Building an AI Centre of Excellence: A Practical Guide for Mid-Market UK Firms

Agentic Business Design

6 April 2026 | By Ashley Marshall

Every consulting firm on the planet is telling you to build an AI Centre of Excellence. Most of the advice assumes you have a 500-person IT department and a seven-figure transformation budget. You probably do not.

What an AI Centre of Excellence Actually Does

Strip away the consulting jargon and a CoE does four things:

1. Sets standards. Which AI tools are approved? What data can feed into them? How do you test before deploying?
2. Shares knowledge. Stops every department reinventing the wheel. Marketing's chatbot learnings help customer service. Finance's forecasting approach informs operations.
3. Governs risk. Ensures AI use is compliant, ethical, and aligned with business goals. Prevents the "shadow AI" problem where teams adopt tools without oversight.
4. Measures impact. Tracks what is working, what is not, and where to invest next.

That is it. It is not a technology lab. It is not a standalone department. It is a coordination function that makes AI work across your business.

The Hub-and-Spoke Model

The most effective structure for mid-market firms is hub-and-spoke. A small central team (the hub) supports AI champions embedded in each department (the spokes).

The Hub (3-5 People)

• AI Lead: Senior enough to influence strategy, technical enough to evaluate tools. Often an existing head of technology or operations who takes on AI as a primary responsibility. Budget for 40-60% of their time.
• Data/Technical Specialist: Handles integrations, data quality, model evaluation, and vendor assessment. Can be a hire, a contractor, or a managed service from an AI consultancy.
• Change Champion: Manages adoption, training, and communication. AI that nobody uses is worthless. This role is often undervalued and always critical.

The Spokes (1 Per Department)

Each major department nominates an AI champion. These people keep their existing roles but spend 10-15% of their time on AI initiatives. They are the bridge between the central team and the front line.

Their job: identify opportunities, test tools, report results, and flag problems. They do not need to be technical. They need to understand their department's workflows and be curious about improvement.

The 90-Day Launch Sequence

Do not plan for six months. You will lose momentum. Here is a practical 90-day roadmap:

Days 1-30: Foundation

• Week 1: Appoint the AI Lead. Get explicit board sponsorship. This needs to be visible leadership commitment, not a quiet IT project.
• Week 2: Draft the CoE charter. One page. Purpose, scope, decision rights, reporting line, initial budget. Templates are useful but keep it short.
• Week 3: Audit current AI use. Survey every department. You will be surprised what people are already using, often unsanctioned tools processing sensitive data.
• Week 4: Select department champions. Brief them on their role. Set up a shared communication channel.

Days 31-60: First Wins

• Weeks 5-6: Prioritise three to five AI opportunities from the audit. Score them on impact versus effort. Pick one or two that can show results within weeks, not months.
• Weeks 7-8: Run structured pilots. Define success metrics before you start. Document everything, including what goes wrong.

Days 61-90: Scale and Standardise

• Weeks 9-10: Review pilot results. Present to the board. Be honest about what worked and what did not.
• Weeks 11-12: Publish your first AI governance framework. Create an approved tools list. Set up a simple request process for new AI initiatives.

Governance Without Bureaucracy

The fastest way to kill an AI CoE is to make it a bottleneck. Here is a tiered governance approach that balances speed with safety:

Tier 1: Self-Service (No Approval Needed)

Pre-approved tools used for their intended purpose. Examples: using approved AI writing assistants for marketing drafts, AI-powered analytics on non-sensitive data, or approved transcription tools for meeting notes. Publish a clear list and update it quarterly.

Tier 2: Light Review (Champion Approval)

New tools or new use cases for approved tools. The department champion reviews against a simple checklist: data sensitivity, cost, integration requirements, and compliance implications. Turnaround target: 48 hours.

Tier 3: Full Review (CoE Approval)

Anything involving personal data, customer-facing AI, significant spend, or integration with core systems. The central team evaluates with a structured assessment. Turnaround target: two weeks.

The key principle: make the right thing easy and the wrong thing hard. If your governance process is slower than just signing up for a free trial, people will bypass it.

What It Actually Costs

For a mid-market UK firm, realistic first-year costs:

• People: Primarily time reallocation, not new hires. Budget 0.5 FTE for the AI Lead role and 0.1 FTE per department champion. If you hire a dedicated technical specialist, expect GBP 55,000 to GBP 85,000 depending on experience and location.
• Tools and platforms: GBP 500 to GBP 2,000 per month for enterprise AI tools, depending on scale. Many start with existing tools used more effectively.
• Training: GBP 5,000 to GBP 15,000 for structured AI literacy programmes. This is not optional. Untrained teams will either over-trust or under-use AI.
• External support: GBP 10,000 to GBP 40,000 for consultancy to help with setup, governance design, and initial use-case identification. This is where an experienced partner accelerates the process significantly.

Total first-year investment for a 200-person company: roughly GBP 80,000 to GBP 160,000 including allocated people time. That sounds significant until you compare it with the cost of uncoordinated AI adoption, duplicated tools, compliance failures, and missed opportunities.

Common Mistakes to Avoid

Making It Too Technical

A CoE that only speaks machine learning will alienate the business. The most effective teams are business-led with technical support, not the other way around.

Waiting for Perfect Data

Your data will never be perfect. Start with what you have, improve as you go. Data quality improvements driven by real AI use cases are far more effective than abstract data cleansing projects.

Ignoring Change Management

Gartner reports that 91% of high-maturity AI organisations have dedicated change management. The technology is the easy part. Getting people to actually use it, trust it, and adapt their workflows is the real challenge.

Over-Centralising

If every AI request goes through a committee, you will kill innovation. The tiered governance model prevents this, but only if you actually trust the lower tiers to work.

Measuring Success

Track these metrics quarterly:

• Adoption rate: What percentage of employees actively use approved AI tools?
• Time saved: Hours reclaimed through AI-assisted processes, measured per department
• Initiative pipeline: Number of AI use cases identified, piloted, and scaled
• Compliance posture: Number of shadow AI tools discovered and brought under governance
• Business impact: Revenue influenced, costs reduced, or quality improved by AI initiatives

Avoid vanity metrics. "We deployed 12 AI tools" means nothing. "Customer response time dropped 40% through AI-assisted triage" means everything.

Frequently Asked Questions

Do we need to hire AI specialists to run a Centre of Excellence?

Not initially. Most mid-market firms start by reallocating existing talent and supplementing with external expertise. The AI Lead role is often an expansion of an existing technology or operations leadership position. Technical specialists can be contractors or part of a managed service agreement with an AI consultancy.

How long before we see ROI from an AI Centre of Excellence?

Quick wins from the first pilot cycle (60-90 days) typically cover tooling costs. Broader organisational ROI, including reduced duplication, better compliance, and scaled automation, usually becomes clear within 6-12 months. The key is choosing high-impact first use cases that deliver visible results early.

What is the biggest risk of NOT having an AI Centre of Excellence?

Shadow AI. Without governance, teams adopt tools independently, feeding sensitive data into unvetted systems, duplicating costs, and creating compliance exposure. A 2026 Gartner survey found that 68% of enterprises discovered unsanctioned AI tools processing customer data. The CoE does not exist to slow things down. It exists to make AI safe and effective.