How to set up an AI centre of excellence led by domain experts, not engineers
What a centre of excellence is for
An AI centre of excellence, or CoE, is a small, named group with a standing remit. It decides how the organisation uses AI, picks and proves the first use cases, sets the guardrails and helps everyone else adopt what works. The idea is to concentrate scarce judgement early, when nobody yet knows which uses are worth the risk.
The point of having one is durability. Launching an AI pilot is easy now. Keeping it alive once the novelty wears off is the hard part. Gartner's June 2025 research found that only 45% of organisations with high AI maturity keep their AI projects operational for at least three years. So even among the organisations that are supposedly good at this, more than half still can't make it stick. A centre of excellence is meant to close that gap, which is why how you staff it matters more than what you call it.
The default blueprint staffs for the wrong problem
Search for how to build one and the big vendors all give you much the same answer. Microsoft's Cloud Adoption Framework, for instance, sets out a multidisciplinary CoE. It spans data scientists, machine-learning engineers, governance, security and platform specialists, with business representatives alongside them. That is sound advice for the company it was written for, one that builds and runs its own models and infrastructure.
Most companies are not that company. They buy AI off the shelf: Claude, ChatGPT, Copilot and whatever their existing software vendors are bolting on. The job is rolling it out across their own teams and getting people to use it well on everyday work. Staff a centre of excellence with engineers for that and you have hired specialists for a problem that is mostly about judgement and working habits. Two things follow. The engineers spend their days on requests they are overqualified for, and everyone with a real use case queues up behind them. The centre of excellence becomes the bottleneck it was supposed to remove.
Why domain experts should lead it
The people who know where AI will pay off are the ones who own the work. Your finance lead knows which reconciliations eat the month-end. Your product manager knows which research never gets done, and your operations manager knows which handoffs break. They can tell a genuine improvement from a slick demo, and an engineer parachuted into their process can't.
The old objection was that these people couldn't build anything themselves. They had to hand their ideas to engineers and wait. That objection has mostly gone. With tools like Claude and Claude Code, a properly trained domain expert can turn a described problem into a working prototype. They don't have to write production code from scratch. The constraint has moved from the tooling to the training.
We have seen it happen. With a training partner, we co-designed and delivered a six-week programme for around twenty senior product managers and UX designers at a Fortune 500 tech company. They were domain experts, not engineers. The sessions were hands-on and used their own PRDs, user research and prototypes. Over the six weeks they went from first prompts to building reusable assets with Claude and Claude Code. For a centre of excellence, note where the capability ended up: with the people who owned the work. That is exactly where you want it.
Who belongs in a lean centre of excellence
You don't need a department. For most mid-sized organisations, five to seven people is enough to start, and who they are matters more than how many. Keep it dominated by people who run the work day to day, with just enough specialist support to keep it safe and accountable.
| Role | Why they are there |
|---|---|
| A single accountable owner (often the Head of AI, or an executive sponsor) | Someone whose job is to set direction, unblock decisions and carry the remit upwards. Without one, the group stalls at the first disagreement. |
| Two to four domain experts from the priority areas | Most of the group. They bring the use cases, judge what is genuinely better and take the capability back to their teams. They need training that gets them building. |
| One data or security voice | Part-time is fine. They settle what data can go where, and who checks output before it leaves. Their role is a guardrail, not a gatekeeper. |
| A change or L&D lead | Adoption is a behaviour-change problem as much as a technical one. Someone needs to own how the rest of the organisation learns what the group proves. |
There is no standing team of engineers on that list, and that is deliberate. Bring engineering in for the specific build that needs it, instead of making the whole centre of excellence wait on their calendar. Still deciding whether to hire or grow the person at the top? We've written separately on what a Head of AI does and how to hire one or grow one internally.
What it should own in the first 90 days
New centres of excellence most often fail by trying to govern everything before they have proved anything. Keep the first quarter narrow and concrete.
- A shortlist of two or three use cases, picked because a domain expert in the room is confident they are real, not because they sound impressive.
- A one-page set of guardrails: which data can go into which tools, what must never go in, and who signs off output that leaves the building.
- A baseline for each use case, taken before anyone changes anything, so later you can tell whether it worked instead of guessing.
- A small shared library of the prompts, commands and worked examples the group builds, so the next team doesn't start from a blank page.
- One honest review at the end that reports what changed, including the use cases that didn't pan out.
That's a quarter's work for a part-time group, and it leaves you with something to point to. Our piece on why most AI pilots fail covers the failure modes this sequence is built to avoid. For more on the baseline, see how to measure AI ROI without fooling yourself.
How to stop it becoming a bottleneck
The standard advice is to start centralised, with the centre of excellence doing the work. As the organisation matures, you shift to an advisory model: the group sets standards and coaches while the teams build their own. That sequence is right, but most organisations leave the switch too late. The central group then hardens into a permanent approval queue.
A group led by domain experts makes the switch easier, because its members already belong to the teams that will own the AI work later. Each of them can double as the AI champion for their own team. Part of their job from day one is to make themselves unnecessary: prove a use case, write down how, hand it back, move on to the next. If the centre of excellence is still building everyone's solutions in year two, it has failed at its real job, even if every individual project succeeded.
Resourcing it: train the experts you already have
Finding domain experts is the easy part, because you already employ them. The hard part is getting them fluent enough to build with AI, which takes more than knowing how to use a chatbot. That's a training problem, and a specific one. It needs to be hands-on and built on the team's own work. It also needs to cover when not to use AI, as well as how.
That's the work we do. Our UK AI training is scoped around each team's own tasks. That is how a domain expert becomes someone who can carry a centre of excellence. If you're setting one up, book a discovery call and we'll talk through who to put in the room and how to get them building.
A note on credentials. Fautons' own AI certification is currently waitlist-only, so don't set up your centre of excellence expecting a badge to do the work. What holds, in our experience, is the capability itself, and the evidence is what the group ships. If you need to prove the training worked, we've written about how L&D leaders measure what stuck.
Frequently asked questions
What is an AI centre of excellence?
An AI centre of excellence is a small, named group with a standing remit. It sets the standards for how the organisation uses AI. It also chooses and proves the first use cases, defines the guardrails for data and output, and helps everyone else adopt what works. The aim is to concentrate scarce judgement early, before anyone knows which uses are worth the risk.
How many people should be in an AI centre of excellence?
For most mid-sized organisations, five to seven is enough to start. Who is in it matters more than the headcount. Most of the group should be domain experts who own real work. Around them you need a single accountable owner, a part-time data or security voice and a change or learning lead. You rarely need a standing team of engineers. Bring engineering in for the specific builds that need it.
Should an AI centre of excellence be led by engineers?
Usually not, if your organisation buys AI tools rather than building its own models. The people who can tell a real improvement from a demo are the ones who own the work. With tools like Claude, a trained domain expert can prototype without writing production code, so the constraint has moved from engineering capacity to training. Keep engineers on call for the specific builds that need them.
How long does it take to set up an AI centre of excellence?
You can set up a useful one in a quarter if you keep the remit narrow. Pick two or three use cases, write a one-page set of guardrails and take a baseline before you change anything. Then keep a shared library of what the group builds, and hold an honest review at the end. The common failure is trying to govern everything before proving anything.
Does an AI centre of excellence need a Head of AI?
It needs a single accountable owner. That is often a Head of AI, but it can be an executive sponsor who gives the group its mandate. Without someone whose job is to set direction and unblock decisions, the group tends to stall at the first disagreement. Whether you hire that person or grow them internally is a separate question, which we cover elsewhere.
How do you keep a centre of excellence from becoming a bottleneck?
Treat it as temporary scaffolding. Start with the group doing the work, then deliberately shift to an advisory model where it sets standards and coaches while teams build their own. A group led by domain experts makes this easier, because its members already belong to the teams that will own the work later. Part of their job is to make themselves unnecessary.
