We may have tolerated the expense and dependency of those decisions for compute. But artificial intelligence is not just someone else’s computer, it’s someone else’s mind.
What government is now outsourcing is not storage or processing power. It is judgement itself.
When a department uses a managed AI service to help determine whether you receive a benefit, how your immigration case is prioritised, or whether a fraud flag is placed on your account, a question arises that policy has not yet answered honestly. Who is actually making that decision?
If a model belongs to a vendor, the version running this morning may not be the version running tomorrow, and the department may not be notified. The training data, the internal logic, the decision-making systems all sit behind proprietary APIs, accessible to government only as outputs. The model could even disappear tomorrow, leaving government scrambling.
The department owns the consequences, while the vendor owns the stack.
Accountability you can’t actually demonstrate
Since March 2024, central government has been legally required to document and be accountable for automated decision-making tools that have a significant influence on decisions with public effect.
For AI running on open, inspectable infrastructure, that standard can be met with evidence. For a managed API, you fill in the compliance form using whatever the vendor told you. Critical parts of the audit chain remain vendor-controlled. The department cannot independently verify which version processed a given case, or reproduce the logic behind a specific outcome.
One of these approaches enables departmental accountability, the other mimics it.
The counterarguments
Some will point to the ability to import custom model weights into these managed platforms as evidence of agency, choice, and control. It is not. Owning the weights while the vendor controls the inference pipeline, the orchestration layer, and the surrounding compute is roughly analogous to owning the recipe while the restaurant controls the kitchen, the ingredients, and the bill. The department may determine what the model is, but the vendor determines how, where, and at what cost it runs. The dependency has simply moved one level down.
Those who defend the status quo tend to invoke UKCloud – a domestic provider that collapsed in 2022, costing the Cabinet Office £17 million and stranding government workloads across the country. This failure is frequently presented as proof that concentration around the large hyperscalers is an unfortunate market reality rather than a policy choice. But that interpretation arguably gets the lesson backwards. UKCloud failed because its clients were locked into its proprietary platform. Any platform lock-in is fragile, therefore the cure is portability and interoperability, not resignation.
An open stack for AI, running on commodity compute, can be picked up and moved. Change the provider, keep the capability. The infrastructure becomes interchangeable, which is, in many ways, what Cloud First was originally intended for.
A contradiction at the heart of policy
Which makes the government’s current position particularly difficult to defend. With one hand, it is making smart investments in sovereign AI, with supercomputers in Bristol, GPU clusters designated as national infrastructure, and a stated ambition to make Britain the global home of open source AI talent. With the other, it is signing procurement agreements that build new proprietary dependencies on top of old ones.
The government has committed over £2 billion to sovereign compute infrastructure on the premise that Britain would have genuine AI capability of its own. But if departmental AI workloads are locked into managed AI services then they cannot move. Our sovereign compute investment will return nothing if the AI workloads running across government are structurally incapable of migrating to it.
You cannot have a sovereign AI strategy and a managed-API procurement default for government. These two positions are oil and water.
The fix is not complicated. Departments are required to demonstrate a viable exit strategy before technology spend is approved (albeit this is frequently overlooked). That requirement should be extended to AI, with one clarification that exit means being able to move the entire working system to different infrastructure without rebuilding it. Not owning the model weights or switching between models on the same vendor’s platform. Actually being able to pick it up and run it somewhere else. That is a testable, demonstrable standard – and it is not currently being met. For high-stakes decisions, including benefits, immigration, fraud, that standard should be non-negotiable.
None of this requires leaving the cloud. Renting someone else’s tin is fine. But letting someone else own your systems of judgement for the sake of efficiency is a category error we will spend the next decade correcting, and regretting.
This is not a question of whether we should use powerful infrastructure. That should be a given. But, should using someone else’s compute leave us nationally exposed to someone else’s strategy, someone else’s commercial decisions, and, when a citizen needs to understand why an algorithm changed their life, someone else’s silence.
For a government accountable to its citizens, the answer has to be no.
It was no in 2013, when the policy so many individuals are quick to invoke was created, and it should be no now.
Except now the stakes are much, much higher.
Read More Regulation
