How to Build an AI Centre of Excellence: A Practical Guide for UK Organisations

Not every organisation needs a formal AI CoE, and building one too early is as problematic as building one too late.

Build a CoE when: you have more than five active AI pilots or deployments across different functions; you are spending more on AI in aggregate than a single function can govern effectively; you are experiencing duplication of effort, inconsistent governance standards, or conflicts between AI initiatives; or your board is asking for consolidated AI oversight that no existing function can provide.

Do not build a CoE when: you are at the earliest stage of AI exploration, where the overhead of the CoE will exceed its value; you do not have the senior sponsorship to give the CoE meaningful authority; or the purpose of the CoE is primarily to create the appearance of AI governance without the substance.

For most mid-sized UK organisations (FTSE 350 range), the right time to establish a formal CoE is when the AI portfolio has grown to a point where ad hoc governance is clearly insufficient, typically 12 to 18 months into a serious AI programme.

The most common structural error in AI CoEs is overstaffing at the technical level and understaffing at the change management and governance level.

An effective AI CoE requires four capability streams:

Technical AI capability: data scientists, ML engineers, and AI specialists who can evaluate AI tools, build custom applications where needed, and provide technical guidance to business functions.

Change and adoption capability: change managers, communication specialists, and training designers who can build the workforce adoption infrastructure that determines whether AI investments deliver value.

Governance and risk capability: AI governance specialists, risk managers, and legal/compliance support who can develop and maintain AI policy, manage the regulatory compliance dimensions of AI deployment, and provide the board oversight function.

Business liaison capability: people with strong business relationships in the major functions, who translate between technical AI capability and business need.

Many UK CoEs have the first stream and lack the second and third. The result is technical capability without adoption or governance infrastructure.

A CoE without authority is a cost centre that frustrates rather than enables.

The CoE mandate should be agreed with the board and embedded in governance documents, not just described in a launch communication. The minimum authority required for a functional CoE:

Approval authority over new AI deployments above a threshold (for example, any AI deployment involving personal data, or any AI investment above a defined threshold).

Policy-setting authority for AI use standards, data handling requirements, and risk parameters across the organisation.

Budget for central change management and adoption support, separate from individual business function budgets.

Reporting line to the C-suite, not to IT. Positioning the CoE as an IT sub-function limits its authority in business change discussions. The most effective CoEs report to the CEO or a C-suite leader with broad business authority.

Without this authority, the CoE becomes an advisory body that business functions can ignore when convenient, which is exactly what will happen.

The CoE's operating model should distinguish between its three main functions:

Governance: maintaining AI policy, reviewing new AI deployments for compliance with standards, monitoring ongoing deployments for risk. This function needs regular cadence (weekly or bi-weekly governance reviews) and clear documentation.

Enablement: providing the change management, training, and adoption support that individual business functions need to deploy AI effectively. This function is demand-driven and project-based.

Innovation: scanning for new AI capabilities, running horizon-scanning exercises, evaluating new tools against the organisation's use cases. This function benefits from a quarterly strategic review cadence.

For a mid-sized CoE, a weekly operations meeting for governance decisions, a monthly all-CoE review, and a quarterly strategic review provides the right rhythm. Add a quarterly report to the board covering AI portfolio status, risk posture, and adoption progress.