Bias / Consistency Review of HR Artifacts
Review HR artifacts for biased language and inconsistent standards, then produce redline corrections and a bias audit report.
Before you start
What you’ll need
- Access to approved artifacts from the policy handbook update, performance review, promotion calibration, and compensation review workflows
- Familiarity with your organisation’s bias review criteria and governance standards
- Understanding of cross-functional HR artifact dependencies and approval workflows
- Bias Review Criteria
- Consistency Standards Checklist
Who’s involved
- HR Governance Analyst — Assembles upstream artifacts, runs the bias and consistency review workflow, and coordinates corrections through to approval.
- Diversity Reviewer — Validates flagged bias findings against organisational equity standards and confirms correction appropriateness.
- HR Director — Approves the final bias audit report and authorises redlined corrections for implementation.
Safe use
- Data Handling — This workflow processes performance review narratives, promotion outcomes, compensation rationale, and policy drafts containing sensitive employee and organisational data. Do not paste these inputs into public or unapproved GenAI tools.
- Verification — GenAI may over-flag neutral language as biased or miss subtle bias patterns. Have a governance reviewer verify every flagged finding and proposed correction against the source artifact context before finalising.
Execution steps
Inputs
- Policy Draft RedlinesdownloadFrom HR10 — Policy Update
- Performance Review NarrativesdownloadFrom HR11 — Performance Review
- Promotion Panel OutcomesdownloadFrom HR13 — Promotion Calibration
- Compensation Recommendation RationaledownloadFrom HR16 — Comp Review
- Bias Review Criteriadownload
- Consistency Standards Checklistdownload
- Confirm the policy draft redlines from the handbook update workflow are available and approved
- Confirm performance review narratives from the performance review workflow are available
- Confirm promotion panel outcomes from the calibration workflow are available and finalised
- Confirm compensation recommendation rationale from the compensation review workflow is available
- Confirm your organisation’s bias review criteria document is current and accessible
- Confirm the consistency standards checklist covers all four artifact types
- Confirm the bias review criteria document defines the bias categories to scan for
Prompt
Scan HR artifacts for biased or exclusionary language
CONTEXT You will be provided with the following source documents: 1. Policy Draft Redlines 2. Performance Review Narratives 3. Promotion Panel Outcomes 4. Compensation Recommendation Rationale 5. Bias Review Criteria 6. Consistency Standards Checklist TASK Scan each artifact for biased language, including gendered wording, subjective qualifiers without evidence, culturally loaded phrases, and exclusionary terminology. For each finding, identify the source artifact, the specific passage, the bias type, and a brief explanation of why the language is problematic. OUTPUT FORMAT Return a markdown table with the following columns: | # | Source Artifact | Passage | Bias Type | Explanation | |---|---|---|---|---| Bias Type must be one of: Gendered Language, Subjective Qualifier, Cultural Bias, Exclusionary Term, Vague Justification. After the table, include a section titled "Prevalent Patterns" with a one-paragraph summary of the most common bias patterns across all artifacts. CONSTRAINTS Do not suggest replacement language in this step — focus on identification only. Do not flag language that is factual and evidence-backed merely because it uses strong terms. Only flag issues supported by the specific passage text provided.
Outputs
- Confirm every flagged finding references a specific passage from one of the four source artifacts
- Verify each bias type classification aligns with the definitions in the bias review criteria
Verification: Verify the AI did not over-flag neutral professional language as biased or miss context-dependent bias patterns that require domain understanding.
Inputs
- Policy Draft RedlinesdownloadFrom HR10 — Policy Update
- Performance Review NarrativesdownloadFrom HR11 — Performance Review
- Promotion Panel OutcomesdownloadFrom HR13 — Promotion Calibration
- Compensation Recommendation RationaledownloadFrom HR16 — Comp Review
- Bias Review Criteriadownload
- Consistency Standards Checklistdownload
- Language Bias Scan Results
- Confirm the language bias scan results have been reviewed for completeness
Prompt
Identify inconsistent standards across HR artifacts
CONTEXT You will be provided with the following source documents: 1. Policy Draft Redlines 2. Performance Review Narratives 3. Promotion Panel Outcomes 4. Compensation Recommendation Rationale 5. Bias Review Criteria 6. Consistency Standards Checklist 7. Language Bias Scan Results TASK Compare the standards, criteria, and justification patterns used across the four artifact types. Identify where the same role, decision category, or evaluation criterion is described using inconsistent standards, conflicting justifications, or mismatched language. For each inconsistency, state the two conflicting artifacts, the discrepancy, and its potential impact on equitable decision-making. OUTPUT FORMAT Return a markdown table: | # | Artifact A | Artifact B | Discrepancy | Potential Impact | |---|---|---|---|---| After the table, include a section titled "Cross-Cutting Patterns" with 2–4 bullet points summarising the systemic consistency issues found across artifacts. CONSTRAINTS Do not evaluate the correctness of any individual artifact’s conclusion — focus on cross-artifact alignment only. Do not fabricate connections between artifacts where the source data does not overlap. Only flag discrepancies supported by direct evidence in both artifacts.
Outputs
- Consistency Assessment Report
- Confirm each discrepancy cites specific evidence from both referenced artifacts
- Verify the cross-cutting patterns reflect themes from the findings table, not new unsupported claims
Verification: Verify the AI did not fabricate cross-artifact links where the source documents address unrelated topics or roles.
Inputs
- Language Bias Scan Results
- Consistency Assessment Report
- Policy Draft RedlinesdownloadFrom HR10 — Policy Update
- Performance Review NarrativesdownloadFrom HR11 — Performance Review
- Promotion Panel OutcomesdownloadFrom HR13 — Promotion Calibration
- Compensation Recommendation RationaledownloadFrom HR16 — Comp Review
- Bias Review Criteriadownload
- Consistency Standards Checklistdownload
- Confirm the language bias scan and consistency assessment have both been reviewed and accepted
Data Handling: Do not paste raw employee performance ratings or individual compensation figures into the prompt; use the summary-level findings from the scan and assessment only.
Prompt
Draft redlined corrections for bias and consistency issues
CONTEXT You will be provided with the following source documents: 1. Language Bias Scan Results 2. Consistency Assessment Report 3. Policy Draft Redlines 4. Performance Review Narratives 5. Promotion Panel Outcomes 6. Compensation Recommendation Rationale 7. Bias Review Criteria 8. Consistency Standards Checklist TASK For each flagged bias issue and consistency discrepancy, draft a redlined correction of the affected passage. Present the original text alongside the proposed revision and include a one-sentence rationale linking the change to the specific finding. OUTPUT FORMAT For each correction, use this structure: ### [Source Artifact] — [Section or Passage Reference] **Finding Reference:** [Bias Scan #X or Consistency #X] **Original:** [existing text] **Revised:** [corrected text] **Rationale:** [one sentence explaining the change and its source finding] EXAMPLE ### Performance Review Narrative — Section 3: Leadership Assessment **Finding Reference:** Bias Scan #4 **Original:** She demonstrates a nurturing leadership style that supports team morale. **Revised:** This employee demonstrates an inclusive leadership approach that measurably supports team engagement and retention. **Rationale:** The original language used a gendered qualifier that is disproportionately applied to certain demographics; the revision uses evidence-anchored terms (Bias Scan #4). CONSTRAINTS Do not introduce changes beyond those identified in the bias scan or consistency assessment. Do not alter factual conclusions or performance ratings — only correct biased or inconsistent language. Do not reference specific organisations, employee names, or proprietary frameworks.
Outputs
- Draft Redlined Corrections
- Confirm every redline maps to a specific finding from the bias scan or consistency assessment
- Verify no correction alters a factual conclusion, performance rating, or compensation decision
- Check that rationale statements reference the specific finding number
Verification: Verify the AI did not alter factual performance conclusions or introduce new language that shifts the meaning of the original assessment beyond the identified bias or inconsistency.
Inputs
- Draft Redlined Corrections
- Language Bias Scan Results
- Consistency Assessment Report
- Confirm all draft redlined corrections are complete and reference their source findings
Prompt
Classify severity and validate redline corrections
CONTEXT You will be provided with draft redlined corrections, the language bias scan results, and the consistency assessment report. TASK Review each draft redline correction and assign a severity classification. Flag any corrections that may be false positives or that alter the factual meaning of the original artifact. For each item, state whether the correction should be accepted, modified, or rejected with a brief justification. OUTPUT FORMAT Return a markdown table: | # | Finding Ref | Severity | Recommendation | Justification | |---|---|---|---|---| Severity must be one of: Critical, Major, Minor, Informational. Recommendation must be one of: Accept, Modify, Reject. After the table, include a summary section with total counts by severity and by recommendation. CONSTRAINTS Do not resolve disagreements between the bias scan and consistency assessment — flag them for human review. Do not approve corrections that change factual content or performance conclusions. Frame all recommendations as inputs for the human reviewer, not final decisions.
Outputs
- Validated Redline Package
- Confirm every correction has a severity classification and recommendation
- Verify rejected items include a clear justification referencing the source artifact context
Inputs
- Validated Redline Package
- Consistency Assessment Report
- Language Bias Scan Results
- Confirm the validated redline package reflects all human reviewer decisions from the previous step
Prompt
Consolidate findings into structured bias audit report
CONTEXT You will be provided with a validated redline package containing reviewed corrections with severity classifications, the consistency assessment report, and the language bias scan results. TASK Compile all findings into a structured bias audit report. The report should summarise the scope of review, key findings by category, severity distribution, recommended corrections, and systemic patterns requiring broader governance attention. OUTPUT FORMAT Structure the report with these sections: **1. Audit Scope** List the four artifact types reviewed and the total number of items scanned. **2. Findings Summary** A markdown table: | Category | Critical | Major | Minor | Informational | Total | |---|---|---|---|---|---| **3. Key Findings** For each critical and major finding, provide: - **Finding:** [one sentence] - **Artifacts Affected:** [list] - **Recommended Action:** [one sentence] **4. Systemic Patterns** 2–4 bullet points describing cross-cutting governance concerns that extend beyond individual corrections. **5. Recommendations** A numbered list of 3–5 actionable recommendations for governance improvement. CONSTRAINTS Do not include personally identifiable information or specific employee data. Do not provide definitive legal or compliance conclusions — frame recommendations as inputs for governance review. Do not repeat the full redline text; reference finding numbers only.
Outputs
- Bias Audit Report
- Confirm the findings summary table totals match the validated redline package counts
- Verify systemic patterns are supported by multiple findings, not single instances
- Check that recommendations are actionable and specific to the identified patterns
Verification: Verify the AI did not inflate severity counts or fabricate systemic patterns not supported by the underlying findings data.
Inputs
- Bias Audit Report
- Validated Redline Package
- Confirm the bias audit report has been reviewed by the diversity reviewer
- Confirm all critical and major findings have been addressed or formally escalated
Outputs
- Confirm all critical findings have accepted corrections or documented escalation rationale
- Verify the final redlined corrections do not alter factual conclusions from the source workflows
- Check that the approved audit report is complete and suitable for governance records
Reference
Guardrails
- Evidence-Based Flagging Only — Only flag bias or inconsistency issues supported by specific passages in the source artifacts; do not rely on AI inference or assumed organisational context.
- Cross-Artifact Comparison Required — Review all four artifact types together to identify cross-cutting inconsistencies; do not assess each artifact in isolation.
- Scan Before Correction — Complete the full bias scan and consistency assessment before drafting any redline corrections to prevent premature edits that miss systemic patterns.
Pitfalls
- Accepting AI bias flags at face value without verifying them against the original artifact language and surrounding context
- Pasting full employee performance narratives or individual compensation figures into the prompt instead of using summary-level inputs
- Allowing the AI to apply a single bias definition uniformly across all artifact types without accounting for differences in document purpose and audience
- Skipping the consistency assessment and proceeding directly to redlines based on the language bias scan alone
Definition of Done
- Every flagged bias or inconsistency issue maps to a specific passage in one of the four source artifacts with a cited finding reference
- The consistency assessment covers all four artifact types and identifies cross-artifact patterns, not just within-document issues
- The bias audit report includes a severity classification for every flagged item with supporting evidence from the validated redline package
- No personally identifiable information or specific employee data appears in any generated artifact
GET FULL ACCESS
Unlock every step, prompt, and downloadable example — for this playbook and the rest of AGASI AiOS, our GenAI capability framework.
We'll send a magic link — no password needed.
AGASI AiOS · HR18 v1.0 · Apr 8, 2026