Creating an AI Centre of Excellence Without a Big Company Budget

What the big-company version gets right

Large organisations set up AI centres of excellence with dedicated staff, a budget line and a slide deck of pillars. Strip away the scale and they do four useful things: stop every team from solving the same problem differently, keep one set of rules about data and tools, spread what works, and put scarce expertise where it counts.

A 120-person business has exactly the same problems, just smaller. The marketing team found a good way to draft product descriptions and nobody in sales knows. Finance is paying for one AI tool and operations for a different one that does the same thing. Three people have independently asked IT whether they can connect an assistant to the shared drive.

You do not need a department to fix that. You need a small group with a clear job.

Who should be in it

Three to five people, drawn from where AI is actually used rather than by seniority.

• A lead with enough authority to make decisions stick, often an operations director or a senior manager who already uses AI seriously
• Someone from IT or systems who understands access, integrations, security and what the current software can do
• Two or so heavy users from different parts of the business, chosen for curiosity and good judgement rather than job title
• Occasionally, a person who handles data protection or compliance, present for decisions touching personal data

Avoid making it a group of enthusiasts only. You want at least one member who is naturally sceptical. They will save you money.

What it actually does

The approval role is the one people worry will become bureaucracy. Keep it fast. A form with five questions, a default answer of yes for low-risk uses on approved tools, and a promise of a response within a week. The moment approval takes a month, people stop asking and start using personal accounts, which is exactly the shadow AI problem the group exists to prevent.

How much time it takes

For an illustrative 150-person business, a realistic commitment looks like this:

• A 45-minute meeting every fortnight for the whole group
• Two to four hours a week from the lead
• An hour or two a week from IT for approvals and access
• Occasional time from members when a larger project is being scoped

That is roughly a fifth of one person's time spread across several people. It is small enough to sustain and big enough to matter. If it grows well beyond that, you probably have enough AI activity to justify a dedicated role, which is a good problem.

The first ninety days

1. Weeks 1-2: build the inventory of AI tools and uses across the business. Check invoices and expenses as well as asking.
2. Weeks 3-4: agree approved tools per data sensitivity level and publish a one-page summary. If you have no usage policy yet, write one; our guide to writing a company AI usage policy gives a starting structure.
3. Weeks 5-6: launch the intake form and commit to response times.
4. Weeks 7-10: collect the ten best existing uses from staff and write each up in a paragraph with an example. Share them.
5. Weeks 11-13: review the list of bigger ideas and pick one or two to take forward properly, with a sponsor and a measurement plan.

By the end, the business should have fewer duplicate tools, a clear route for new ideas and a visible set of things that work. That is a centre of excellence, whatever you call it.

Where outside expertise fits

The honest limitation of a small internal group is depth. Members are good at spotting opportunities and setting sensible rules, but most will not have built a production AI system, evaluated a model rigorously or integrated an agent with an ERP.

That is where an outside partner is cheaper than hiring. At SpiderHunts we often act as the technical bench for groups like this: joining a meeting a month, giving a view on whether an idea is feasible and roughly what it costs, and building the pieces that need engineering through our enterprise AI service. The internal group keeps ownership of priorities and standards. That split works well, as long as the partner is prepared to say no to ideas, including ones it could bill for.

The group's job is judgement and consistency. Engineering depth can be borrowed; knowing the business cannot.

When not to bother

If you have fewer than about 30 people, one sensible person owning AI tools and policy is enough. A committee would outnumber the decisions.

It also fails when set up without authority. If the group can recommend but not approve, and every decision goes back to the managing director anyway, members will stop turning up by month three. Give it a real remit, a small budget for tools, and the right to say no, or do not set it up.

Finally, beware the centre of excellence that becomes the only place AI is allowed to happen. Its job is to make good AI use easier across the business, including internal copilots for teams, not to own all of it.

Frequently asked questions