TL;DR
The UK government says the AI tools it builds in-house are not tied to any one model supplier. A spokesperson told Computer Weekly, in a report published on 9 November, that tools from the Government Digital Service (GDS) and the Incubator for Artificial Intelligence (i.AI) use “decoupled, model-agnostic architectures” that let officials swap or self-host models. The statement follows a freedom of information response which found US-built models behind 10 of the 11 AI tools central government uses.
The government’s answer
The FOI, requested by the UK AI firm Aventail, listed Anthropic, Google, Microsoft and OpenAI as the model sources, reached either directly or via AWS and Azure. The one non-US supplier was PolyAI, a British company whose models serve the shared services provider SSCL. We reported the FOI findings last week.
The spokesperson said that keeping data pipelines, interfaces and retrieval separate from the underlying model means “we avoid hardcoding government applications into any single vendor’s ecosystem”. Commercial models under secure enterprise agreements, they said, had let civil servants prototype quickly, while modular API and orchestration layers keep the “operational resilience to evaluate, swap, or self-host alternative models as capabilities mature”.
The Cabinet Office also named Google’s Gemini as its main authorised enterprise generative AI tool. It said it has a contingency plan, but withheld the details on crime-prevention grounds.
Aventail’s test
Joe Ward, Aventail’s co-founder and chief executive, welcomed the approach. “Model-agnostic design is the right approach and the government deserves credit for thinking about model flexibility,” he said. But he set a harder test: “if the US president revoked access to every US model overnight, would our AI systems stay afloat? Resilience depends on both adaptability and speed.”
What it means
The statement fits the line the AI minister, Kanishka Narayan, took in a speech last month, as quoted by Computer Weekly: “We want the world’s leading companies to invest here, but we will not accept dependence as the price of that investment.”
In our view, the government has answered the architecture question but not the operational one. Being able to swap a model in principle is different from having a tested alternative ready. GOV.UK Chat, for example, is built on Anthropic’s Claude, according to a separate ministerial answer reported by PublicTechnology, which makes it an obvious first test of how quickly a switch could be made. For public bodies outside central government buying AI tools, the useful question to put to suppliers is the one Ward raises: how long would a switch actually take?