What an AI Center of Excellence is actually for

The AI Center of Excellence usually starts with a mandate and a slide. Leadership has decided the organization needs to "get serious about AI," someone capable is put in charge, and a small team forms with a name that promises a lot. Excellence. Center. The words do work before anyone has done anything.

Six months later, most of these teams have quietly become one of two things. Either a documentation team that maintains a wiki of prompts and guides that fewer people open each month, or a governance team that writes policies about what you may and may not do with AI. Both are busy. Both produce artifacts. And both have drifted away from the one job the CoE was created to do, without anyone deciding they should.

This piece is about what that job actually is, and why it is easy to lose.

The job, stated plainly

A Center of Excellence exists to make a good way of working spread across an organization, and to be able to show that it did. That is the whole remit, and every part of it is load-bearing.

A good way of working, not a pile of tools or a training catalog. The CoE is not there because people cannot access AI. Access is a solved problem. It is there because access alone produces a thousand private, inconsistent approaches, and someone has to turn the best of those into a standard.

Spread across an organization, not documented in one place. Writing the standard down is a step, not the finish line. A standard that lives in a document the right people never open has not spread. It has been filed.

Show that it did. A CoE that cannot demonstrate impact is one budget cycle from being reorganized into nothing. The proof is not optional. It is the difference between a function leadership invests in and a cost center leadership tolerates.

Hold those three together and the drift becomes obvious. A documentation team does the second thing badly and skips the third. A governance team does neither and mistakes control for progress.

Why it drifts into a documentation team

The pull toward documentation is strong because documentation feels like the job. Someone works out a good way to draft a board update or triage a support ticket, and the natural instinct is to capture it: write it up, add it to the library, tag it well. The CoE becomes the team that keeps the library tidy.

The problem is that a library is storage, and storage was never the constraint. The good prompt exists. It is written down. And the forty people who do that same job across the organization still do not use it, because it sits in a space that belongs to the team that wrote it, and they will never go looking. The library grows. Adoption does not. This is the circulation gap, and a documentation team is optimized to make it worse: more content, filed more neatly, opened by no one new.

Why it drifts into a governance team

The other drift is toward control. AI raises real risks, so a natural role for a central team is to set the rules: which tools are approved, what data may go where, what is forbidden. This is necessary work. It is also not the CoE's reason to exist, and it is a comfortable place to hide.

Governance is comfortable because it is legible and defensible. You can show a policy. You can point to a review process. What you cannot show is a single person doing their job better because of it. A CoE that becomes a governance team has traded the hard, ambiguous job of making practice spread for the easy, concrete job of writing rules about it. The rules matter. They are not the mission.

The one thing a CoE cannot fake

Here is the test that cuts through both drifts. A Center of Excellence can produce documents nobody reads and policies nobody feels, and still look busy. What it cannot fake is this:

Can you point to a specific way of working that started in one part of the organization, and show that it is now used by people doing that same work somewhere else, who did not invent it and were not told to adopt it?

That is spread, and spread is unfakeable. You either have examples or you do not. A team copying a practice they were mandated to use does not count, because compliance is not adoption. A prompt used heavily inside its origin group does not count, because that is not spread, it is just local habit. The thing that counts is a practice that traveled to the people who do that kind of work, on its own, because it reached them and it was good.

If a CoE can show that, and can show it across several kinds of work, it is doing the job. If it cannot, no amount of documentation or governance changes the verdict.

What a good CoE actually owns

The shift is from owning the library to owning the loop. A CoE that is working owns four things, and none of them is a wiki.

• Capture that is a byproduct of the work, so the good stuff surfaces without anyone being asked to donate their time to the program.
• Promotion, the act of turning a proven individual practice into the standard for a kind of work, based on evidence that people are already reaching for it rather than on who lobbied hardest.
• Circulation by kind of work, routing the standard to everyone who does that job wherever they sit, which is the move a team-scoped or department-scoped library structurally cannot make.
• Evidence, the measures that let you show reuse, spread, contribution, and coverage to the leadership that funds you.

A CoE built around that loop looks different from one built around a library. It spends less time writing content and more time watching what people actually adopt, promoting the good of it, getting it to the right people, and reporting where it landed.

That is what the name was always supposed to mean. Not a center that holds excellent documents, but a center that makes an excellent way of working spread, and can prove it did. If you want to see where your own program sits on the path from documented to spread, the maturity assessment is a three-minute read on exactly that.