What's the right AI operating model?

Three common shapes

1. Centre of excellence (CoE)

A central AI team owns strategy, standards, tooling, and often delivery itself. This works when the organisation is starting from near-zero AI capability: nobody in the business units knows what a model card is, procurement doesn't have a clause for AI vendors, and if you don't centralise, you get chaos instead of consistency. The trade-off shows up fast. Business units start treating AI as "someone else's job." Requests pile up in a small team's backlog. Six months in, the CoE is the bottleneck it was meant to prevent.

2. Federated

Capability sits inside each business unit or function. This suits organisations where the units are genuinely different businesses — a bank and an insurer under one group, a manufacturing arm and a logistics arm — where domain context matters more than technical uniformity. The risk is duplication: three teams solving the same document-classification problem with three different vendors, three different data retention policies, and no one person who could tell you how many AI systems the organisation actually runs.

3. Hub-and-spoke

A small central hub owns platforms, sets standards, and runs governance. Spokes inside business units do the delivery, close to the domain problem. Most organisations that survive past their first eighteen months of AI activity end up here, whether they planned it or not.

The transition nobody plans for

Here's the pattern. A mid-sized retailer sets up a CoE: four people, a shared model, a policy document. For a year it works — every AI initiative in the business runs through them, and quality is high because there's one team building judgement.

Then it doesn't scale. Marketing wants a content pipeline. Supply chain wants demand forecasting. Customer service wants a triage assistant. The CoE can build one of these well per quarter. The business wants all three now. Units start going around the CoE — a manager buys a chatbot tool on a card, a regional team hires a contractor. Nobody decided to become federated; it happened because the centre couldn't keep up.

This is the moment that matters, and it's the one most organisations don't plan for because nobody schedules a review of the operating model — they only notice it's broken once shadow AI is already running in production somewhere. The fix isn't heroics from the CoE. It's a structural change: the CoE stops being the delivery team and becomes the hub. It keeps the platform, the standards, and the governance board, and it hands delivery to spokes embedded in supply chain, marketing, and customer service. Those spokes report technically into the business unit but are accountable to the hub for anything touching data, models, or risk.

The mechanics of that handover are where it goes wrong. Spokes need people who understand both the domain and enough of the platform to not reinvent it — that's a hiring and training problem, not just an org chart change. And the hub needs the authority to say no to a spoke's architecture choice, which means the governance board has to have teeth before the transition, not after.

Governance is the constant

Whichever shape you're in, governance doesn't move. It's not owned by a business unit, and it's not something the hub delegates away when spokes appear. A cross-functional board, meeting on a fixed cadence, reviewing new initiatives, incidents, and policy changes — that structure needs to exist on day one and survive every reshuffle after. If governance is redesigned every time the operating model changes, you've built policy on sand.

Specific pitfalls

• No funding decision. Central pot, chargeback, or hybrid — pick one explicitly. Leaving it ambiguous means every project starts with a fight about who pays.
• Talent strategy as an afterthought. Decide up front who's hired centrally, who's embedded in a spoke, who's borrowed for a quarter. Without this, your best people end up wherever shouted loudest.
• Treating the operating model as permanent. It isn't. Revisit it every 18 to 24 months, or sooner if you notice initiatives routing around the structure rather than through it.
• Standing up spokes without hub authority. A spoke that can ignore the platform standard isn't a spoke, it's federation with extra paperwork.

Where Florence helps

Florence doesn't dictate an operating model, but the Strategy step is where you record the decision — CoE, federated, or hub-and-spoke — alongside the readiness assessment that tells you which one fits where you actually are, not where you'd like to be. Insights surfaces the gaps between your current structure and your ambitions (a federated set-up trying to run enterprise-wide governance, for instance). And because the decision sits with the rest of the strategy, not in a slide deck nobody opens again, it's there to revisit at the 18-month mark — before the transition catches you by surprise rather than after.