; appeals process | 2h module | biennial | ### Acceptance criteria - Each decision type names a concrete qualification, not "trained staff". - Reviewers have completed system-specific training on the AI's factors and known limits. - Qualifications are tracked with expiry dates and recertification. --- ## Override-rate monitoring spec _What to measure so a near-zero override rate gets investigated._ ### Template | Metric | Definition | Expected range | Investigate when | |---|---|---|---| | Overall override rate | overrides / total reviews | | outside range for 2 weeks | | Override rate by reviewer | per reviewer | within X points of team median | near zero, or 3x median | | Time-per-review distribution | median and 10th percentile | 10th pct above a floor | 10th pct near zero | | Agreement on divergent cases | rate reviewers accept AI when it conflicts with the record | context-specific | trending toward 100% | | Post-decision reversal rate | decisions later overturned on appeal | low | rising, especially for one reviewer | ### Worked example | Metric | Value this month | Range | Status | |---|---|---|---| | Overall override rate | 4% | 8-18% (from calibration) | Investigate: low | | Lowest reviewer override rate | 0.6% (u-3390) | within 6 pts of 11% median | Investigate | | 10th-percentile review time | 12s | above 60s floor | Investigate: too fast | | Appeal reversal rate | 2.1% | under 3% | OK | Finding: the low override rate tracks with the 60s auto-advance and the handle-time target. Both flagged to the workflow redesign. ### Acceptance criteria - Expected ranges come from a calibration exercise, not a guess. - Per-reviewer rates are monitored, and near-zero rates trigger a review of that reviewer's workload and tooling. - The dashboard is actually watched, with a named owner and a cadence. --- ## Governance controls this kit produces evidence for - **HOC-002**: The workflow checklist and decision template are the design of the human approval gate for consequential decisions. - **HOC-003**: The decision documentation template is the output review workflow record. - **HOC-005**: The qualification requirements artifact is the reviewer competency standard. - **HOC-004**: The override-rate spec is the automation-bias detection mechanism. - **ALC-001**: Per-case reviewer decisions and rationale extend decision logging to the human step.