An AI centre of excellence in marketing is a small team that owns the shared parts of AI capability — standards, tooling, evaluation, and training — while the rest of the marketing organisation applies them in their own work. Its purpose is to raise the whole team's capability, not to do the AI work on their behalf. When it becomes a gatekeeper, it has failed, however competent it is.
Whether you need one at all depends on scale and spread. A single team of ten can share standards by sitting together. Once AI use is spread across several teams, brands, or regions, the cost of everyone solving the same problems separately becomes real, and a small central capability starts to pay for itself.
When a centre of excellence is worth it
A CoE earns its existence when the same work is being repeated across teams: the same tool evaluations, the same prompt problems, the same governance questions, the same mistakes. At that point, a shared function reduces duplication and stops each team from re-learning what another already knows.
It is not worth it when the organisation is small enough to coordinate informally, or when the real problem is that nobody is using AI yet. In that case the first job is a pilot, not a centre of excellence. Building a central team before there is demand produces a function with no customers and a mandate it cannot fulfil. The how to pilot AI piece covers that starting point.
Three operating models
The main design choice is where decisions sit. The table below sets out the three models and when each fits.
| Model | How it works | Best when |
|---|---|---|
| Centralised | One team does most AI work | Early, few use cases, need control |
| Hub-and-spoke | Central sets standards; teams apply | Scaled, many teams, mixed needs |
| Federated | Teams own their own AI work | Mature, high capability everywhere |
Most organisations end up at hub-and-spoke, because it balances consistency with speed. Centralised is the right default when you are starting and need to establish practice before spreading it. Federated only works once capability is genuinely distributed, and it tends to lose the coordination benefits that justified a CoE in the first place.
What the centre owns
The CoE owns the things that are better done once than many times: standards and guardrails, the evaluation of new tools, a shared library of prompts and workflows, training, and the governance questions that need a consistent answer. It is the keeper of the operating model's shared layer, and it should be measured on how much the rest of the organisation can do without it. The aim is the same one I made about reporting years ago: the job is not to produce more output, but to get people to the good insights faster and without repeating each other's mistakes.
It does not own the work itself. The campaign teams own their campaigns; the CoE owns the conditions that make those campaigns better. The clearest test of a CoE is whether the surrounding teams get more capable over time, or whether they simply route more requests to the centre. The first is success; the second is a bottleneck forming.
The bottleneck failure mode
The failure is easy to fall into. A capable central team is asked to help, helps well, and slowly becomes the place all AI work must pass through. Requests queue, the centre hires to keep up, and the organisation's capability concentrates instead of spreading. The CoE has become a service desk, and the operating model is worse, not better.
Avoid it by making enablement the explicit mandate and by refusing to be the default executor. Publish the standards, run the training, host the examples, and let teams apply them. If a request is really a capability gap, teach it rather than doing it. That is slower the first time and much faster every time after.
Staffing the centre
A good CoE is small and senior, not large and junior. It needs people who understand the marketing work deeply enough to know what good looks like, and who can build workflows rather than only use tools. Two or three strong people who ship and teach beat a larger team that reviews. The centre's product is capability, and capability transfers through people who can show, not tell.
Bring in the disciplines the work needs: a builder who can wire workflows and tools, an operator who knows the campaigns, and someone who owns governance and evaluation. Resist the temptation to staff it entirely with technologists; a CoE that does not understand the media and marketing reality will build standards nobody can use.
Measuring the centre
Measure the CoE on enablement, not throughput. How many teams can now do work they could not before, how much of the shared library is actually used, how many repeated questions have stopped recurring, and whether the organisation's AI output is getting better and more consistent. A centre that is busy but leaves everyone else dependent has not succeeded.
Counter-metrics matter here too. Track whether the centre has become a queue, whether teams are building workarounds, and whether the standards are being followed because they help or because they are enforced. The health of a CoE is visible in how the surrounding teams talk about it — as a source of capability or a toll gate.
A counter-metric the centre should watch
Throughput is the metric a CoE drifts toward and the one that misleads. A centre can look busy and still be failing, so pair adoption with a counter-metric: how many teams solved a shared problem without the centre's help this quarter. Track the drop in repeat questions, the share of the shared library used by teams that were not in the room when it was built, and whether the centre's own request queue is shrinking or growing. If it is growing while adoption rises, the centre is becoming a service desk; if it is shrinking while adoption rises, capability is actually spreading. That single ratio — capability up, dependency down — is the clearest signal the model is working, and it is something the centre's own workload can never show on its own.
A first 90 days
Start by inventorying what teams are already doing and where they repeat each other. Pick two or three shared problems — tooling, prompts, evaluation — and solve them once, publicly, as templates others can copy. Publish a light standard and the governance rules that go with them. Then measure how many teams adopt the shared work without being told to.
Keep the scope deliberately narrow for the first quarter and let demand pull the CoE outward. A centre that starts by owning everything will be resented; one that starts by making a few things obviously better will be asked to do more. The marketing AI policy and governance work is a natural first product, because it is needed once and used everywhere.
The CoE and the teams that resist it
Not every team will welcome a central function, and some resistance is legitimate. Teams that have already built good AI workflows may see a CoE as overhead that slows them down; teams that are sceptical of AI may see it as pressure to adopt. Both reactions carry information, and a centre that ignores them will be worked around.
The response is to start by making something obviously better for the sceptics — a tool evaluation that saves them a bad purchase, a template that removes a step they hate — rather than by mandating adoption. Earn credibility through usefulness, and let the teams that are ahead keep their autonomy. A CoE that respects the teams it serves gets adopted; one that positions itself above them gets resisted, and the resistance is usually right about something. The other trap is measuring the centre by how much work it does. A CoE that ships a lot of assets is a production team wearing a strategy title; a CoE that ships standards and capabilities is doing its job. Watch for the day the centre becomes the busiest team in marketing, because that is the signal it has drifted from enabling to executing. The healthiest version is one whose own workload goes down as the organisation's capability goes up.
FAQ
Do we need a centre of excellence?
Only if AI use is spread across enough teams that the same problems are being solved repeatedly. Below that scale, a shared standard and a few templates do the job. Build the centre when demand exists, not in anticipation of it.
Should the CoE do the work or enable it?
Enable it. The centre's job is to make the rest of the organisation more capable, and doing the work for them concentrates capability instead of spreading it. If requests are a queue rather than a learning opportunity, the model has gone wrong.
How big should it be?
Small and senior. Two to three people who can build and teach usually outperform a larger review team. Grow only when there is a genuine shared problem to own, and hire for the ability to transfer capability, not to hold it.
How do we know it is working?
Teams do more without the centre, the shared library gets used, and repeated questions stop recurring. If the centre is the busiest team in marketing, it is probably a bottleneck rather than a capability multiplier.
The short version
An AI centre of excellence in marketing owns standards, shared tooling, evaluation, and training so the rest of the organisation can apply them. Choose hub-and-spoke for most scaled organisations, keep the team small and senior, and measure it on enablement rather than throughput. The Execution guide places the CoE in the wider operating model, and the for in-house teams track works it through for a client-side team.
If you have built one, I would like to hear how you kept it from becoming a bottleneck.
Cheers, Chandler