Ai governance council

AI Governance Council

Purpose

Designs the decision architecture for human oversight of AI development and deployment: who has authority to approve, delay, or kill an AI launch, how conflicts of interest are controlled, how incidents escalate, and how the whole thing is judged. This is organizational governance, not a multi-model technical pattern.

Core thesis: do not create a discussion forum and call it a council. A council without explicit authority, membership rules, conflict-of-interest controls, evidence standards, and audit trails is a discussion forum wearing a governance costume. Judge a council by launch quality, incident severity and recurrence, and how early dissent gets surfaced — never by how often it meets.

When to use

Use when the user wants to:

• Stand up an AI governance board/council/committee from scratch.
• Draft a charter, decision-rights matrix, or conflict-of-interest policy for such a body.
• Define a model/feature launch approval process or sign-off chain.
• Design an incident escalation path for AI harms.
• Reconcile AI oversight across multiple jurisdictions (US/EU/UK/China).

Trigger even when the user never says "governance council" — e.g. "we need a review process before shipping models," "who signs off on AI launches," "no one here can say no to a risky launch."

When NOT to use

• Multi-model LLM ensembles/orchestration ("have GPT-4 and Claude debate and vote," "cross-check outputs across models") — that is a sibling llm-council-* skill, a technical pattern, not a governance body.
• Generic corporate governance (board composition for stock plans, general fiduciary duties unrelated to AI) — out of scope.
• GDPR / data-privacy compliance questions with no AI-oversight-body angle — out of scope.
• Writing a SECURITY.md or a code-level vulnerability-disclosure process — unrelated document.
• Technical safety evals (red-teaming, jailbreak testing, benchmark scoring) — this skill designs who decides, not how models are tested.

Workflow

1. Scope the org. Ask (or infer): how many AI launches/year, is the org building frontier models or deploying third-party AI, which jurisdictions it operates in, whether a board already exists. Skip ceremony a small org doesn't need — a 15-person startup does not need a 5-person board committee with quarterly deep-dives; scale the hybrid model down instead of skipping oversight entirely.
2. Recommend the hybrid model as the default (see references/governance-council-design.md#the-hybrid-model for full detail):
   • Management AI Council — 7–10 cross-functional members (engineering, product, safety, security, privacy, legal, compliance, risk), staggered 1–3 year terms. Routine review.
   • Board Safety Committee — 3–5 members, majority independent, ≥1 technical. Binding authority on frontier launches, waivers, severe incidents.
   • External Review Panel — 2–5 conflict-screened outside experts. Mandatory input for defined trigger cases (frontier capability, novel harm class, regulator inquiry).
   • Responsible AI office — runs the process and artefacts day to day.
   • Escalation ladder — documented path from management council up to board committee. All three decision layers (operational review, escalation authority, independent challenge) must exist together — see the failure modes below.
3. Assign decision rights by risk tier, not by default consensus-seeking. Use the tiered matrix in references/governance-council-design.md#decision-rights-matrix: low-risk update → delegated approval; medium feature launch → council vote; high-impact/regulated → full review + board notice; frontier capability/safeguard waiver/severe incident → mandatory board committee. Supermajority required to override a documented "do not launch" from safety, security, or privacy leads; dissent recording is mandatory, not optional.
4. Build in conflict-of-interest controls from day one — annual and per-item disclosure, a recusal register, a chair who does not own a shipping target, compensation independent of decision outcomes, financially disinterested external reviewers. See references/governance-council-design.md#conflict-of-interest-minimums.
5. Specify the evidence artefacts the council must produce, not just meet about: charter + decision-rights matrix, AI inventory, impact assessment, launch decision memo with dissent log, incident log + post-incident review, quarterly board report, annual public summary. See references/governance-council-design.md#evidence-artefacts.
6. Set cadence to the minimum that keeps evidence current — monthly ordinary, quarterly deep-dive, ad-hoc emergency within hours, annual charter review plus one severe-incident simulation per year. Do not recommend more meetings as a proxy for more rigor.
7. Add jurisdiction annexes, not parallel processes — one global charter, per-jurisdiction annexes for US/EU/UK/China, "stricter control wins" as the conflict rule. See references/governance-council-design.md#jurisdiction-lanes and flag every jurisdiction fact as time-sensitive: state it, then tell the user to verify current thresholds/dates against the official source before relying on it.
8. Report the KPIs that matter (targets in references/governance-council-design.md#kpis): review coverage, decision latency, dissent visibility, launch-condition completion, incident recurrence, audit-trail completeness. Explicitly do not report meeting frequency as a success metric.
9. Hand over templates, not just advice — pull the charter, COI policy, decision matrix, incident escalation flow, meeting agenda, and roadmap templates from references/governance-council-design.md and adapt them to the org's scope from step 1.

Output spec

A complete answer includes: which of the three layers apply at the user's scale, a decision-rights matrix sized to their risk tiers, the minimum evidence artefacts they must start producing, a COI baseline, and (if relevant) a jurisdiction annex plan with a verify-before-relying-on-this caveat. Point to templates in references/ rather than inlining all of them.

Failure modes to name explicitly

Each missing layer has a documented failure pattern — cite these when a user's proposed design is missing a piece:

• Operational review only, no escalation authority → advisory theatre: the council recommends, nothing binds, risk gets shipped anyway.
• Escalation authority without informed operational review → uninformed executive veto: the board says yes/no without the operational detail to do it well.
• External panel bolted on without a legitimacy model → brittle, performative challenge that collapses under pressure (see Google ATEAC in references/governance-council-design.md#lessons-from-real-implementations).

Gotchas

• Don't let "consensus" quietly mean "no one ever votes no" — require a recorded-dissent mechanism even when the group reaches agreement.
• A council that only ever meets is not evidence of governance; ask for the artefacts (decision memos, incident logs) before declaring the design done.
• Jurisdiction facts in this domain move fast (see AI Act phase-in dates) — never present a date/threshold as evergreen; always attach "verify against the official source."
• Cite only named frameworks/organizations (NIST AI RMF, ISO/IEC 42001, EU AI Act, named companies' public responsible-AI programs); never fabricate deep URLs.
• This skill designs governance structure; it is not legal advice. Route jurisdiction-specific obligations (EU AI Act classification, state ADMT laws, China filings) to qualified counsel before relying on them.

References

• references/governance-council-design.md — full hybrid-model detail, decision-rights matrix, COI policy, evidence artefacts, cadence, KPI table, jurisdiction lanes, and templates (charter, COI policy, decision matrix, incident escalation flow, meeting agenda, implementation roadmap).