Imagine a food group where three artificial intelligence projects have started without consulting one another: one in production planning, another in delivery rounds, the last in quality control. Each department has chosen its supplier, opened its own access to a language model and set its own testing rules. The CIO then proposes to the executive committee that an AI centre of excellence be created.
An AI centre of excellence is a cross-functional team that pools what several AI projects have no interest in redoing separately: scarce skills, a technical foundation, an evaluation method. The term is a matter of usage, and its content is set in a mandate. Such a team is justified when several departments are running projects in parallel. It pools the foundation and leaves the business units their use cases. From its creation, its mandate carries a response time and a review date. The questions that follow are those this committee would have to examine.
Is an AI centre of excellence needed, or are project teams enough?
The answer can be read in what the current projects have already built or bought. We propose a test: draw up for each project the list of its foundations, that is, access to a model, security rules, the evaluation set, the supplier contract and the skills mobilised. An element present in several lists is a candidate for pooling. In the starting scenario, access, contracts and testing methods are among them.
If no element recurs, the question is premature. With one or two projects in the same department, a project team and a champion can be enough. A structure created before the projects it is meant to serve does not yet have any demand to respond to.
Several forms are found in practice, and they can be combined: a central team bringing together the specialists, a network of champions who remain in their department and devote part of their time to it, teams based in the business units and bound by common rules. The centre of excellence corresponds to the first, supplemented or not by the second. Our AI strategy page places this organisational choice among the trade-offs of the strategy, alongside use cases and sourcing.
Which department should it report to?
The reporting line says to whom the centre is accountable, and who decides when two departments ask it for the same capacity.
Reporting to the IT department, the centre is close to the foundation, security and operations. It may be perceived as a technical team, called upon once the need has been settled. Reporting to a data department, where one exists, it is close to the sources and the models, with the same reservation. Reporting to general management or to a transformation department, it has an arbiter between the business units and depends on other teams for everything to do with the foundation. Reporting to the business department that runs the most projects, it is close to usage, and the other departments may see it as serving just one of them.
The criterion we propose is arbitration: the centre is placed where a member of the executive committee can settle two competing requests, or it is given a sponsor with that power.
The funding model also matters. A centre funded from a central budget must publish its priority rule, otherwise order of arrival takes its place. A centre recharged to departments may see its services chosen by those that can pay, and low-value requests risk no longer reaching it.
What should be pooled, and what left to the business departments?
The elements to pool are those that are costly to redo and benefit from being identical from one project to the next:
- access to models and the technical foundation (accounts, security, logging);
- the evaluation method and the tools that support it;
- the relationship with suppliers and the contracts;
- reusable components and the memory of past projects;
- upskilling of internal teams, through pair working, documentation and workshops.
On this point, Article 4 of Regulation (EU) 2024/1689, known as the AI Act, in its wording resulting from Regulation (EU) 2026/1744, asks providers and deployers of AI systems to take measures to foster the development of AI literacy among their staff. The centre can be the function responsible for these measures.
The business departments retain the elements that require knowing the process and being accountable for it: the choice of use case and its funding, the commitment to the gain, the description of the work to be transformed, the judgement on what a good answer is, adoption by the teams.
One boundary calls for a precaution, the one separating support from control. The register of systems, their classification by risk level and authorisation to put them into service fall under AI governance. The centre can prepare the files. The authorising body is best kept separate from it, since a centre that has built a system would have to rule on its own work. Our article on types of PMO describes, for a project office, the scale running from support to control and then to direct management. It can be transposed: the mandate states at which level the centre sits for each service it provides.
How do you prevent it from becoming a bottleneck?
The mechanism is arithmetic: if every project must go through a team of fixed size, the queue grows with the number of projects. Departments wait, or do without the centre, and the uses that develop outside its remit may join what is known as shadow AI.
Several arrangements address this. The foundation is made usable without the centre’s intervention: documented access, project templates, annotated examples. The departments’ champions handle routine requests and pass on only what is new. Each of the centre’s services comes with an announced response time. The priority rule is written down, and applying it falls to the sponsor rather than to the centre itself. Low-risk uses follow a fast authorisation track, without having to go through the central team.
A few measures show whether these arrangements hold: the time between a request and its first response, the number of pending requests, the share of time the centre spends doing work in place of the teams, the number of projects launched outside the common foundation.
When should it be changed or dissolved?
A centre of excellence responds to a moment: skills are scarce, the foundation does not exist, the rules are not written. When these gaps are filled, its purpose changes. The mandate therefore provides for a review on a fixed date, at which the committee goes back over the services still requested, the departments that have become autonomous and the elements of the foundation stable enough to be run as a routine service.
The foundation can move to the operations teams, with the pipeline described in our MLOps and LLMOps page. The central team can shrink to a core that runs the network of champions and maintains the evaluation method. The centre is dissolved when each of its functions has found a permanent owner, named in the dissolution decision: the foundation, the contracts, the common reference set.
Certain signs may lead to bringing the review forward: the centre measures its activity by the number of projects it delivers itself, or departments consult it out of obligation without expecting an answer.
On the date set in the mandate, the committee of the group imagined at the outset would go back over the list of what its three departments do not yet do on their own, and the centre’s fate would be read there.