What an AI centre of excellence is, and when you need one

What an AI centre of excellence is

An AI centre of excellence (AI CoE) is a permanent team that owns how your organisation builds and runs AI. It holds scarce skills in one place, sets the standards everyone else follows, and delivers or supports the AI work that business units ask for. The word permanent matters. A project team forms to deliver one thing and then disbands. A CoE stays, and what it learns on one piece of work carries into the next.

A CoE can be small. A handful of people with the right mix of skills and a clear mandate is a CoE. A large group with no say over standards or priorities is a department with a new name.

What it does day to day

Most of the work is unglamorous. The CoE takes in requests from the business and sorts them: which are worth doing, which need AI at all, and which a rule or a report would solve. It builds the chosen use cases, or pairs with business-unit teams who build them. It runs what is already live, watching accuracy, cost and failures, and retraining or fixing as data changes.

It also keeps the shared pieces in order: data pipelines, evaluation methods, reusable components, approved tools and models. And it handles the questions no single project wants to own, such as how the organisation meets its privacy and compliance requirements when using AI, who may approve a model for release, and which outputs a person must review.

• Intake and prioritisation of use cases
• Building and releasing AI systems
• Monitoring what is in production
• Shared platforms, components and standards
• Training and advice for business teams

The roles it needs

The mix matters more than the headcount. You need a lead who can talk to executives and to engineers, and who can say no to weak ideas. You need data scientists or machine learning engineers to design and build models. You need data engineers, because most AI work is limited by the state of the data. You need software engineers to turn a working model into a dependable system, and someone responsible for running it afterwards.

Add a product or business analyst who can turn a loose request into a defined problem with a measure of success. Add part-time input from legal, security and the business areas you serve. In a small CoE one person can cover two roles. No role can stay empty for long.

• CoE lead
• Data scientists and machine learning engineers
• Data engineers
• Software and operations engineers
• Product or business analyst

Signs you need one, and signs you do not

A single project team is the right answer when you have one or two use cases, a clear owner and no plan beyond them. Do the project and learn from it. A standing organisation built around a first experiment is overhead.

A CoE starts to earn its place when the same problems appear in several places. Different teams buy overlapping tools. Each project rebuilds the same data connections. Pilots work in a demonstration and never reach production. Nobody can say which models are live or who is responsible for them. People with AI skills are scattered, isolated and hard to keep. If several of these sound familiar, you are already paying for the lack of a centre.

• Several teams doing AI separately
• Pilots that never reach production
• Duplicate tools and data work
• No clear owner for live models
• More demand than skilled people

Where it sits in the organisation

There are three usual arrangements. A central CoE does all AI work itself, which keeps standards high and can turn into a queue. A hub-and-spoke CoE keeps a central core for standards and hard problems, with AI people placed inside business units. A federated model leaves delivery with the units and keeps only governance and shared platforms in the centre. Many organisations start central and move towards hub-and-spoke as skills spread.

Whichever you choose, tie the CoE's priorities to business owners. Every use case needs a named sponsor outside the CoE who will use the result and answer for the outcome. Without that sponsor, the CoE builds things that are technically sound and that nobody adopts.

How an AI CoE relates to a GCC

A global capability centre (GCC) is a team in another country that your company owns, set up to run chosen functions. An AI CoE is defined by what it does. A GCC is defined by where it is and who owns it. The two often overlap. Many companies house their AI CoE inside a GCC, because that is where they can hire the range of data and engineering skills in one place. A GCC can also hold much more than AI, such as finance operations, engineering and support.

Both raise the same build question. You can assemble the team yourself, or have a partner build and run it and then hand it over (build-operate-transfer). Data Prophets builds AI CoEs and GCCs on that model and has handed over two centres, each with 80+ people.

How to measure it

Judge the CoE on what reaches production and what changes as a result. Useful measures include the number of use cases live and in regular use, the time from an approved idea to a working release, the pilots stopped early for good reasons, and the business measure each use case was meant to move. Count reuse as well: components, pipelines and patterns that a second project picked up without rebuilding.

Avoid judging it on the number of models built or experiments run. Those figures rise easily and prove little.

• Use cases live and in regular use
• Time from approved idea to release
• Business measure moved per use case
• Components reused by later projects

What to do next

List the AI work under way or requested across your organisation, with an owner and a status for each. If the list is short, run a project. If it shows duplication and stalled pilots, sketch a small CoE: a lead, the core roles and a mandate. Our free AI readiness report can give you a starting view.