Template 04 — Responsible-AI Checklist

Purpose

The 10-item checklist that every Medium- and High-risk agent passes before Pilot. This is where ethical principles become operational checks. Low-tier agents may skip with CoE Lead discretion.

This is paired with the Security review (template 05) — the two together produce the M10 sign-offs in the In Action roadmap.

• When you use it: At M10 in the roadmap, after Agent Card written and before Step 11 Pilot. Re-run at every material scope change.
• Who fills it: CoE Lead (asks the questions, records answers). Builder provides evidence per item. Legal signs.
• Time to complete: 60–120 minutes total — including evidence-gathering.
• Output: Signed checklist attached to the registry entry.

Worked example (AP Accountant invoice reconciliation)

"Reviewed 2026-05-02 by Morteza Moradi (CoE Lead) + John Smith (General Counsel). Agent is Medium tier."

Agent: finance-invoice-recon v1.0

Sign-off

Conditional approvals / open items

None. All 10 items pass or are documented as N/A with reasons. Cleared to proceed to template 05 (Threat Model) → Step 11 (Pilot).

Blank template (copy below for your agent)

# Responsible-AI Checklist — [Agent Name]

**Agent ID:** [agent-dept-slug]
**Agent version:** [X.X]
**Tier:** [Medium / High]
**Review date:** [YYYY-MM-DD]

| # | Item | Status | Evidence | Notes |
|---|---|---|---|---|
| 1 | Fairness / bias — if the agent makes decisions about people, has it been tested across demographic groups? Known limitations documented? | [✅ Pass / ⚠️ Conditional / ❌ Fail / N/A] | [Link to evidence] | [Reasoning] |
| 2 | Privacy / PII — data classification done; retention rules defined; masking applied where required; user deletion path available | [✅ / ⚠️ / ❌ / N/A] | | |
| 3 | Data residency — LLM calls hit providers in approved jurisdictions? | [✅ / ⚠️ / ❌ / N/A] | | |
| 4 | Reliability / safety — failure modes documented; worst-case action assessed and acceptable? | [✅ / ⚠️ / ❌] | | |
| 5 | Transparency / disclosure — when the agent talks to external parties, is its AI nature disclosed? When internal, is it labeled as AI-generated? | [✅ / ⚠️ / ❌ / N/A] | | |
| 6 | Explainability — for agents affecting individuals, can we explain why a given output was produced? | [✅ / ⚠️ / ❌ / N/A] | | |
| 7 | Inclusiveness — does the agent serve all user groups, or does it exclude / disadvantage some by design or data? | [✅ / ⚠️ / ❌ / N/A] | | |
| 8 | Accountability — single named human owner; clear path to override or stop the agent | [✅ / ⚠️ / ❌] | | |
| 9 | Audit log retention — period defined per the company's data policy and regulation (EU AI Act: high-risk logs ≥ 6 months) | [✅ / ⚠️ / ❌] | | |
| 10 | Synthetic-content labeling — if synthetic content generated (image, audio, video, text that imitates), labeled per EU AI Act Article 50 | [✅ / ⚠️ / ❌ / N/A] | | |

## Sign-off

| Role | Name | Date | Signature |
|---|---|---|---|
| AI CoE Lead | | | |
| General Counsel (or delegate) | | | |
| Department Champion (informed) | | | |

## Conditional approvals / open items

[List any ⚠️ items + the conditions / open items + due date to resolve]

Per-item guidance

1. Fairness / bias

When it applies: Agent makes decisions about people — hiring, promotion, lending, pricing, eligibility for services, performance review. EU AI Act Annex III categories. NIST AI RMF MAP 5.1.

What to check:

• Has the agent been tested across demographic groups (race, gender, age, etc. as relevant)?
• Are approval / rejection / outcome rates documented across those groups?
• Have known limitations been written down?
• Is there a process for affected individuals to challenge an output?

Tools: IBM AI Fairness 360, Microsoft Fairlearn, Aequitas, Google What-If Tool.

N/A criteria: Agent makes no decisions about people. Document the reasoning — auditor will check this isn't being skipped.

2. Privacy / PII

When it applies: Any agent that processes personal data — including mild PII like internal employee names.

What to check:

• Data classification per source documented (Agent Card §5)
• Retention period defined (typically ≥ 6 months for High tier; longer if sector regulation requires)
• Masking / redaction applied to logs where appropriate
• User deletion path exists (GDPR Article 17 right to erasure)

Tools: Microsoft Purview, BigID, OneTrust, TrustArc.

Pass criteria: Each bullet above has a documented answer.

3. Data residency

When it applies: Any agent that processes data subject to residency requirements — EU GDPR, Schrems II, sector regulations.

What to check:

• LLM provider's processing region matches the data's residency requirement
• Vector store + observability platform regions also confirmed
• Cross-border transfer mechanisms in place if data crosses borders (SCCs, adequacy decisions)

Tools: LLM providers with regional endpoints (Anthropic, OpenAI, Bedrock, Vertex, Azure AI Foundry).

Pass criteria: Each data flow documented with region; cross-border transfers have legal basis.

4. Reliability / safety

What to check:

• Failure modes from Agent Card §10 reviewed
• Worst-case action assessed: is it acceptable?
• If worst case is irreversible (e.g., autonomous deletion), additional controls in place
• Kill switch tested

Pass criteria: Worst-case is acceptable AND mitigations are wired AND kill switch is tested.

5. Transparency / disclosure

When it applies:

• External parties (customers, partners, regulators) — disclosure required
• Internal stakeholders consuming AI-generated content — labeling recommended
• EU AI Act Article 50 specifically requires:
  • AI systems that interact with humans must disclose AI nature
  • Synthetic content (deepfakes, generated audio/video/text) must be marked machine-readable

Pass criteria: Required disclosures are wired into the agent's output.

6. Explainability

When it applies:

• For agents affecting individuals' rights, jobs, finances, access to services
• EU AI Act Article 13 obligations for high-risk

What to check:

• Can the company explain why a given output was produced?
• Is the explanation accessible to affected individuals (not just engineers)?
• Are reasoning logs retained?

Tools: SHAP, LIME, Captum (for ML models). Reasoning logs / chain-of-thought capture for LLM agents.

Pass criteria: Explanation mechanism exists and is documented.

7. Inclusiveness

What to check:

• Does the agent serve all user groups equally?
• Language support — does it work for non-English speakers if relevant?
• Accessibility — does its output work with screen readers, alternative inputs?
• Demographic coverage in training/eval data

N/A criteria: Limited internal use case with homogeneous user group. Document the reasoning.

8. Accountability (named owner, override path)

What to check:

• One named human Owner (Agent Card §1)
• Override path: how can a human stop or correct the agent in real time?
• Kill switch tested

Pass criteria: Named owner + working override path + tested kill switch.

9. Audit log retention

Baseline: 6 months for Medium and High tier (EU AI Act Article 19 baseline).

Longer if:

• Sector regulation requires it (SOX = 7 years for financial controls; HIPAA = 6 years; etc.)
• Internal data retention policy requires longer
• Contractual obligation with customers requires longer

Pass criteria: Retention period documented per the agent's tier + regulatory exposure.

10. Synthetic-content labeling

EU AI Act Article 50 obligations:

• AI systems that interact with humans → disclose AI nature
• AI-generated audio, image, video, or text content imitating a person or another entity → mark as AI-generated, machine-readable where feasible
• Emotion recognition or biometric categorization → disclose to subjects

Tools: C2PA content credentials, watermarking libraries.

N/A criteria: Agent doesn't generate synthetic content (e.g., it only parses or ranks). Document the reasoning.

Usage notes

• N/A is a real answer. Several items legitimately don't apply for many agents. Document the reasoning — "Not decisions about people" is enough, but it must be written.
• The 4 always-applicable items: §2 Privacy, §4 Reliability, §8 Accountability, §9 Audit retention. These cannot be N/A.
• EU AI Act Article 50 is broader than people think. It applies to any AI system that interacts with humans, not just chatbots. Read it carefully.
• The checklist is not just for the first review. Re-run on any material change to scope, data sources, or autonomy stage.
• Conditional pass (⚠️) requires a documented resolution path — open item + owner + due date. Don't carry ⚠️ items into pilot.

Common pitfalls

Framework cross-references

• framework.md §18 (this checklist's specification)
• framework.md §10 (risk tier — Medium+ triggers this review)
• framework.md §17 (privileged identities — feeds §8)
• framework.md §22.1 EU AI Act (Articles 13, 14, 19, 50)
• framework.md §22.2 NIST AI RMF (MEASURE function, trustworthy AI characteristics)
• framework.md §22.3 ISO/IEC 42001 (Annex A fairness + transparency controls)
• workflows.md Step 7 (Responsible-AI review)
• workflows.html → In Action view → node M10 (Reviews — Medium+)