We are your AI Center of Excellence

Intake in this seat is not a form on a share drive. It is the live answer to what counts as AI work, who may start it, and what happens when they start it anyway. When a business unit wants a copilot, a vendor offers a platform, or a team has already shipped a prompt into production, the CoE decides whether that work enters, reuses something we already have, or is refused. We hold that with the operating owner: a short queue with reasons, a pattern to reuse, and a date we will know if it is working. If it is not, the unit kills it.

Most companies already have AI standards. They will not call them that. A page in Confluence, a security review that happens if someone remembers, an eval that was run once. The CoE treats standards as a short list the unit will still be in the room to enforce: what “good” means, what must be evaluated on real cases, what must be logged, when a human has to see it, what is forbidden. A project produces a binder. A unit runs a standard — exceptions, reviews, and the uncomfortable no when someone wants to skip it.

The second team always wants to rebuild. The vendor always wants to be the platform. The CoE exists so the company has a memory: we already have a support pattern, a retrieval pattern, an intake pattern. Reuse is a decision, not a suggestion. Pointing at the pattern and refusing the duplicate is how spend stops multiplying. If nobody can name what we already have, you do not have a CoE. You have a collection of first times.

A pilot without a kill date is a permanent employee with worse benefits. The CoE puts criteria on anything that is not yet a decided capability: what “working” means, who judges, what it may cost, when it dies. Killing work is a success when the criteria said it should die. The unit takes the heat so the requesting team does not have to. If no one has ever killed an AI pilot, you do not have a CoE. You have a museum of hopes. Deeper AI governance and compliance work exists as its own service. On this page, kill-criteria are simply how the unit stays honest.

Enablement here is not a curriculum we were paid to write. It is the CoE sitting with the people doing the work: this is the pattern, this is the standard, this is why last week’s request was refused. People get fluent because the unit is still there next month. If you need a workshop series, that is training and enablement. If you need the people who run the door to also make the rest of the company better at using it, that is this seat.

A Chief AI Officer owns the agenda with the CEO and the board: spend at the top, vendors at the top, the story the board hears. The CoE owns the operating door underneath that story. You can have a CAIO and still need a CoE. You can have a CoE and not yet have a CAIO. We sell both, on different pages, as different seats. Do not buy this page if what you wanted was the officer. Do not buy the officer if what you wanted was the queue.

The job happens on your calendar. We join the intake review, the operating forum where AI work is accepted or sent back, and the sessions where a pattern is about to be rebuilt or a pilot is about to become a zombie. We do not invent a steering committee so we have somewhere to present. We arrive having read the queue, the standards, and last week’s kills, and we use the meeting to decide.

A CoE without rights is consulting with a better acronym. Before we start, we write the mandate: what the unit can accept, reuse, refuse, or kill; what goes to the CAIO, the CIO, or the CEO; spend and vendor thresholds that belong to the unit versus the officer. You cannot hold a CoE responsible for a queue it was not allowed to touch. If you want a smart guest, do not buy this. If you want the unit, the unit has to be able to say no, and have it stick.

If the same pattern was rebuilt, a request sat for six weeks with no owner, a standard was skipped in public, or a pilot that should have died is still burning cloud, that is our problem. Units do not get to say “we recommended.” They decide, and they live with it. Accountability is also the weekly record — decisions, owners, dates — so the same people who made the call are still there to own it.

A pass through the intake queue. At least one reuse or refuse. At least one kill-criteria conversation on something that is not yet a capability. Time with the people doing the work — that is the enablement. A short written trail. And the unscheduled thing, which is why you wanted a unit instead of a playbook. Some companies need a day a week. Some need more around a board cycle or a wave of vendor renewals. We set that in the open.

Week one is the mandate: decision rights, the queue, what we can refuse, the name on the standard. We meet the operating owner on their cadence. We take the keys that exist — requests, vendors, shadow work, the Confluence page nobody uses — the way a finance team takes a close, not as researchers gathering quotes. By month’s end the door is open, leadership can see the queue, and we have made at least one real refusal or kill so the unit is not ceremonial. We will not run an interview tour, publish a maturity score, or produce a backlog for our engineers.

There is a standing intake. The queue is shorter, because some things were refused and some things died. There is a reuse list someone will defend. Standards are a habit: a review, an exception, something stupid stopped before it is public. Enablement is happening on the work in the queue, not in a classroom. At 90 days we can tell leadership what entered, what reused, what we killed, and what we will decide next. If we cannot, we have been busy, not in the unit.

The CoE should now be boring in the way a good finance close is boring: cadence, rights, a queue, fewer surprises. New ideas go through the door. Duplicates get pointed at the pattern. If the company has grown into a full-time bench, we will say so. If it has not, we stay — that is a success, not a failure to convert. One hundred and eighty days is not the end of a project, and it is not the moment we reveal systems we would like to build.

A COO, CIO, or operating partner who wants an AI operating unit and will give that unit rights on intake, reuse, and kills. A company that already has AI in the building — licenses, pilots, vendors, shadow work — and no door that owns the whole of it. A mid-market or growth-stage company that knows the public Head of AI band is not this year’s hire, and is done pretending a working group is the same thing. A company that tried “establish a CoE” with a large firm and still has the same unowned queue. A company that will let us sit with the people doing the work, not just the innovation team.

It is also for leadership that wants the truth more than the theater. If you want a unit that will refuse work, kill pilots, and tell the operating owner the unvarnished version, we can do that. If you want a Center of Excellence for the annual report, we cannot.

Anyone shopping for a custom-build, an “AI OS,” or a development partner and using CoE language to start the funnel. Anyone who wants an audit, a gap list, and a proposal to close the gaps. Anyone who wants a standup workshop or a target operating model they can file. Anyone who will not grant decision rights. Anyone looking for a full-time employee, a recruiter, or a contractor to convert. Anyone whose real problem is that they want a demo by Thursday.

It is also not for a company that already has a strong CoE and wants extra hands. That is staff augmentation. We are not extra hands. We are the unit. If the unit is filled, we should not be in it. And it is not the Chief AI Officer seat. If you need the officer, that page already exists.