The requests land individually. Someone in operations wants to trial an AI contract reviewer. A department head has already signed up for three tools and is asking for retroactive approval. Finance wants Copilot for the team. Legal has concerns. Each request takes a morning, or a week, and then the next one arrives. You become the person standing between your organisation and every AI decision it needs to make. That is the bottleneck the steering group exists to resolve.
What is an AI steering group that actually decides?
An AI steering group that decides is a small, standing body with genuine authority to approve tools, retire tools, set guardrails, and close out intake requests. Small means three to five people, including the delegate and at least one member with budget authority. Its mandate covers saying yes, saying no, and recording which it chose and why. That authority is the whole point.
The distinction that matters is decision authority versus advisory authority. An advisory committee can surface good analysis but cannot close the loop. Requests come in, get discussed, get passed up or deferred, and the queue grows. A group with genuine authority closes requests on the day they are reviewed, or commits to a specific timeline with a named owner.
Spencer Stuart’s research on AI delegation makes this concrete. When founders delegate AI but retain sign-off on every tool choice, they create a governance structure where the delegate advises but cannot act. The alternative is a forum with a defined scope, a documented mandate, and the power to act within it. That power must come from above, confirmed clearly at the outset, and not gradually eroded when decisions become uncomfortable.
Why does the delegate-alone model create a bottleneck?
When one person holds all AI approval authority, every request queues behind their attention. Approvals slow, departments start working around the system, and shadow AI grows in the gaps. The delegate is simultaneously responsible for moving the business forward on AI and for slowing it down by being the single point of friction on every individual decision that crosses their desk.
The Korn Ferry research on AI leadership readiness notes that operators handed AI mandates often lack the supporting structures to make good decisions at speed. When approval volume outpaces one person’s bandwidth, quality drops. Requests get rubber-stamped to clear the queue, or held too long while the delegate gathers confidence that never quite arrives.
The other failure mode is departments that stop submitting requests altogether. They discover the process is slow and uncertain, so they solve the problem themselves. A free trial becomes a paid subscription. A workaround becomes embedded in workflow. By the time the delegate finds out, the tool is load-bearing and the conversation about approval is theoretical. BCG’s research on AI adoption identifies unclear governance as a consistent driver of higher rates of unmanaged shadow AI in organisations. The steering group is the structural answer to that problem, but only if it has the authority and the process to act on what comes through.
Where does the group do its real work?
The artefacts are where the group’s real work sits, not the agenda itself. The three that matter are the AI tool register (what has been approved, what has been retired, and why), the risk log (open issues, mitigations, and named owners), and the intake queue (incoming requests from departments, triaged before the session begins). The meeting exists to resolve them.
The standing agenda flows from these artefacts rather than being designed fresh each time. A typical session works through the intake queue first, reviews any open risk items, and then runs a keep-or-kill assessment on any tool that is up for review. The bulk of preparatory work happens before the meeting, with intake items pre-read, the risk log updated, and the register current. The session resolves rather than deliberates.
The intake queue carries a specific weight here. It is the front door that changes the relationship between the group and the rest of the business. Once departments know there is a defined channel for AI requests, the informal one-off approach to the delegate stops being the default. A CIO planning framework describes this precisely as standing up “a lightweight programme that owns AI across the company plus an intake process for new AI projects so departments don’t each freelance.” That shift from informal to formal is the practical outcome of getting the intake right, and it is the change that most directly reduces the delegate’s approval burden.
When should it convene, and how does it avoid becoming theatre?
A steering group that convenes monthly with no standing agenda, no pre-circulated material, and no recorded decisions is a meeting, not a governance forum. The pattern that earns its place meets fortnightly or monthly based on intake volume, circulates the agenda at least 48 hours before the session, runs to a hard time cap, and records every decision with a named owner before anyone leaves.
The EY Board Matters research on AI governance frames this at board level, but the operating principle carries down. Governance forums require rigour in how decisions are recorded, not just notes that summarise the conversation. Ownership needs to be explicit. Schellman’s research on AI implementation failures identifies unclear decision ownership as a recurring driver of deployment problems in real organisations. If the group meets and disperses without naming who is accountable for each outcome, the accountability sits nowhere.
Scope clarity matters as much as meeting discipline. The group should have a defined boundary for what it can decide without escalating. Approved-tool additions, risk-item closures, and tool retirements should sit within that scope. Questions about whether to start or stop the AI programme, whether to hire a permanent AI lead, or how to restructure the function belong above the group. Getting that distinction right means the group acts quickly on what it is built to handle, and the delegate does not spend session time relitigating strategic calls that should have been settled elsewhere.
What does this sit alongside in a working AI governance structure?
The steering group’s job is to decide. The adoption work, building enthusiasm in departments, and running tool training sit with the champions network, the named people in each function who know the approved toolkit and help colleagues use it day to day. The delegate connects both layers, attending or chairing the steering group and commissioning the champions to carry adoption into the business.
The Harvard Law School Corporate Governance Forum’s research on AI risk disclosures found that reputational risk is the top AI concern cited, with boards increasingly expecting formalised oversight structures. For owner-managed businesses, the structure is lighter, but the accountability principle holds. Someone needs to be clearly responsible for what AI is doing in the business, and that accountability needs a structure to sit inside. PwC’s AI Predictions research identifies governance gaps as primary risk factors across the adoption lifecycle, a finding that applies as readily to a 60-person professional services firm as to a listed company.
The steering group provides the formal accountability layer. The champions network handles informal adoption. The delegate holds the thread connecting both. Remove any one of the three and the others lose coherence. The group can approve tools that nobody uses well; the champions build energy around tools the group has not properly vetted; the delegate tries to manage everything alone and the bottleneck re-forms. All three working together is what a functioning AI governance structure looks like in an owner-managed business.
Build the group small, give it a documented scope and the authority to act within it, tie every session to the register, the risk log, and the intake queue, and record every decision with a named owner before the room empties. When departments know there is a defined channel, they use it. When the group has genuine authority, it acts. That is the version worth building.
