Reading Time: 12 minutes

Quick Answer: An AI center of excellence is the team that holds an organisation’s AI agents in one place, keeps a register of what is running, and measures what each agent saves. Many organisations set up a committee that approves tools instead. Approval tells you what staff may use; a centre tells you what is working, and what to stop.

In a global poll of more than 3,400 digital trust professionals run by ISACA in 2026, 38 per cent said their organisation had a formal, comprehensive AI policy. Only 22 per cent said AI’s return had met or beaten expectations, and another 22 per cent did not know what it had returned. Policy is being written faster than anyone counts results. That gap is what an AI center of excellence is for.

Most organisations already have something they call a centre of excellence. Often it is a committee: a few senior people who meet monthly, review requests and keep a list of approved tools. A centre that holds the agents, knows who built each one and can say what each returns is a structural choice of a different kind. It exists to run everyday business agents, not to permit them.

Two Things Called an AI Center of Excellence

Most AI governance advises or records rather than runs anything. In one 2026 study of 351 organisations, which its authors call exploratory and not representative, 285 described the authority of their most senior AI governance role. About one in five of those answering said it could approve or veto AI systems; the rest called it advisory, limited to documentation and compliance, or symbolic, or were not sure. That role answers a different question from the one a centre answers.

 Approval committeeCentre of excellence
Question it answersMay staff use this tool?What is each agent returning?
Where work startsA request formConversations with each team
What it keepsA list of approved toolsEvery running agent, with an owner
What it reportsApprovals grantedHours per agent against a starting count
What a request costs a teamA wait for the next meetingA short conversation
What it can stopA purchaseAn agent that is not paying its way

The committee is not wrong; it answers an earlier question, and many organisations need both. The trouble starts when it is the only body and approval is mistaken for knowing what is running. A centre is one form of an in-house team: smaller than an AI department, closer to the work, and judged on what its agents return.

Why an Approval-Only Committee Falls Behind

Approved lists describe tools, but the work is increasingly done by agents built on top of them. In one industry survey commissioned by a security vendor, 82 per cent of 418 IT and security professionals had found AI agents they did not know about in the past year, 41 per cent more than once. Yet 68 per cent rated their visibility of agents highly, and only 21 per cent had a formal process for retiring agents. Four habits of the committee model explain why an AI centre of excellence needs more than the right to approve.

  • It approves tools, but the value sits in agents. One approved platform can hold dozens of agents built by different people, all of them off the list because none is a new tool.
  • Approval is a moment; an agent runs for years. The committee signs off once, and nobody comes back to ask what the agent returned or whether anyone still uses it.
  • The wait has a cost. In a small survey of 152 IT and security leaders, 83 per cent said security or compliance reviews had delayed AI projects, and two-thirds of those delayed waited a month or more.
  • Nothing gets switched off. Without a register or a measure, there is no point at which an agent is judged, so stalled pilots keep running alongside the ones that work.

What an AI Center of Excellence Holds and Measures

The Scottish Government offers a public example of a centre that builds rather than only approves. A ministerial update to a Scottish Parliament committee in March 2026 says its Intelligent Automation Centre of Excellence delivers the government’s AI, automation and agentic capability, owning delivery capability, shared platforms and operational assurance. A freedom of information release says it was set up in 2021 and has cost about £4 million. Neither document publishes what it has returned, the step most centres skip. Whatever its size, an AI center of excellence holds three things.

A Register Built From What People Already Run

The register starts with teams, not products. The centre asks each team what it uses and what it has built, by interviewing each team rather than sending a form. Every entry records the agent’s owner, the work it does, where it runs and its agent access rights to systems and data. Agents found in personal accounts move into the organisation’s own first, because an agent that lives in one login is gone the day a colleague resigns.

One Figure per Agent, Agreed Before It Goes Live

Each agent gets one measure, agreed with its team before launch: usually hours returned against the before figure, counted the same way each month. An organisation-wide score helps the board but cannot say which agent to keep. A running score built up agent by agent can, because every figure in it belongs to something a person can decide about.

Keep, Extend or Retire

Every agent gets one of three decisions on a fixed cycle: keep it as it is, extend it to another team, or retire it. Retiring is part of the job, not a failure. An agent that saved time last year may be redundant once a process changes, and a centre that only ever adds agents is running a collection, not a portfolio.

Who Sits in the Centre, and Who Stays Out

An AI centre of excellence can start small and grow as AI maturity rises. What matters is that it can build and decide, not its headcount. A workable first team looks like this:

  • A lead with budget authority, who reports to the board rather than only to IT.
  • One or two builders, who set up, maintain and retire agents.
  • A finance partner, who signs off the hours figure so nobody marks their own homework.
  • Risk and legal on call, consulted on agents that touch client data or decisions about people, not present at every meeting.
  • An agent owner in each team. The centre holds the agent; the team owns the work, knows the unwritten rules it follows and says when its output is wrong.

A centre that does all the building itself becomes the new queue. Its job is to make building safe and visible, then step back. Settling whether to build or buy each agent belongs here too, as does the point that an agent built for someone’s job is usually the organisation’s in law, even when it sits in a personal account.

One Law Firm, Five Agents

Take a law firm of around 450 staff across four offices. Its partners’ AI committee had approved three tools in a year. When the firm set up a small AI centre of excellence and asked each team what it actually ran, it found more agents than tools. The table the partners were shown did more for winning partners over than any policy paper.

AgentBuilt byWhere it was runningWhat the centre foundDecision
Lease clause comparisonA property associatePersonal accountUsed daily by the property team; nobody outside it knew it existedMove to a firm account; extend to the commercial team
Conflict check summariesThe risk teamApproved platformSaved real time each week; no named ownerKeep; owner named in risk
Client onboarding chaserBusiness servicesApproved platformChases identity documents and engagement letters reliablyKeep
Time entry narrativesThree partners, separatelyMixedLittle used after the first month; most drafts rewrittenRetire
Pitch pack builderBusiness developmentPersonal accountReused past pitches well, but drew on old matter detailsMove, narrow its sources, then extend

Two of the five sat in personal accounts, and one was retired within a quarter. The committee had seen none of them, because none was a new tool. Partners still sign off all client work; the centre decides only which agents keep running, and on what evidence. The same pattern holds at enterprise scale: the larger the organisation, the more agents sit between the tools on the approved list.

Setting Up an AI Centre of Excellence

Six steps cover the first quarter.

  • Ask before you approve. Start with a readiness check of what is already running, team by team, before writing any list or policy.
  • Give it a mandate to build. A centre with sign-off rights but no budget or builders becomes another committee, so settle the platform shortlist early and give builders somewhere to work.
  • Promise a turnaround. State how many days a request takes to answer, so waiting is never the reason a team goes elsewhere.
  • Agree the measure at launch. One figure per agent, checked by finance, so the result is one the board will accept.
  • Review on a fixed cycle. Keep, extend or retire every agent each quarter, and record why.
  • Report in hours, not approvals. The board should see what agents returned, not how many requests were cleared. Smaller organisations can run the same loop with a lighter playbook and one part-time lead.

Where GrowthNation Fits

GrowthNation’s AI agent interviews every team, by voice or text and in confidence, to find the work worth handing over and the agents already running. The agents it builds are served centrally and belong to the organisation, not to one person’s account. Hours saved are tracked per agent against the starting count, the figure behind every keep, extend or retire decision. If you need only a policy and an approved-tools list, a committee may be enough; GrowthNation is for a centre that has to produce working agents and show what they return.

Give Your Centre Its First Register

Ask one team to list every AI agent it ran last week, and who built each one. Most cannot answer in full, which is the place to begin. An AI center of excellence starts with that list, not with a policy, because nothing can be measured or retired until it is known.

Book a walkthrough and GrowthNation will interview one team, then hand you three things:

  • The agents already running, and who built them. Including any in personal accounts, moved where the organisation can keep them.
  • The hours each one returns. Counted the same way every month, agent by agent.
  • A working agent, held by your organisation. Built from the interviews, owned centrally and ready for its first review.

Hours are measured against where you started, so the saving can be checked.

Book a walkthrough: one team, one register, one working agent. No deck.

Join the GrowthNation waitlist. New organisations are taken on in batches, so joining now puts your team in line for the next slot.

Frequently Asked Questions

What is an AI center of excellence?

An AI center of excellence is the team that holds an organisation’s AI agents in one place, keeps a register of what is running and who owns it, and measures what each agent returns. It decides which agents to keep, extend or retire, which is what separates it from an approval committee using the same name.

What is the difference between an AI centre of excellence and an AI governance committee?

A governance committee sets policy and decides which tools staff may use. A centre of excellence runs the agents built on those tools: it records them, measures them and decides their future. Many organisations need both.

Does a mid-sized organisation need an AI centre of excellence?

Not a large one. A lead with budget authority, one builder and a finance colleague who checks the figures can find what is running, hold it centrally, measure each agent and review it every quarter. Authority to build and to retire matters more than size.

Who should lead an AI centre of excellence?

Someone with budget authority who reports to the board, rather than only a technical lead. The role decides which agents run and on what evidence, so it needs standing with department heads. Pairing the lead with a finance partner who signs off results keeps the figures credible.

Should an AI centre of excellence build agents or only set rules?

It should be able to build, or it becomes another approval queue, but it should not build everything. The centre holds, maintains and measures agents; each team owns the work an agent does and says when its output is wrong.

How do you know a centre of excellence is working?

Look at hours returned per agent against a count taken before launch, agents retired as well as added, and how long a team waits for an answer. A centre reporting only approvals or agents built is measuring activity, not results.