In plain terms
Most companies have a handful of people who really understand AI and dozens of departments that want it. A centre of excellence puts those few people in one place and gives them one job: make it easy and safe for everyone else. They choose the tools, write the rules of use, build the parts that every project needs and coach the teams. The departments still own their projects and their results.
Why it matters
It stops ten departments from buying ten tools, negotiating ten contracts and repeating the same mistakes. The design question is the split: the centre owns what is common, and business units own what is specific to their work. Drawn wrongly, that line produces one of two failures. As a bottleneck, the centre approves and builds everything, and its queue grows to months. As an ivory tower, it produces frameworks and demos that no business unit asked for. The check for an executive: judge the centre by what the business units have put into production.
Example
A logistics group sets up an AI centre of eight people. In its first year the centre tries to build every request itself: 45 requests arrive, 5 are delivered and the average wait is seven months. In the second year the group moves to hub-and-spoke: the centre runs the platform, the standards and the training, and each of six divisions names two AI leads who build with its support. That year 18 use cases reach production.
Most often confused with
CoE vs. Central AI or data-science team
A central delivery team takes requests and builds solutions; its capacity is the limit of what the company can do. A centre of excellence multiplies capacity elsewhere: its products are standards, shared components, training and advice. Many centres start as delivery teams, which is reasonable while skills are scarce. The warning sign is a centre that, two years on, is still the only place where AI gets built.
Under the hood
Three organisational models. Centralised: all AI specialists sit in one unit that builds for the whole company; quick to start, consistent, and a bottleneck as demand grows. Hub-and-spoke: a central hub owns the platform, standards and governance, while spokes in the business units own use cases and delivery; the usual target for larger organisations. Federated: business units run their own AI teams and a small central group coordinates standards; it suits groups with very different businesses and carries a risk of duplication. What the centre should own: model and vendor contracts, the shared platform (model access through a gateway, evaluation tooling, logging, guardrails), reference architectures and reusable components, policy and risk review together with legal and security, training, and the view of the whole portfolio. What it should leave to the business: choosing and funding use cases, process redesign, adoption and the benefit case. Measures: use cases in production in the business units, reuse of shared components, time from request to production, and adoption.