What should it actually produce?

The answers nobody should have to work out twice: which tools are approved and for what material, what a review will ask, patterns that worked, and a fast route for a team that is stuck. Those are cheap to produce once and expensive when ten teams each derive them separately.

How big should it be?

Small enough that it cannot take on delivery, because a team with capacity to deliver ends up delivering and then has a backlog. The constraint is doing its actual job rather than modesty, and it is the first thing to go when the function appears to be succeeding.

Not until several teams are doing this independently and answering the same questions differently. Before that it is overhead, and one team's own experience serves better. The signal is repetition across groups rather than the amount of AI activity in total.

Glossary

AI centre of excellence

A centre of excellence is a small central team that sets standards and supports AI work across an organisation, without owning every project itself.

In plain terms

A few people in the middle who know more about this than everybody else and whose job is to make everybody else better at it. They decide what is approved, help teams who are stuck, and keep the things that should only be worked out once from being worked out ten times. What they are not is the department that does all the AI work.

Why it matters

Because the alternative is every team solving the same problems separately, badly and at different standards, and because the obvious remedy of centralising the work is worse than the problem. Getting the shape right is what decides whether a central team accelerates an organisation or becomes the queue everything waits in, and the two outcomes look identical for the first few months.

How it works

The useful work is the part nobody should repeat. Which tools are approved and why, what a review process asks for, patterns that turned out to work, the answers to the questions every team reaches eventually. Producing those centrally and giving them away is worth more than any individual project the same people could have delivered.

It fails in two directions and the failures are opposite. Too much authority and it becomes a gate everything queues at, which teams route around exactly as they route around procurement. Too little and it becomes an advisory group nobody consults, producing guidance that is not read. The workable position is narrow and has to be held deliberately.

Staying small is what preserves it. A team that grows takes on delivery, and a team doing delivery is a department with a backlog rather than a support function, at which point the organisation has centralised the work it meant to enable. The size limit is not modesty; it is the mechanism that keeps the role intact.

It works when it is a route rather than a checkpoint. Teams should want to come because it is faster than not coming: a question answered in a day, a pattern already worked out, a supplier already reviewed. Where the incentive runs that way the standards spread on their own, and where it does not no amount of mandate produces them.

The narrow position, and the two ways off it

Both ends of this line are stable and the middle is not, which is the honest difficulty with the whole idea. A team drifts towards the left when it has no way to make itself useful quickly, and towards the right whenever something goes wrong and the organisation responds by requiring its involvement. Neither drift feels like a mistake at the time; the right-hand one in particular is usually a reasonable reaction to a genuine incident. Holding the middle takes continuous effort and one specific discipline, which is measuring the team on how quickly it unblocks somebody rather than on how much it has reviewed. That measure pulls against both drifts at once, and it is the closest thing to a mechanism available for a position that otherwise depends entirely on the judgement of whoever is running it.

Seen in the wild

Deciding centrally which assistant is approved for which material, so each team does not negotiate the question separately.
ChatGPT
Keeping the working patterns for automation in one place, so a team building its fifth workflow does not start from nothing.
n8n
Owning a deployment that reaches across systems, since access questions of that kind cannot sensibly sit in one department.
Glean

Common misconceptions

People assume

It should do the AI work.

In fact

Then it is a delivery department with a queue, and the teams closest to the work are further from the tools than before. The point is to make other people effective, which means giving away everything it learns rather than accumulating the interesting projects.

People assume

It needs authority to be effective.

In fact

Authority makes it a gate, and gates get routed around exactly as procurement does when it is treated as an obstacle. What works better is being genuinely faster to consult than to bypass, at which point the standards spread because using them saves time rather than because using them is required.

Telling them apart

Centre of excellence vs AI strategy

Centre of excellence

Who helps, and where the reusable answers live.

AI strategy

What the organisation has decided to do and not do.

One is a team, the other a document. A central team without decisions to apply mostly writes more documents.

Questions

What should it actually produce?: The answers nobody should have to work out twice: which tools are approved and for what material, what a review will ask, patterns that worked, and a fast route for a team that is stuck. Those are cheap to produce once and expensive when ten teams each derive them separately.
How big should it be?: Small enough that it cannot take on delivery, because a team with capacity to deliver ends up delivering and then has a backlog. The constraint is doing its actual job rather than modesty, and it is the first thing to go when the function appears to be succeeding.
Do we need one?: Not until several teams are doing this independently and answering the same questions differently. Before that it is overhead, and one team's own experience serves better. The signal is repetition across groups rather than the amount of AI activity in total.