Five people times 2 hours bi-weekly times 26 meetings equals 260 person-hours per year. About 0.15 FTE distributed across senior leaders. Small relative to the risk being managed.

The AI Ethics Board That Actually Works (and Why Most Don't)

TL;DR

Why most AI ethics boards fail:

No decision rights — they review, the org ignores.
Wrong composition — too many ethicists, too few operators.
Wrong cadence — quarterly meetings can’t keep up with weekly deployments.
No teeth — recommendations are non-binding.
No artifacts — no record of what was decided or why.

What works: a small standing committee with named decision rights on a defined scope, monthly cadence, written record of decisions, and explicit escalation path for the cases they can’t decide. Five people, two hours a month, real authority.

Most AI ethics boards are theatre. The structural reasons why, and what a working board looks like.

The “AI ethics board” became a popular response to AI governance pressure starting around 2023. Most of these boards have produced little. The pattern is consistent: announced with fanfare, populated with prestigious names, meets quarterly, makes recommendations that get filed but rarely implemented. This piece is the structural critique and the alternative model.

Why most ethics boards fail

1. No decision rights

The committee reviews; the organization decides. Recommendations get noted, sometimes implemented, often forgotten. Without authority over specific decisions (which agents go to production at which autonomy level, which use cases are approved), the committee is advisory in name only.

2. Wrong composition

Common pattern: heavy on ethicists and academics, light on operators. The conversations tend toward principles; the work needed is operational. The right composition for working boards is mostly internal operators (CTO, CISO, AI program lead, function leads) with one or two outside advisors, not the reverse.

3. Wrong cadence

Quarterly meetings can’t govern weekly deployments. By the time the board meets, decisions have been made; the board is reviewing, not deciding. Monthly is the minimum; bi-weekly is better for active programs.

4. No teeth

Recommendations are non-binding. The committee says “we’re concerned about X”; the org ships X anyway. Without veto authority on specific scoped decisions, the committee is theatre.

5. No artifacts

Most committees don’t produce written records. Decisions, rationale, dissent — none of it documented. When the committee’s work is needed in an incident, regulatory inquiry, or audit, there’s nothing to show.

What works: the operational AI committee

Different name, different design. A small standing committee with specific authority.

Composition

Five people. Specifically:

AI program lead (chair).
CTO or VP Engineering (technical authority).
CISO or security lead (security authority).
General Counsel or AI Counsel (legal/regulatory authority).
Senior business leader (rotating, from the function whose use cases are being reviewed).

Plus 1–2 outside advisors (academic, industry, civil society) on retainer for input but not voting.

Decision rights

Specific. Examples:

Approval for new agents at Level 2 or higher.
Approval for use cases involving regulated functions.
Approval for vendor selections above a threshold.
Recommendation (binding) on incident response actions for material incidents.

The matrix from AI Governance: The Operating System for Responsible AI maps cleanly to this committee’s authority.

Cadence

Bi-weekly, 90 minutes. With async review for time-sensitive decisions (the chair can call for written votes between meetings).

Artifacts

For every decision: a one-page record. Decision, rationale, dissent if any, conditions if any. Filed to the governance archive; available to regulators and counsel.

For every meeting: a public-facing (internal) summary of topics discussed. Without specifics that compromise confidentiality, but enough that the broader org knows what governance is doing.

Escalation

Cases the committee can’t decide (or that exceed scope) escalate to:

CEO for cross-function strategic decisions.
Audit committee or risk committee of the board for material risk decisions.
External advisors for specialized expertise gaps.

The escalation path is explicit, documented, and used.

What about external ethics boards?

Some companies maintain external ethics boards for legitimacy and outside perspective. These work when:

The external board’s role is advisory, with explicit scope.
Internal operational committee handles real decisions.
External board reviews the operational committee’s record annually.
External board can publish independent observations.

External boards become theatre when they’re positioned as decision-makers but lack actual authority, or when they’re populated for credentialing without substantive expertise.

What to do this quarter

Audit your current AI ethics board (if any). Five-question test: does it have decision rights, the right people, the right cadence, real teeth, and written artifacts? Most boards fail at least three.
Redesign as the operational AI committee if needed. New charter, new composition, new cadence.
Document decision rights specifically. The committee decides X, recommends on Y, doesn’t decide Z.
Schedule the first meeting. Operational momentum matters more than perfect setup.

FAQ

Should we still have a “Chief AI Ethics Officer”? Usually no. The role becomes ceremonial when the operational work is owned by the AI program lead. If the title is needed for external positioning, it can be a hat the AI program lead wears, not a separate role.

What if our ethics board is already in place and is theatre? Restructure. Painful but necessary. The status quo of theatrical governance produces real risks (incidents, regulatory exposure) that the structural fix prevents.

Should we publish our committee’s existence? Acknowledge yes (transparency). Publish all decisions: no (operational confidentiality). Publish summary statistics quarterly: increasingly common practice.

Do we need the committee if we’re a small company? Smaller orgs don’t need a formal committee; the AI program lead with input from leadership covers the function. The committee structure scales up around 500–1000 employees.

What’s the cost? Five people × 2 hours bi-weekly × 26 meetings = 260 person-hours / year. Plus ~50 hours of agenda prep. About 0.15 FTE distributed across senior leaders. The cost is real but small relative to the risk being managed.

Working with JAIN on AI governance design? We help executive teams design the operational committee that actually governs, not the ethics board that doesn’t. Book a 30-minute call.

Related reading: