borghei/Claude-Skills / c-level-advisor/chief-ai-officer-advisor
chief-ai-officer-advisor
UnverifiedAI leadership advisor on AI strategy, governance, risk, investment, and org design. Use when defining an AI strategy, building an AI governance program, scoring AI maturity, or drafting an AI risk register.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 11.7 KB |
| references/ai-org-and-talent.md | markdown | 8.9 KB |
| references/ai-risk-and-governance.md | markdown | 12.5 KB |
| references/ai-strategy-framework.md | markdown | 8.8 KB |
| scripts/ai_investment_planner.py | script | 13.7 KB |
| scripts/ai_maturity_assessor.py | script | 16.8 KB |
| scripts/ai_risk_register_generator.py | script | 16.0 KB |
references/ai-risk-and-governance.md
View on GitHub# AI Risk & Governance Reference
Practical reference for setting up an AI governance program and the
underlying risk taxonomy. Maps to NIST AI RMF 1.0, ISO/IEC 42001:2023
(AIMS), the EU AI Act, and adjacent expectations from financial-services
model risk management (SR 11-7) and FDA AI/ML guidance.
## 1. Frameworks at a glance
| Framework | Scope | What it cares about |
|-----------|-------|---------------------|
| NIST AI RMF 1.0 | Voluntary, US-origin, sector-agnostic | Govern / Map / Measure / Manage outcomes |
| ISO/IEC 42001:2023 (AIMS) | Certifiable AI management system | Org-level processes, policies, controls |
| EU AI Act | Mandatory in EU, extraterritorial | Risk tiers, conformity, post-market monitoring |
| SR 11-7 (US Fed) | Banks, model risk management | Model lifecycle, validation, documentation |
| FDA AI/ML guidance | Medical software | Predetermined change control, real-world performance |
| ISO/IEC 23894 | Risk management for AI | Risk-management process detail |
| ISO/IEC 5338 | AI system lifecycle | Lifecycle stages and activities |
If you operate across multiple regions, build your control set around
**ISO 42001 + NIST AI RMF**, then layer EU AI Act + sector regs on top.
## 2. NIST AI RMF — 4 core functions
The RMF organizes activities into four functions. A useful checklist:
### Govern
- Risk management roles and accountabilities are defined and resourced.
- Policies are published, signed, and communicated.
- Workforce has appropriate AI literacy for their role.
- Incident reporting and review processes exist.
- Third-party AI relationships are governed.
### Map
- Context for each AI system (purpose, users, deployment setting) is documented.
- Categorization of risk is performed at system intake.
- Impact assessment considers individuals, groups, and society.
- Tradeoffs between accuracy and other values are documented.
### Measure
- Quantitative and qualitative metrics are defined per system.
- Pre-deployment testing (eval) is performed against documented criteria.
- Tracking of performance, fairness, and robustness post-deployment.
- Independent review for high-risk systems.
### Manage
- Risk responses are prioritized and assigned.
- Continuous monitoring detects drift and degradation.
- A documented decommissioning process exists.
- Mechanisms for human override, escalation, and incident response are tested.
## 3. ISO/IEC 42001 — AIMS minimum viable program
A pragmatic ISO 42001 implementation (12-week sprint pattern):
| Week | Workstream | Deliverable |
|------|-----------|-------------|
| 1–2 | Scope + context | AIMS scope statement, interested parties register |
| 3–4 | Policy + leadership | AI policy, roles + responsibilities (RACI) |
| 5–6 | Risk + impact | Risk register, AI impact assessment template, AIIAs for top systems |
| 7–8 | Controls (Annex A) | Statement of applicability across Annex A controls |
| 9–10 | Operations | Operating procedures: change management, supplier management, incident response |
| 11 | Internal audit | First internal audit pass, NCs logged |
| 12 | Management review | Management-review record, corrective actions, certification readiness |
See `ra-qm-team/audit-prep/aims-audit` for the deeper audit-prep playbook.
## 4. EU AI Act — risk tiers and obligations
### Risk tiers
| Tier | Examples | Status |
|------|----------|--------|
| Unacceptable risk | Social scoring, real-time biometric ID in public (with narrow exceptions), manipulative AI | Prohibited |
| High-risk | Annex III categories (employment screening, credit scoring, critical infra, law enforcement, education, medical devices) | Heavy obligations |
| Limited risk | Chatbots, deepfakes | Transparency obligations |
| Minimal risk | Spam filters, video game AI | Voluntary codes of conduct |
| GPAI (general-purpose AI) | Foundation models | Provider obligations; systemic-risk tier adds extra |
### High-risk obligations (Title III, Chapter 2)
A high-risk system requires:
1. **Risk management system** (continuous, throughout lifecycle)
2. **Data and data governance** (representative, relevant, free of errors)
3. **Technical documentation** (Annex IV content)
4. **Record-keeping** (automatic logs of operation)
5. **Transparency and user information**
6. **Human oversight**
7. **Accuracy, robustness, and cybersecurity**
8. **Conformity assessment** before placing on the market
9. **EU declaration of conformity + CE marking**
10. **Post-market monitoring** (Article 72)
11. **Serious incident reporting** (Article 73) — typically 15 days; immediate for fundamental rights breaches or critical infra
### Timeline checkpoints (as amended by the Digital Omnibus, Regulation (EU) 2026/1744)
- 1 Aug 2024 — Act enters into force
- 2 Feb 2025 — Prohibitions on unacceptable-risk practices apply
- 2 Aug 2025 — GPAI obligations apply; governance bodies operational
- 2 Aug 2026 — Article 50 transparency obligations apply
- 2 Dec 2027 — High-risk obligations for Annex III systems apply (moved from 2 Aug 2026)
- 2 Aug 2028 — High-risk obligations for AI in products under EU product-safety law apply (moved from 2 Aug 2027)
Check the [Official Journal text](https://eur-lex.europa.eu/eli/reg/2026/1744/oj) before relying on a date.
## 5. AI Impact Assessment (AIIA) — what to include
Whether you call it AIIA, AI risk assessment, FRIA (Fundamental Rights
Impact Assessment), or model card, a credible assessment includes:
1. **Purpose and deployment context** — who uses it, where, why, with what
2. **Data lineage** — sources, consent, licensing, sensitivity, sovereignty
3. **Modeling approach** — model family, training process, eval methodology
4. **Performance** — headline metrics + subgroup performance + failure modes
5. **Risk identification** — what could go wrong (taxonomy in §6)
6. **Mitigations** — controls in place to reduce each risk
7. **Residual risk** — what remains after mitigations; explicit acceptance
8. **Human oversight** — review/override/escalation paths
9. **Monitoring plan** — what gets watched, with what threshold, by whom
10. **Decommissioning** — when do we kill it; what's the rollback
For high-risk systems, an AIIA is non-negotiable. For low-risk internal
productivity tools, a one-page summary is usually enough.
## 6. AI risk taxonomy (use this for the register)
Group risks for easier tracking. The taxonomy below is a pragmatic
union of NIST, ISO, and common red-teaming categories.
### Model risk
- Inaccuracy / hallucination
- Bias and unfair subgroup performance
- Distribution shift / drift
- Adversarial robustness (input perturbations, jailbreaks)
- Reproducibility and versioning failures
### Data risk
- Data quality (missing, stale, biased)
- Data lineage gaps (can't prove provenance)
- Data leakage (training data exposed in outputs)
- Sensitive data in training without consent
- Cross-border data transfer issues
### Security risk
- Prompt injection (direct and indirect)
- Data exfiltration via the model
- Model theft / weight extraction
- Supply chain (compromised model from a hub)
- Insecure agent tool use
### Privacy risk
- Re-identification of training subjects
- Membership inference attacks
- PII in inputs being logged
- Disclosure of business-confidential info to third-party LLM
### Operational risk
- Vendor lock-in / single foundation-model dependency
- Cost runaway (token explosion, agent loops)
- Latency / availability regressions
- Capability concentration in a single team
### Compliance risk
- EU AI Act / sector regulation non-compliance
- Failure to meet contractual AI/data terms
- IP risk (training on copyrighted content, model output IP ownership)
- Mandatory disclosure obligations missed
### Reputational and human-rights risk
- Harmful or offensive outputs
- Discrimination, disparate impact
- Lack of transparency to users
- Misuse / dual-use
## 7. AI governance committee structure
A two-tier structure works for most companies past the early stage.
### Tier 1 — AI Council (executive)
- **Members:** CEO chair, CAIO sponsor, CTO, CISO, CFO, GC, business-line heads
- **Cadence:** monthly or quarterly
- **Mandate:** strategy approval, budget, principles, top-risk acceptance
- **Out of scope:** technical architecture decisions
### Tier 2 — Model Risk / Technical Review Board
- **Members:** CAIO chair, head of ML/data, head of security, privacy lead, product reps
- **Cadence:** biweekly or weekly
- **Mandate:** approve new models, review eval results, accept residual risk
- **Authority:** can require additional eval; can block deployment; can require fallback
Plus working groups as needed: red team, AI policy, vendor management.
## 8. Model approval workflow
A clean approval workflow (target: ≤2 weeks for low/medium risk):
```
Intake → Risk classification → Eval design → Eval execution
→ Risk review → Approval / Conditional / Reject → Deployment gate
→ Post-deployment monitoring → Quarterly review
```
Approval decisions are recorded in the **model registry** with:
- Model card / fact sheet
- Eval results vs published threshold
- Residual risks and acceptance
- Approver(s) and date
- Review trigger conditions (drift, new use case, new regulation)
## 9. AI incident response
AI incidents are different enough from generic security incidents to deserve
their own playbook.
### Detection triggers
- Eval pass rate drops below threshold
- User-reported harmful output above noise floor
- Drift detector fires on input distribution
- External disclosure (researcher, customer, regulator)
- Cost or latency anomaly
### Response steps
1. **Contain** — disable the model or feature behind a feature flag
2. **Triage** — severity (S1–S4), data-subject impact, regulatory implication
3. **Investigate** — pull logs, reproduce, identify root cause (data, model, prompt, integration)
4. **Notify** — internal (exec, GC, CISO), customer (if contractually required), regulator (if applicable)
5. **Remediate** — fix and re-eval, then re-enable
6. **Post-mortem** — within 5 business days; publish to the AI council
### Regulatory clocks
- EU AI Act Article 73: serious incidents — 15 days (immediate for fundamental rights / critical infra)
- GDPR Article 33: personal data breach — 72 hours to supervisory authority
- Sector-specific (FDA, EMA, banking) — varies; pre-publish your matrix
## 10. Vendor and third-party AI governance
Most AI risk now arrives via vendors. Govern accordingly.
### Vendor intake checklist
- Model provenance (base model, training data summary, evals)
- Data handling: do they train on your inputs/outputs? (default: no)
- Data residency: where are inputs processed and stored?
- Security: SOC 2 / ISO 27001; AI-specific controls
- Contract: indemnity for IP claims, audit rights, exit clause
- Incident notification SLA
- Roadmap visibility and deprecation notice period
### Foundation model contracts — must-haves
- No-training clause covering inputs and outputs
- Defined data residency and processing locations
- Indemnity for output IP claims
- 30-day deprecation notice for any model used in production
- Service credits or refunds for material SLA breaches
- Right to audit summarized governance evidence (or accept a third-party report)
## 11. Mapping table — controls across frameworks
| Control area | NIST AI RMF | ISO 42001 | EU AI Act |
|--------------|-------------|-----------|-----------|
| Governance | Govern 1-6 | Clauses 5, 9 | Art. 17 (QMS) |
| Risk assessment | Map 1-5 | Clauses 6.1, 8.2 | Art. 9 (RMS) |
| Data quality | Measure 2.1, 2.2 | Annex A.7 | Art. 10 |
| Documentation | Govern 1.3, Manage 4 | Clauses 7.5, 8.5 | Art. 11 + Annex IV |
| Human oversight | Manage 3.2 | Annex A.6.2.6 | Art. 14 |
| Robustness | Measure 2.5-2.7 | Annex A.6.2.4 | Art. 15 |
| Post-market monitoring | Manage 4.1 | Clause 9.1 | Art. 72 |
| Incident reporting | Manage 4.3 | Clause 10 | Art. 73 |
A control once, comply many times: design controls to satisfy the union,
not one framework at a time.
## 12. Common pitfalls
- **AI governance committee with no decision rights.** Either give it real authority over approvals and spend, or disband it.
- **Risk register that never gets updated.** Tie ownership to the model registry; review quarterly.
- **Approval workflow that becomes a tollbooth.** Publish SLAs; fast-path low-risk patterns.
- **AIIA template that's 40 pages.** No one fills it out. Keep it to 3–5 pages with deeper appendices.
- **Conflating safety with security.** They share tools but have different threat models and owners.
- **Outsourcing red-teaming once and calling it done.** Continuous adversarial testing should be part of the eval harness, not a one-time engagement.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation). | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |