A portfolio of pilots is not a strategy. The transition that matters is from interesting experiments to a system the organisation actually runs.
Most organisations we meet do not have an AI problem. They have a pilot problem. There are projects in every function, a respectable amount of spend, and a slide that lists them all. What there is not is an operating model — a way of working in which AI is something the organisation runs, owns, and improves, rather than a collection of experiments that each belong to whoever started them.
A portfolio of pilots is not a strategy. It is a list of options, and options expire.
The transition from pilots to an operating model is the single hardest step in any AI programme, and it is the one that separates the organisations that compound from the ones that stall. It is hard because it is not technical. It is organisational: ownership, accountability, funding, and the unglamorous infrastructure that makes the tenth deployment faster than the first.
It begins with ownership. A pilot can live in an innovation team; an operating model cannot. The systems that matter have to be owned by the part of the business that depends on them, with a named person accountable for the outcome and the budget to keep it running. Until that ownership exists, every success is fragile and every problem is someone else's.
Then comes the platform. The first three AI systems an organisation builds will each reinvent the same plumbing — data access, evaluation, deployment, observability, guardrails. The shift to an operating model is the decision to build that plumbing once, properly, so the fourth system inherits it. This is what turns AI from a series of bespoke projects into a capability.
There are projects in every function, a respectable amount of spend, and a slide that lists them all.
Funding has to change too. Pilots are funded as experiments: a fixed amount, a fixed window, a binary outcome. Operating models are funded as products: a baseline to run and maintain what exists, plus investment to extend it. Organisations that keep funding AI as a series of experiments keep getting experiments.
Governance moves from project to standard. Instead of each team negotiating controls from scratch, the organisation defines once how AI systems handle data, permissions, evaluation, and audit — and every new build inherits that posture. This is how governance stops being a tax on each project and becomes a shared asset that speeds all of them up.
And the work has to be sequenced rather than scattered. A good operating model does not chase every idea at once; it picks the builds that are valuable, feasible, and governable, ships them, and uses what it learns to make the next ones cheaper. Momentum compounds when the organisation finishes things, and dissipates when it starts everything.
The payoff is a different kind of organisation. The second deployment is faster than the first because the platform exists. The risk conversation is shorter because the governance is standard. The business owns the outcomes because the accountability is clear. AI stops being a thing the organisation is trying and becomes a thing the organisation does.
Pilots prove that something is possible. An operating model is how a serious organisation makes it ordinary — and ordinary, repeatable, governed AI is where the value actually lives.
