Ai Risk Governance

AI Risk & Governance

When to Use

• Classifying AI use cases by risk tier and impact
• Drafting AI acceptable-use policies and governance frameworks
• Building AI risk registers with likelihood/severity/mitigation tracking
• Preparing model cards, system cards, or DPIAs for AI deployments
• Reviewing third-party LLM vendors (data terms, fine-tuning, safety commitments)
• Mapping AI products to frameworks (NIST AI RMF, ISO 42001, EU AI Act concepts)
• Aligning product and engineering teams with compliance requirements
• Designing human-in-the-loop oversight for consequential AI decisions

When NOT to Use

• Implementing RAG pipelines, agents, or production features → ai-engineer
• Running jailbreak tests or adversarial campaigns → ai-redteam
• SOC 2/ISO evidence automation and technical control mapping → compliance-engineer
• General SOC 2 IT controls without AI scope → cybersecurity
• Commercial/enterprise AI solution architecture → applied-ai-architect-commercial-enterprise
• Skills portfolio governance and publish gates → ai-skill-manager

Related skills

Core Workflows

1. Use-case intake and classification

Capture:

Risk tier (example):

See `references/risk_classification.md` for EU AI Act–oriented mapping (non-legal).

2. Risk assessment

Use structured worksheet:

1. Identify hazards (bias, hallucination, leakage, misuse, dependency) 2. Estimate likelihood and severity 3. Define mitigations (technical, process, legal) 4. Assign owner and review date 5. Residual risk acceptance sign-off

See `references/risk_assessment.md` for worksheets and NIST AI RMF functions.

3. Documentation artifacts

See `references/documentation.md` for model card sections and change log.

4. Policy and oversight

• Acceptable use policy (prohibited uses, approval paths)
• Human oversight rules for consequential decisions
• Escalation for policy violations and serious incidents
• Training for builders and reviewers

See `references/policy_oversight.md` for governance committee cadence.

5. Vendor and third-party models

Review: data processing terms, subprocessors, retention, fine-tuning on customer data, safety commitments, breach notification, exit plan.

See `references/vendor_review.md` for vendor questionnaire topics.

When to load references

• Tiering and regulation mapping → references/risk_classification.md
• Assessments and frameworks → references/risk_assessment.md
• Model cards → references/documentation.md
• Policies and committees → references/policy_oversight.md
• Vendors → references/vendor_review.md

Documentation

Table of contents

1. Model card sections 2. Change log

Model card sections

• Intended use and out-of-scope uses
• Training/fine-tune data summary
• Evaluation results and limitations
• Ethical considerations and biases observed
• Environmental impact (optional)

Change log

Track: model version, prompt version, retrieval index version, deploy date, approver.

Policy and oversight

Table of contents

1. AUP topics 2. Committee cadence

AUP topics

• Prohibited uses (e.g., covert surveillance, illegal content)
• Required human review domains
• Data handling and retention
• Third-party model restrictions

Committee cadence

Risk assessment

Table of contents

1. Worksheet 2. NIST AI RMF mapping

Worksheet

| Hazard | Likelihood | Severity | Mitigation | Residual | Owner |

Review quarterly or on material change.

NIST AI RMF mapping

Risk classification

Table of contents

1. Tiering factors 2. EU AI Act orientation

Tiering factors

Score each 1–3: impact, autonomy, data sensitivity, scale, reversibility.

Sum → Low / Medium / High governance path.

EU AI Act orientation

Not legal advice. Map use case to:

• Prohibited practices checklist (reject early)
• High-risk annex domains (enhanced documentation)
• GPAI provider vs deployer obligations for third-party models

Escalate ambiguous cases to legal/compliance.

Vendor review

Table of contents

1. Questionnaire topics 2. Exit plan

Questionnaire topics

• Subprocessors and data residency
• Training on customer data (opt-in/out)
• Retention and deletion SLAs
• Safety evaluations and incident history
• API logging and enterprise controls (SSO, SCIM)

Exit plan

Alternative model/provider, data export, contract termination assistance.