The UK public sector has awarded £1.41 billion in AI-related contracts so far in 2026, already the highest annual total on record, according to Tussell’s public sector AI procurement tracker. Adoption is moving ahead quickly, even as the abolition of the Department for Science, Innovation and Technology changes the way AI policy is coordinated across government.
MPs have asked for clarification on where DSIT’s responsibilities will sit. Public bodies, meanwhile, will continue agreeing contracts that could shape essential services for years. This makes sovereignty a practical procurement question. Institutions need to know how much control they will retain over the systems they buy and the conditions under which they could change provider.
Government policy has placed considerable emphasis on buying British. The location of a supplier is relevant, particularly when sensitive public data is involved, but it cannot show whether a customer can inspect a system, manage access to it or run it elsewhere. Those are questions of operational control.
British-owned does not always mean British-controlled
It is tempting to treat a local supplier as a sovereign choice by default. In practice, the relationship between the customer and the technology matters just as much as the address of the company providing it.
A council, hospital or court could use a model supplied by a British company while leaving most operational decisions with that company. The supplier may need access to maintain the service, but the public body should determine the terms of that access, be able to audit and verify it, and withdraw it when necessary.
The same principle applies to the model itself. Can the institution inspect how it has been configured, can it decide which information the model can use and could another qualified team operate the system if the commercial relationship ended? A British supplier may answer those questions well, although its nationality alone provides no assurance.
Overseas providers raise a further consideration because they remain subject to decisions made in their home jurisdictions. In June, a US export-control directive led Anthropic to suspend access to two of its models worldwide, because it could not verify users’ nationality in real time. Access was restored within weeks, but the episode showed that a provider can face legal constraints outside the customer’s control. Loss of control is more likely to emerge gradually than through a sudden interruption. A department may find that it needs the incumbent supplier to assess an update, reproduce a configuration or move a workload. Each dependency makes a later change more involved, particularly if the customer has not retained the relevant documentation and expertise.
Control will not come at the expense of capability
Public bodies often have different requirements from general commercial users. A model may perform well on public benchmarks while remaining difficult to evaluate for a particular hospital, court or government department.
For a defined public-sector task, a smaller model running in a controlled environment may offer a better fit. It can be evaluated against the organisation’s own requirements, tested under the conditions in which it will operate and supervised according to the consequences of an error.
There is no need for one organisation to control every part of the system. AI services rely on several layers, from infrastructure and computing capacity to models and applications. Different providers can be responsible for each. Public bodies should decide where direct control is necessary and where a managed dependency is reasonable.
For some services, control of the deployment environment may be enough. The organisation can set the model’s permissions, determine how its outputs are reviewed and retain the records needed to evaluate its performance. It can then consider models from several providers without changing the rest of the service each time.
Open-weight models and portable infrastructure can make that arrangement easier to maintain: an organisation can move a model between providers, examine its behaviour and replace individual components. Secure infrastructure, experienced staff and effective governance remain necessary, whatever licensing model is used.
]The value of portability becomes clearer over time. Models improve, prices change and regulations develop. A public body that can introduce another model without rebuilding its service has more scope to respond.
Making sovereignty part of public procurement
Government investment should support deployment environments that public institutions can operate or oversee directly. Where sensitive data is involved, those environments should also be verifiable: hardware attestation can now prove that a workload ran unmodified and that data stayed protected, even from the operator of the infrastructure. These environments could accommodate models from different providers while keeping access controls, evaluations and operational records under the authority of the institution responsible for the service.
This would be more useful than trying to select one sovereign model for every part of government. Smaller models may suit narrowly defined areas of government work, especially where the relevant data, users and risks are well understood. They can be adapted for a particular task and evaluated on representative material.
Larger commercial models will remain appropriate elsewhere. Public bodies need the technical capacity to make that choice rather than having it settled by the design of an existing platform.
Procurement rules deserve equal attention. Bidders should explain how a workload could be transferred, which parts of the system the customer can inspect, what evidence of data handling it can verify independently, and what information another provider would receive. Access to logs, testing methods, configuration records and technical documentation can be agreed at the start of the contract.
Britain will continue to depend on research, infrastructure and products developed overseas. Sovereign AI should give public institutions a clearer view of those dependencies and enough control to manage them.
As government responsibilities move beyond DSIT, the opportunity is to set consistent expectations across the public sector. Institutions should be able to use capable AI systems while retaining the knowledge, access and authority needed to change course. That is what sovereignty should mean in practice: not where a supplier is based, but whether an institution keeps control of its data, its systems and the knowledge they create.
