The most common first move in enterprise AI transformation is also the most dangerous one: hiring a Head of AI and giving them ownership of all AI initiatives.

It feels like the right call. You're signalling commitment. You're creating accountability. You're getting a senior person in the room when AI decisions get made.

What you're actually doing, in most cases, is creating a bottleneck.

"We created CDOs. Then everyone waited for the CDO to do digital. We're doing the same thing with AI."

The Centralisation Trap

When you create a Head of AI and put all AI initiatives under one function, a predictable set of dynamics kicks in.

Every AI idea in the organisation now has to go through the AI team. The AI team, being a small group with limited bandwidth, becomes a queue. Teams across the business stop experimenting on their own — they wait for the AI team to come to them. The AI team, meanwhile, gets protective of its portfolio.

The business units don't build AI fluency. They build a dependency.

The typical model
Centralised AI Function
  • All AI runs through one team
  • Business units wait to be served
  • Head of AI owns all AI decisions
  • AI fluency stays in one function
  • Bottleneck grows as demand grows
  • Innovation requires central approval
The winning model
Distributed AI Capability
  • AI embedded in every function
  • Business units build their own AI
  • Shared infrastructure, distributed ownership
  • AI fluency is a baseline expectation
  • Capacity scales automatically
  • Innovation happens at the edge

The CDO Parallel

This is not a new failure mode. We ran this exact experiment during the digital transformation era.

Companies that hired CDOs and created central digital functions largely failed to transform. The CDO became a political player managing a separate team that did "digital things" while the rest of the organisation continued operating exactly as before. Digital fluency never diffused into the business.

The companies that got digital transformation right took a different approach: they embedded digital capability into every function, set expectations that business leaders would own digital outcomes, and used a small central team for shared infrastructure and standards — not for doing the work.

The AI transformation parallel is exact.

What the Head of AI Role Should Actually Look Like

None of this means you shouldn't have a Head of AI. It means the role needs to be defined differently.

The right job description
Head of AI — What This Role Should Own
1
Shared infrastructure. The model serving layer, data pipelines, the observability stack. The plumbing every team uses but no single team should own.
2
Standards and guardrails. How AI gets evaluated before it goes to production. Security, compliance, the rules of the road — not the driving.
3
Capability building. Making every PM, analyst, and ops lead in the company more capable of building AI into their own work. Train the organisation, don't replace it.
4
Killing the role itself. The explicit goal: make AI so embedded in every team that a centralised AI function is no longer necessary.

The Metric That Matters

Here's the test for whether your AI transformation structure is working: can the Head of AI go on a two-week holiday without any AI initiative in the company slowing down?

In a centralised model, the answer is no. In a distributed model with strong embedded capability, the answer is yes.

"The Head of AI should be killing their own job description — not by doing more AI, but by making AI so embedded that nobody needs a centralised function anymore."

That's a harder job to hire for. It requires someone who is genuinely excited about giving capability away, about making other people's work better rather than owning the most interesting work themselves.

But it's the only Head of AI worth hiring.

The Transformation Layer

A weekly 300-word essay on AI transformation, infrastructure, and what executives get wrong.