Centralise the Reusable. Distribute the Problem Solving.

There is an interesting paradox in building specialist AI teams.

You may need one precisely so that, eventually, you don’t need one.

Right now, concentrating scarce capability can make a lot of sense. A small specialist team can move quickly, learn faster, establish patterns, build shared components and demonstrate what is possible.

The risk comes if AI remains something that specialist team does on behalf of everybody else.

That can accidentally teach the organisation that AI is a specialist activity rather than an increasingly normal part of how digital, data, technology and operational teams solve problems.

I think the distinction is between centralising capability and centralising problem solving.

There are things it makes sense to build once and reuse widely: secure access to models, identity and permissions, observability, cost controls, evaluation patterns, common integrations, safety mechanisms and perhaps some shared context.

Those are parts of the organisational harness.

But understanding the problem should usually remain close to the people who understand the service, the users and the work.

Centralise the capability that should be reusable. Distribute the ability to solve problems.

That also avoids another trap: building a large central platform before the organisation has learned what it actually needs.

Reusable capability should emerge from delivery.

Solve real problems. Notice what keeps recurring. Turn the recurring parts into shared capability. Then let more teams use them to solve the next set of problems.

Over time, the specialist centre changes shape.

Its value becomes less about being the only place that can build with AI and more about making the rest of the organisation better at doing so responsibly.

That might mean fewer bespoke AI products from the centre and more shared tools, patterns, evaluations, controls and expertise that other teams can draw on.

The end state isn’t necessarily an organisation with a very large AI department.

It may be an organisation where AI has become sufficiently normal that the specialist capability can focus on the genuinely difficult, novel and reusable parts.

You might need a specialist AI unit precisely so that, eventually, you don’t need a specialist AI unit.