AI sovereignty isn’t about place – it’s a switching cost

For a few days in June, a number of British businesses discovered something about their AI arrangements that may not have been on the risk register. The US government imposed export controls on Anthropic; the company withdrew its most capable models worldwide; and organisations that had quietly built working processes on top of them found those processes had stopped working. Access was restored within weeks and the controls were lifted. As disruptions go, it was mild.

In a sense, it was a pretty cheap fire drill.

The reaction has largely been to argue about nationality: whose models, whose data centres, whose cloud. That argument has its place, but it isn’t the one most businesses need to have first. If you are using AI seriously, the honest position is that losing it altogether would hurt. That is rather the point of adopting it. Nobody restructures a claims process, a customer service function or a first-line support desk around a technology they could do without. So the question worth asking is narrower and much more answerable than the geopolitical one: if a model you depend on became unavailable tomorrow, how long would it take you to move to another, and what would that cost you?

The comforting assumption is that models are interchangeable: same interface, similar capabilities, swap the endpoint and carry on. That is roughly true for casual use and roughly false for anything you have invested in.

The friction sits in the layers around the model rather than the model itself. Prompts get tuned to one system’s quirks over months. Evaluation suites, where they exist at all, are calibrated against one baseline. Agent scaffolding, tool definitions and output formats get shaped by what one provider does well. Assurance and procurement sign-off was granted for a named supplier, and the paperwork does not transfer. Somewhere in the middle sits your data (the enterprise context that makes the system useful) and the ease of moving that varies enormously depending on decisions made early and casually. Ethan Mollick has made the point that as the frontier models advance, the small differences between them are becoming amplified.

None of this is exotic. It is ordinary supplier lock-in, of the kind we have understood in IT for thirty years. What is unusual is how fast the dependency formed and how little of it went through the normal disciplines.

The obvious response is to own the whole stack. This deserves a fair hearing and then a realistic costing. McKinsey’s work puts sovereign AI offerings at a perceived 10 to 30 percent premium over global alternatives, and finds demand for them is real but selective: sovereignty earns its cost where data is sensitive, regulatory exposure is material, or the service is genuinely critical. For everything else, the premium buys reassurance rather than resilience.

There is also a trap in the sovereignty pitch itself. Some of the loudest commentary comes from vendors with a direct interest in you replacing one dependency with theirs. Palantir’s chief executive has been a widely quoted example. France has already noticed the pattern, moving to favour domestic providers over Palantir and describing the relationship as a new strategic dependency. Independence achieved by signing a longer contract with a different single supplier is not independence.

The useful conclusion is that sovereignty matters more for some workloads than others. Some processes genuinely need to run somewhere you control. Most need to be able to run somewhere else at short notice, which is a different and considerably cheaper requirement.

The most practical thinking on this is not coming from the sovereignty debate at all. It is coming from operational resilience regulation, where the questions have been worked through properly.

In July, HM Treasury designated the first Critical Third Parties under the UK regime, following a Treasury Committee recommendation that major AI and cloud providers be brought into scope by the end of the year. The FCA has been explicit that dependencies on model providers must be mapped and governed, and its one-year review of the resilience regime found a telling weakness: firms had mapped their internal technology but treated external providers as endpoints, rather than mapping through them to the services underneath.

That is exactly the gap June exposed, and the framework for closing it is public, mature and free to borrow. Financial services firms will have it applied to them. Everyone else would be sensible to adopt it voluntarily, because nobody has produced a better tool.

Government policy is worth keeping in view as one input among several. The new administration has signalled more interest in domestic ownership and tech sovereignty than its predecessor, and some existing arrangements are under review; however, the direction is contested and the detail is not yet settled.

Start by mapping dependencies: which business services would degrade, and by how much, if a named model became unavailable? In an era where you’ve outsourced thinking and doing to AI, is it even feasible to fall back on your remaining humans?

Then set a switching tolerance for each of those services, in the same spirit as an impact tolerance. Four hours for one, a fortnight for another, never for a third. The numbers will be uncomfortable, which is the useful part.

Keep a second option warm rather than theoretical. An alternative model that has been run against your evaluation set, on your real data, at least once, is worth more than a page of vendor-neutral architecture diagrams.

And separate the layers deliberately, particularly your data and your orchestration, so that changing a model is a configuration decision rather than a big project. You might also want to invest in some old-fashioned knowledge management with humans, as a backup for when the AI isn’t there.

None of this requires a view on where AI should be built or by whom. It requires knowing what you would do on the morning it stops working. In June, a lot of organisations found out they hadn’t thought about it. The next time it happens, that will be harder to describe as bad luck.