Enterprise AI adoption & strategy
The AI Steering Committee That Actually Meets: Seats, Cadence, Decision Rights and a 90-Day Charter
· By AIHQ Team
The committee that exists and the committee that decides
Most AI steering committees in Malaysian and Singaporean organisations are set up in the same fortnight: a CIO or Chief Data Officer circulates a draft Terms of Reference, eight to twelve names are added, the first meeting is scheduled for the following month, and then the calendar quietly slips. Six months later you have a governance body that has met twice, produced one risk register, and blocked nothing and approved nothing.
Meanwhile engineers keep waiting. A team wants to put an internal retrieval-augmented generation (RAG) service on a document store and needs a decision on data classification. Another wants to buy 40 Copilot licences and needs a budget line and a security review. A third has a working prototype stuck in a repository nobody can deploy. In every case the request has a clear owner on the technical side and no owner on the governance side.
That gap is almost never a people problem. It is a design problem. The committee was given a mandate to "oversee AI" but not the three things that make oversight real: specific seats, a fixed cadence, and written decision rights bounded by thresholds. Below is a structural model you can adapt — with concrete seat definitions, a RACI, escalation paths to the board and a 90-day charter skeleton you can paste into a document and edit.
Why most AI steering committees stall
Four failure modes account for most of the stalled committees we see across corporate, public sector and regulated environments.
- The committee is a briefing forum, not a decision forum. The agenda is a series of vendor presentations and departmental updates. No item requires a decision, so no decision gets made, and the people who can decide stop attending.
- The seats are assigned by seniority, not by decision relevance. A committee of seven VPs with no one who can speak to data architecture, security posture or employment implications defers every question to a working group that meets less often.
- There are no thresholds. Every AI request, from a free chatbot subscription to a customer-facing agent, is treated as equally weighty, so the committee drowns in small approvals and starves the ones that matter.
- There is no escalation path. When the committee cannot decide — usually a risk appetite question — the item goes back to the requester with a request for more information. It never reaches the board, so the board never sets the appetite.
A fifth, subtler failure: the committee owns nothing operationally. If no member is accountable for moving a pilot into production, the committee can approve a pilot and still watch it die. This is why the governance discussion has to connect to the adoption work — the same reason a well-scoped enterprise AI roadmap cost model usually gets revised the moment governance starts making decisions, because budget lines only become real when approvals do.
Who should hold a seat — and who shouldn't
A working AI steering committee is usually seven to nine voting members. More than that and attendance collapses; fewer and you cannot cover the four decision domains (business value, technology and data, risk and compliance, people and capability).
The core seats:
| Seat | Why they are there | Typical decision they own |
|---|---|---|
| Executive sponsor (CEO, COO or a named C-suite leader) | Only person who can settle risk-appetite disputes and reallocate budget | Final tie-break; approves Tier 3 use cases |
| CIO / Head of IT or Engineering | Owns architecture, integration, security posture and vendor contracts | Approves platform and tooling standards |
| Head of Data (or Data Governance lead) | Owns data classification, quality and access rules | Approves data-access and retention decisions |
| Risk, Compliance or Legal lead | Owns regulatory exposure and customer-facing risk | Signs off on external-facing AI and personal-data use |
| HR / L&D lead | Owns role impact, capability gaps and training commitments | Approves role-change and training scope |
| Business owner (one rotating department head) | Owns the workflow and the outcome metric | Owns the business case and adoption target |
| AI programme lead / delivery owner | Owns the portfolio, pipeline and reporting | Runs the pipeline; no veto power |
Seats to think twice about. A vendor or systems integrator should be present for defined agenda items only and should never be a voting member. A board member can be an observer and escalation recipient, but putting the board inside the operating committee blurs the line the board is supposed to hold. A single "Head of AI" without a team is a seat with responsibility and no capacity — if you create that role, give it a named backfill and a defined scope before the first meeting.
For organisations with large technical estates, the security architect is usually the highest-leverage attendee you can add, and often the one most frequently left off. If your risk tiering will be built around data classification, someone has to define those classes. That is also the ground the internal chatbot governance checklist covers in detail — permission-aware retrieval questions belong in front of this committee, not discovered at launch.
The decision rights: a RACI for AI
The single most useful artefact you can produce in the first thirty days is a decision-rights table. Below is a starting model. R = responsible (does the work), A = accountable (owns the outcome), C = consulted, I = informed.
| Decision | Committee | Exec sponsor | CIO/IT | Data lead | Risk/Legal | Business owner |
|---|---|---|---|---|---|---|
| Adopt a standard tool (approved list) | A | I | R | C | C | I |
| Approve Tier 1 use case (internal, non-sensitive) | I | I | A | C | I | R |
| Approve Tier 2 use case (internal, sensitive data or agentic) | A | C | R | C | C | R |
| Approve Tier 3 use case (customer-facing, regulated, or high financial exposure) | C | A | R | C | R | R |
| Data access and retention rules | C | I | C | A | C | I |
| AI policy wording | C | A | C | C | R | I |
| Training and capability budget | A | C | I | I | I | R |
| Pause or retire an AI system | A | C | R | I | C | I |
Two design notes. First, Tier 1 approvals should not come to the committee at all — they should be delegated to the CIO with a standing approved-tool list and a published exclusions list. Second, the committee should have explicit authority to say no within a service-level target (we suggest five working days for Tier 2), or the pipeline will treat the committee as a delay rather than a gate.
Tiering itself should be definitional, not case-by-case. A workable three-tier model: Tier 1 is internal-only, no personal data, no external output, cost under a defined threshold. Tier 2 adds personal data, agentic action, or system-to-system integration. Tier 3 covers customer-facing output, regulated decision support, cross-border data flows and anything with financial exposure above your materiality line.
Cadence: what a committee that meets actually does
Monthly is the minimum viable rhythm for a committee that wants to unblock work. Quarterly governance with monthly operational delivery is a common structure, but only if the operations group has delegated authority — otherwise the quarterly committee becomes a bottleneck by default.
A 90-minute monthly agenda that works:
- Decision log review (10 min) — every open decision, its owner and its due date. Anything past due escalates automatically.
- Blocked items (25 min) — requests stuck on a governance question, each presented with the specific decision required.
- Pipeline and portfolio review (20 min) — use cases moving between tiers, pilots with dates, anything moving to production.
- Risk and incident review (15 min) — incidents, near-misses, policy exceptions granted since the last meeting.
- Capability and adoption metrics (15 min) — active users, workflows changed, training completions by role.
- One deep-dive (5 min to table, 20 min async) — a specific workflow under consideration, read in advance.
Two rules make the cadence stick. A meeting without a decision log entry is treated as not having happened, and the decision log is visible to anyone who requests it. And every item must arrive with a named business owner and a stated decision required; updates without a decision request go in writing.
Escalation to the board in five working days
The board's role is to set risk appetite, approve material capital allocation, and receive assurance — not to adjudicate individual use cases. That distinction only holds if there is a written path between the two.
- Risk appetite. The board approves an annual statement covering, at minimum: which categories of AI use are in scope, which are out of scope, tolerance for external-facing AI, and data-residency expectations.
- Threshold-based escalation. Anything above a defined financial exposure, anything customer-facing in a regulated product, or any exception to the approved tool list goes to the board risk committee within a defined window (five working days is achievable if the submission is pre-formatted).
- Assurance reporting. A quarterly one-page report to the board with four numbers: use cases in production, use cases paused or retired, policy exceptions granted, and training coverage by role. Four numbers, not forty slides.
- Named escalation contact. A single committee member owns the board relationship and the submission format.
This is the point where governance and regulated-sector oversight converge. Public sector and GLC teams usually need this path to be more formal than a commercial organisation does, and the controls framework for AI agent governance in public sector settings is a useful reference for the specific pre-automation gates to route through this escalation route.
A 90-day charter skeleton
Weeks 1–4 — Foundation. Owner: AI programme lead with exec sponsor. Deliverables: charter signed by the executive sponsor; seat list with named deputies; three-tier definition document; a first-pass inventory of AI tools already in use (you will usually find more than expected — shadow usage across marketing, HR and finance is common); approved-tool list v0.1; decision log template.
Weeks 5–8 — Decision rights and policy. Owner: Risk/Legal lead with CIO. Deliverables: RACI table ratified; AI policy v0.1 covering permitted/prohibited use, data classification and disclosure expectations; a training commitment mapped to roles, noting which roles need only awareness and which need applied workflow training. If any programme is intended to be HRD Corp claimable, confirm eligibility and submission requirements early in this window rather than at invoice time — HRDC claim rejection patterns are almost always created by late scoping, not by bad intentions. Programmes can be structured to be claimable subject to client eligibility, grant approval and HRD Corp submission requirements.
Weeks 9–12 — First cycle and calibration. Owner: committee chair. Deliverables: first three Tier 2 decisions processed within the five-day target; first escalation submitted to the board risk committee in the agreed format; first capability metrics read (active users, workflows changed, training completion by role); a written review of what the committee got wrong — usually that Tier 1 approvals are still consuming agenda time or that a required seat is not attending.
A realistic budget for the governance layer itself, excluding tooling, sits in the MYR 40,000–120,000 range for the first year for a mid-sized Malaysian organisation — largely chair and member time, external policy drafting, and any independent review. Tooling and training sit separately and are the larger lines; a full AI implementation cost model for Malaysia is the right companion for the numbers the committee will approve.
What the committee should be measuring
Keep the metric set small and make it about decisions rather than activity.
- Decision throughput — decisions made per month, and median days from request to decision against the five-day target.
- Queue health — number of open decisions older than target; number of requests returned for information more than once.
- Pipeline conversion — use cases moving from pilot to production, and how many were retired (a healthy committee retires things).
- Exception rate — policy exceptions granted, by category. A rising rate is a signal that the policy is miscalibrated, not that people are misbehaving.
- Coverage — percentage of roles trained against the capability matrix, and role-based usage rather than licence counts. Seat utilisation is not adoption; the AI competency matrix gives you the levels to measure against.
- Escalation timeliness — escalations to the board submitted within the agreed window.
If a metric cannot change a decision next month, it does not belong on the monthly pack.
What this means for a team running its first 90 days
Governance bodies fail quietly. No one announces that the committee has stopped deciding — attendance just thins, the decision log stops growing, and the delivery teams learn to route around it. The fix is structural rather than cultural: seven to nine named seats with defined decision domains, a three-tier model that keeps small approvals out of the room, a five-day decision target, a written escalation path with a board-facing assurance report, and a charter that names owners by week.
Start with the decision-rights table. It is one page, it forces the conversations that the Terms of Reference usually avoid, and it will tell you within a fortnight whether your committee is designed to unblock work or merely to observe it. For teams that want to align the governance structure with practical capability building and, where needed, implementation support, AIHQ's work spans custom AI solutions and role-based capability programmes.
FAQ
How many people should sit on an AI steering committee?
Seven to nine voting members is the practical range. Fewer than seven and you cannot cover the four decision domains (business value, technology and data, risk and compliance, people and capability). More than nine and attendance and decision speed drop. Vendors and integrators should attend for specific agenda items only and should not vote.
How often should an AI steering committee meet?
Monthly is the minimum for a committee that needs to unblock work. A common structure is a monthly operating committee with delegated Tier 1 and Tier 2 authority, plus a quarterly review with the executive sponsor. Quarterly-only governance works only if a separate operations group has real delegated authority.
What is the difference between an AI steering committee and an AI centre of excellence?
The steering committee decides — it sets risk appetite boundaries, approves use cases above defined thresholds, and owns the decision log. A centre of excellence builds and supports — standards, reusable components, training and delivery capacity. Organisations usually need both, and conflating them is a common reason decision-making slows down.
Should the AI steering committee report to the board?
It should have a written escalation path to the board or board risk committee, plus a quarterly one-page assurance report. The board sets risk appetite and approves material allocation; it should not adjudicate individual use cases. A five-working-day escalation window is achievable if the submission format is fixed in advance.
What should be on the first meeting agenda?
The decision log template, the three-tier use-case definition, the approved-tool list, and the RACI table. All four are artefacts the committee needs in order to make any decision at all. Vendor presentations should not appear on the agenda until those four exist.
How do we stop the committee becoming a bottleneck?
Define tiers so that internal, non-sensitive, low-cost requests never reach the committee, and set a published decision target — five working days for Tier 2 is realistic. Track median days to decision and the number of requests returned for information more than once. Both metrics surface bottleneck behaviour early.