An AI steering committee should decide which initiatives receive shared money and people, which cross-department conflicts need executive resolution, and when a proposed expansion changes the company's commitments. It should also settle exceptions within its delegated authority and send other decisions to the people who can legally and operationally make them.
Give the committee a choice that someone in the room can act on. Routine draft checking, technical fixes and delivery scheduling should remain with the teams responsible for that work. Bringing every case to a senior meeting can leave both the committee and those teams waiting.
The wider AI adoption strategy sets the direction. The steering group turns competing proposals into explicit commitments, including decisions to defer or stop work.
Define what the group can actually authorize
“Steering committee” does not itself grant purchasing, data-access or risk-acceptance powers. Before adding meetings, identify the decisions that exceed a workflow owner's authority and establish who can make them.
A useful remit can cover:
- Portfolio priorities. Choose between initiatives competing for the same budget, specialist time or operating capacity.
- Shared commitments. Resolve who will fund support, release staff and maintain a capability used by several departments.
- Expansion or withdrawal. Decide whether evidence and resources justify a larger commitment, a revised trial or an orderly end.
- Unresolved exceptions. Decide only those exceptions within the group's mandate. Refer the rest to the required authority and record the boundary.
NIST's voluntary AI Risk Management Framework calls for clear responsibilities and communication, executive responsibility for AI risk decisions, and risk-management activity proportionate to organizational tolerance. It also addresses safe disengagement when a system's performance or outcomes conflict with its intended use. It does not prescribe a steering committee or a meeting schedule.
The committee's mandate should connect to the adoption leader's charter. If the lead can already schedule a trial within committed capacity, leave that decision there. If two departments claim the same specialist, the lead may need an executive choice before either trial can start.
Compare proposals against the same capacity
Approving each proposal separately can hide the fact that they need the same person. Ask what the company would have to delay, fund or stop to make room.
Suppose an equipment rental company has two AI proposals. The service team wants to draft internal handover notes from approved technician records. Customer services wants to draft answers from the company's rental terms. Both need the systems specialist who also supports the checks on returned equipment.
The operations manager, Claire Bennett, brings both proposals to the steering group. The specialist, Sam Taylor, has capacity to prepare one trial alongside existing duties. Neither proposal includes authority to automate safety decisions or send unchecked customer messages.
Compare the proposals on the conditions needed to do the work, rather than the polish of their demonstrations.
| Proposal | What is ready | What needs a decision |
|---|---|---|
Internal service handover notes | Approved source records, a service owner and a technician check of each draft | Whether to allocate Sam's available trial capacity to this work first |
Customer response drafts | A customer services owner and an existing manual response process | Who will reconcile conflicting rental terms before a drafting trial can use them |
Start both together | Two interested teams | What existing work would move, and who would provide the additional capacity |
In this example, the group chooses the handover trial first. Customer services keeps its manual process while its manager resolves the rental terms. The group sets a date to reconsider that proposal once the source material and support plan are ready.
This is a sequencing decision, not a finding that internal work is always safer or more valuable. The service owner still has to assess whether handover drafts are accurate and whether checking them makes the complete task easier. Use whole-work measurement before turning an attractive draft into a claim of benefit.
Deferral needs a reason and a return condition. “Come back next month” is unhelpful if nobody knows what must change. Equally, protecting an existing safety check is a valid capacity constraint, not evidence that a team is resisting AI.
Separate advice from approval
A committee can coordinate expertise without being the final approving authority. Australia's AI Review Committee terms of reference provide a concrete contrast. For specified government use cases, including those assessed as having high residual risk, it provides non-binding advice. Its terms explicitly say that advice is not endorsement or compliance advice.
Your company's arrangement may differ. Make the distinction visible in the decision record:
- Advice received. What the specialists recommend, including unresolved objections.
- Resource decision made. What the authorized sponsor or group commits, and within which limits.
- Approval still required. Which security, data, legal or operational decision must occur before the proposed use proceeds.
A majority vote cannot substitute for a required approval. Nor should a steering group mark an initiative green simply because nobody objected during the meeting. If a necessary reviewer has not assessed the proposal, that condition remains open.
In the rental-company example, funding Sam's time does not authorize new customer data, automatic messages or changes to equipment safety procedures. Those are separate decisions. Keep them with the relevant owners and reviewers.
Ask for a decision before accepting a status update
In a September 2026 LinkedIn article and its discussion, data leader David Krauza argues that committees often review work already owned elsewhere while blocked funding and cross-function choices stay off the agenda. Olga Maydanchik raises a useful qualification in the replies: a data-quality scorecard can belong there when it supports an investment decision. These are practitioner perspectives, not a controlled test of committee effectiveness.
The distinction is the decision attached to the material. A report on inconsistent rental terms may be a routine update. The same report may justify a choice to fund reconciliation before customer-response drafting. Keep the evidence that informs that choice; avoid making executives supervise individual edits.
For each proposed agenda item, ask the presenter to state:
- The choice needed. What cannot proceed within existing authority or commitments?
- The realistic alternatives. Include continuing the current process, deferring or stopping where relevant.
- The consequence. What work, money, exposure or affected people change under each alternative?
- The action after the meeting. Who implements the decision, and what condition would bring it back?
The ownership guide helps assign implementation responsibility. A committee decision should release an owner to act, with an agreed boundary, rather than make the group the owner of every task.
Examine what is waiting between meetings
The committee's usefulness depends partly on what happens outside the room. Give routine proposals a documented route within delegated limits. Define which changes require escalation, such as a new data source, a larger funding commitment or a use that affects another department.
Then inspect unresolved requests. Record when a decision was first needed, what information is missing, who can decide and what work is waiting. Distinguish three problems:
- Missing evidence. Assign the investigation and state what would make the proposal decidable.
- Unavailable capacity. Choose a smaller commitment, provide capacity or defer the work explicitly.
- Unclear or contested authority. Obtain a decision from the person who can settle the boundary.
Choose the meeting frequency around those decisions. One Reddit user described a monthly cross-functional committee involving technology, marketing, operations, risk and finance, with a CTO sponsor. The account does not establish its results or make monthly meetings a benchmark for other companies.
Agree a response deadline and deputy for decisions that cannot wait until the next ordinary meeting. An unanswered request is still unanswered, not permission to proceed. For incidents, use the established urgent response route rather than wait for portfolio discussion.
Before the next meeting, take one delayed request and identify the specific choice this group can make. If it cannot make that choice, send it to the correct authority. If every item can be handled within existing commitments, use a shorter update or cancel the decision meeting.
Questions and answers
Does every AI use case need steering committee approval?
No. A company can delegate routine use within documented tool, data, budget and workflow limits. A use case needs wider review when it crosses those limits or creates a commitment the local owner cannot make. Write down the escalation triggers and required approving roles before teams begin. The authority boundary is a starting point, not blanket permission for low-cost tools.
Who should attend an AI steering committee?
Include people authorized to settle the choices on its agenda, with the relevant business, finance, technology and risk expertise. Invite a workflow owner or specialist when their evidence is needed; obtain input from affected employees as appropriate. Titles alone do not establish decision authority. Test membership against one real shared-capacity choice, then identify who can commit resources and who must provide a separate approval.
How often should an AI steering committee meet?
Set the cadence around the frequency and consequence of decisions that need the group. There is no universal monthly or quarterly schedule established here. Look at the requests waiting between meetings, agree response deadlines and provide an urgent route for matters that cannot wait. Repeatedly discussing a blocked request is a reason to examine evidence, capacity or authority, rather than simply add meetings.
Can the committee override a security or legal objection?
Only the properly authorized decision-maker can accept a risk or approve a use, and applicable requirements still constrain that decision. The committee's name or majority vote does not create that authority. Record the objection, the advice received, who must decide and which conditions remain unresolved. The advice and approval distinction helps prevent a funding decision from being mistaken for permission to deploy.
