AI Centre of Excellence: The Missing Piece

AI is already embedded across most organisations. Far fewer can say who owns it?

Ownership here means more than who bought the tool or sponsored the first pilot. It means who owns the standards, the risk decisions, the documentation, the learning, and the decisions about what should scale.

That is where many AI programmes start to wobble. Teams experiment locally and governance arrives late. Nobody can give a confident answer when leaders ask, “Are we ready for the EU AI Act?”

This is why the AI Centre of Excellence (CoE) conversation has become urgent.

The question is no longer whether AI needs governance. It does. The real question is whether an organisation has a practical way to make governance work without slowing delivery down.

What an AI Centre of Excellence actually means

Done well, a Centre of Excellence is a working capability, not another committee or a folder of unread policies.

It brings together shared expertise, standards, governance, enablement, and measurement that help the organisation move faster with more confidence.

Without that, teams solve the same problems repeatedly: different risk thresholds, different documentation, different assumptions, and very little shared learning.

Signs AI needs an owner

Most organisations do not announce an AI governance problem. It shows up as avoidable friction.

• Local AI experiments: teams are trying tools independently, with no shared view of what is being tested.
• Duplicate pilots: different teams are solving similar problems with different tools, vendors, and assumptions.
• No shared inventory: leaders cannot easily see which AI systems exist, where they are used, or who owns them.
• Uneven risk decisions: one team treats a use case as low risk while another would apply stronger controls.
• Scaling without repeatability: pilots move into wider use before standards, monitoring, or training are in place.
• Regulatory exposure: AI systems need clearer ownership, documentation, oversight, and risk classification.

The answer is a clearer ownership model, not another pilot for every initiative. That model decides what gets funded, governed, paused, or stopped.

In many organisations, the first five are tolerated as “innovation noise”. The sixth turns them into a board-level concern.

The EU AI Act changes the tone of the conversation

The EU AI Act changes this from a “good governance” discussion into a business readiness issue. It bans unacceptable-risk practices, places heavier obligations on high-risk systems, and introduces lighter transparency duties for lower-risk uses.

Some obligations are already live. AI literacy requirements applied from February 2025, and general-purpose AI model provider obligations from August 2025.

The main wave of high-risk system requirements was set for 2 August 2026. These cover risk management, data governance, documentation, and human oversight. A May 2026 political agreement proposes deferring some use-based high-risk obligations to December 2027. Until it is formally adopted, 2 August 2026 remains the operative date.

Here is the trap: waiting for complete regulatory certainty before building a capability the organisation already needs.

Whether a date moves or not, the work remains. Someone still has to own classification, documentation, oversight, monitoring, and improvement. Without a clear owner, that accountability defaults to the leaders who fund and answer for AI. It usually lands at the worst possible moment.

Why a compliance project is not enough

AI governance is continuous work.

Models change, vendors change, and use cases expand, so a one-off compliance project cannot keep pace. A system that looked low risk at the start can become more significant once it is embedded in real decisions.

A standing CoE owns the work that cannot be left to individual project teams:

• Strategy: keeping AI aligned to business goals.
• Governance and standards: defining risk thresholds and documentation once.
• Capability development: building AI literacy across the workforce.
• Knowledge management: creating one shared source of truth.
• Measurement: tracking value, risk, adoption, and change.

The value of a CoE is simple: less duplication, clearer ownership, lower risk, and faster movement. Teams stop rebuilding governance from scratch every time a new AI idea appears.

A structure that fits most organisations: hub-and-spoke

CoE structure usually follows maturity. A central model helps early consistency. A fully decentralised model gives speed but risks silos. For AI, most organisations need both consistency and reach.

A hub-and-spoke model makes the most sense for many organisations. The hub owns strategy, governance, standards, and shared services. Its spokes keep delivery close to the business, with embedded specialists connected back into the hub.

Deloitte’s 2026 State of AI in the Enterprise survey found that only 21% of organisations have a mature governance model for AI agents. Hub-and-spoke is one practical way to close that gap.

That balance is important because AI adoption is too broad to centralise completely, but too risky to leave entirely local.

Before funding the next AI idea, test readiness

Many generative AI projects fail after proof of concept. Value, adoption, or risk controls were not clear enough before the pilot began.

A readiness framework helps a CoE test AI ideas before budget, time, and governance attention are committed.

• Feasibility: can it be built and operated safely?
• Desirability: does it solve a real problem people will adopt?
• Viability: does it align to strategy, risk, compliance, and ROI?
• Decision: based on the evidence gathered, proceed, pause, or refine.

This is how a CoE earns its place with leaders. It becomes a decision-making capability. Leaders gain the visibility to fund what works and stop what does not.

AI literacy is not a nice-to-have

Value appears when people adopt AI in real workflows, not when tools are deployed.

AI literacy duties applied from February 2025. Organisations that provide or deploy AI must ensure relevant staff understand the systems, their risks, and their limitations.

The literacy gap is real. Microsoft’s 2025 Work Trend Index found that 67% of leaders report familiarity with AI agents, compared with 40% of employees.

AI literacy cannot be a one-off awareness session. People need to know what AI can do, where it fails, when to challenge it, and how to use it safely.

The bigger point: AI governance is a capability

Deadlines matter. But they are not the whole story.

Organisations that scale AI well treat governance as part of the operating model, not a last-minute response to regulation.

A deliberate AI CoE helps them move quickly, responsibly, and consistently as the technology and rules keep changing.

So what should leaders do now?

• Name a single executive sponsor with authority to join up AI governance.
• Use hub-and-spoke when AI risk spans multiple regions, products, or business units.
• Run the CoE as a working team with decision rights, not a policy layer.

A practical test applies. Can leaders explain who owns AI governance, where AI is used, and how people are enabled to use it safely? If not, the CoE conversation is already overdue.

That is the moment to stop asking, “Which AI tool should we buy next?” and start asking, “What capability do we need to build so AI can scale safely?”