CEOs hear "AI risk" at every board meeting in 2026. Boards ask three things: who is accountable, how is risk measured, and what is the kill switch. A governance framework answers all three before they ask. Here is the working version.
Why governance now
Three drivers that turn governance from optional to mandatory:
- Insurance — D&O policies now ask for AI governance evidence.
- Regulators — EU AI Act Article 9 requires risk management systems.
- Investors — proxy advisers add AI governance to ESG-style scorecards.
A formal framework satisfies all three at once.
The charter
The cornerstone document. One page, board-approved. Covers:
- Purpose — what AI is used for, what it is not.
- Principles — values that constrain agent design (e.g., human oversight on irreversible actions).
- Scope — which agents are governed (typically: any in production touching customer data).
- Accountability — who owns AI risk at the executive level.
- Review cadence — quarterly board update, annual external review.
Templates exist; the value is in the conversation that produces it.
The four committees
Most enterprises end up with four governance bodies:
1. AI Steering Committee
Executive-level. Sets strategy, approves budgets, owns the charter. Meets quarterly.
2. AI Risk Committee
Cross-functional (risk, legal, security, engineering). Reviews high-risk agents before launch and after incidents. Meets monthly.
3. AI Ethics Review
Independent voices (sometimes external). Reviews use cases that touch sensitive populations. Meets ad-hoc.
4. AI Operations Forum
Engineering-led. Daily-to-weekly. Owns the operational evidence for the risk committee.
Smaller orgs collapse these into two; very large orgs split each into sub-committees by business unit.
The lifecycle
Every agent passes through five gates:
Idea → Concept → Build → Deploy → Operate → Retire
Gate 1 (Idea → Concept):
AI Risk Committee classifies risk tier.
Gate 2 (Concept → Build):
Approval for high-risk; notify for low-risk.
Gate 3 (Build → Deploy):
Pre-launch review: eval results, risk register entry, audit hooks.
Gate 4 (Deploy → Operate):
90-day post-launch review. SLO compliance, incident count.
Gate 5 (Operate → Retire):
Decommission process; data deletion plan.
Without gates, governance is theatre.
The risk register
Every production agent has an entry:
agent_id: support-bot-v3
risk_tier: limited
top_risks:
- id: r1
risk: hallucination on product capability questions
likelihood: medium
impact: medium
mitigation: source pinning + post-pass verification
owner: support-eng-lead
- id: r2
risk: prompt injection via support tickets
likelihood: medium
impact: high
mitigation: input sanitisation + scope-limited tools
owner: security-team
last_reviewed: 2026-04-15
next_review: 2026-07-15
Reviewed by the AI Risk Committee on cadence. Lives in the compliance toolkit.
Accountability mapping
Three roles, distinct responsibilities:
- Agent Sponsor (executive) — answers for the agent's business outcomes.
- Agent Owner (engineering) — answers for the agent's technical operations.
- Agent Steward (compliance) — answers for the agent's regulatory posture.
Same person can hold two but not all three. Each role has a documented escalation path.
The kill switch
Boards specifically ask about this. Three components:
- Per-agent disable — operator can stop a single agent in seconds.
- Per-tool disable — operator can pull a tool from every agent (e.g., on supplier breach).
- Full estate disable — exec-level button that pauses all agents.
Tested quarterly. Documented in incident response runbooks.
Reporting up
A monthly executive dashboard:
- Number of agents in production by risk tier.
- Open risk register items above threshold.
- Recent incidents and remediation status.
- Spend (see cost optimization).
- Major upcoming launches and their gate status.
A quarterly board pack adds: external benchmarks, regulatory horizon scan, ethical review summary.
What does NOT count as governance
- A document nobody reads — quarterly review or it does not exist.
- Engineer-only committees — without legal and risk in the room, decisions miss critical constraints.
- No teeth on the gates — if Gate 3 has never blocked a launch, it is theatre.
- No external review — internal-only governance fails the third-party test.
Common mistakes
- One framework for all agents — risk-tier the framework; light governance for low-risk.
- No risk register update cadence — entries go stale within months.
- No incident-to-governance loop — incidents must update the register and the framework.
- Skipping the kill switch test — the first time you press it should not be in a real incident.
Where this is heading
Two trends by 2027: standardised AI governance frameworks endorsed by major industry bodies, and AI governance scoring services (Vanta-style) that assess your framework against the standard. Build it now; swap into the standard when it lands.