Should You Set Up an AI Center of Excellence?

An AI center of excellence makes sense when several functions already use AI without shared rules, and it is surplus when almost nobody does. The usual mistake is not building one, it is building it too early: a central layer with nothing to coordinate ends up deciding about processes it does not know.

Carlos Andrés Ramírez · 2026-10-11

An AI center of excellence makes sense when several functions already use AI without shared rules, and it is surplus when almost nobody does. The usual mistake is not building one, it is building it too early: a central layer with nothing to coordinate ends up deciding about processes it does not know.

Start with how the proposal reaches the table. An executive committee sees AI moving on several fronts, with nobody leading it and some disorder, and concludes there should be one place where the knowledge sits. A lead is named, five or six technical people are gathered, and they are asked to help every function. It sounds reasonable, and sometimes it is.

The short answer is that a center of excellence solves a coordination problem, and it only pays off when that problem exists. If three teams are buying similar tools, each with its own data rules and its own idea of an acceptable result, the center saves money and risk. If there is one pilot and two curious people, the center has nothing to coordinate and starts producing the only thing it can: documents, committees and templates.

The symptom

When does an AI center of excellence turn into bureaucracy?

Picture a central team that has just formed and has no cases of its own. To justify itself it defines an approval process: every AI initiative must pass through it before starting. A finance team that wanted to try a reconciliation tool finds it now needs an assessment, a form and a meeting. It takes weeks for what used to take days, and the next time it goes ahead without telling anyone.

That is the pattern that turns a center into bureaucracy. It does not happen out of ill will. It happens because a team that does not execute can only control, and control without knowledge of the process becomes generic requirements that slow down whoever is actually doing something.

The test

The moment shows in three signs, not a hunch

Before deciding, check three things. If all three hold, the center has real work from day one. If one or none do, a lighter alternative serves better.

1. Several functions run AI in production

With two or three teams that already have something working, the center can learn from real cases and spread what it learns. With scattered pilots or ideas on paper, what it spreads is opinion.

2. Those functions contradict each other on something that costs money

They sign with the same vendor on different terms, treat data under different rules, or measure results with indicators that cannot be compared. Each contradiction is a decision someone should take once, and the center is where it gets taken.

3. Someone has authority to decide, not only to advise

A center that only advises depends on functions choosing to listen. One that can set a data standard or stop a duplicate purchase has real influence. Without that authority, the layer exists on the org chart and not in decisions.

A center of excellence does not create the coordination that is missing. It organizes it once there is something to coordinate.

BECOME

The alternative

Before the center, one person and a few rules can be enough

When the three signs are not yet there, the alternative is not doing nothing. It is a minimal version: one person with allocated time, a short register of what is in use and where, and two or three rules for everyone, such as which data never leaves the company and who approves a recurring spend. That covers the main risk without building a structure that must then be maintained.

The minimal setup also does something else: it shows how often functions collide. If after a few months the register shows duplicate purchases and clashing criteria, there is evidence the center is needed and what it has to solve. If the register is still nearly empty, you avoided building a layer for nothing.

The consequence

A well-timed center pays for itself in the decisions it saves

When it is set up at the right time, the center stops being measured by the meetings it runs and starts being measured by the decisions it prevents from being repeated: one negotiation with the vendor instead of three, one data standard instead of four interpretations. Its design changes too: fewer people approving, more people helping functions execute, with cases of its own that it keeps current.

List the functions that use AI in production today and note where they contradict each other. If the list has fewer than three functions or no costly contradiction, start with one person and a register. If it has three or more and a conflict that costs money, the center already has its first job.

Let's check whether your organization is ready for a center of excellence