What the FOI response reportedly shows
According to the FOI response, 10 of the 11 AI tools deployed or bought by the Cabinet Office use AI models developed by US tech firms. The models are hosted in the UK, but their design, access and pricing are controlled by US organisations.
The suppliers named are Google, OpenAI, Microsoft and Anthropic. Computer Weekly says they are used either directly or through Microsoft Azure and Amazon Web Services. The only exception is PolyAI, a UK company whose models are used for the government's shared services provider, SSCL.
The Cabinet Office confirmed that Google Gemini is the primary authorised enterprise-level generative AI tool. Other tools include Hansard AI for parliamentary and legislative research, Minute for meeting notes, and GCS Assist for government communications.
One caveat applies throughout. The underlying FOI document was not available to inspect. The tally of 11 tools and the supplier breakdown are Computer Weekly's account of that response. The figures describe the Cabinet Office only, not the whole of UK government.
The "house of cards" argument
Aventail's CEO, Joe Ward, said that hosting data in the UK helps, but it does not change who controls the models, the pricing or the terms of access. He said he supports government use of the latest US models. His warning was that essential services are fragile if they depend on continued access to any one supplier. His remedy is infrastructure that lets public bodies switch providers and run open-weight alternatives under their own control.
Aventail has a commercial interest here. It sells a platform for model deployment and routing that can run supported models on a customer's own hardware or route requests to approved external AI APIs. That doesn't make Ward wrong. It does mean "house of cards" is an advocate's framing, not a technical audit or a report of an incident.
The continuity question stays unanswered
The FOI request went to the Home Office, which Computer Weekly says manages the government's AI and technology supplier relationships. It asked for business continuity plans, risk assessments or risk register entries covering an outage of, or loss of access to, those providers' services.
The government's reply, as reported, was that the information exists. It declined to disclose it, citing prejudice to the prevention or detection of crime. So we know plans or risk entries exist. We don't know what they say, and the reporting contains no evidence that the government lacks contingency plans.
Outside views
Mark Lewis, an IT outsourcing lawyer at Stephenson Harwood, told Computer Weekly that the concentration goes beyond the model labs. He pointed to the underlying infrastructure, which he said is inevitably US cloud. He also warned that AI labs could lock customers in just as cloud providers did.
Lewis was careful, too. He said it is unknown whether government negotiated smart commercial terms, though he doubted its bargaining position given its procurement record. That is informed scepticism, not evidence about particular contracts.
A worked example: GCS Assist
Government's own Algorithmic Transparency Recording Standard record for Assist shows how this looks in practice. The record was first published on 4 June 2025 and last updated on 30 July 2026. It describes a tool for government communicators doing brainstorming, first drafts and reviewing work.
The record says:
- Assist sends requests to its own API, which accesses a large language model hosted on AWS Bedrock.
- All infrastructure and data processing sits in an AWS environment, and Anthropic, the model's original trainer, never receives Assist data under an AWS "escrow account" arrangement.
- Amazon does not receive user prompts, outputs or analytics for third-party uses, including model training.
- As of July 2026 it uses Claude Sonnet 5 as the core model, with Haiku 4.5 for some background tasks.
- It was available to around 8,500 members of the government communications profession in July 2026.
- It is not designed for automated decision-making. Users remain responsible for what they do with the outputs.
The record also says that during alpha the team compared OpenAI GPT-3.5 and GPT-4 against Anthropic Claude 2.1, then chose Anthropic models for quality against cost. That detail is useful for the lock-in debate. The model layer was chosen by comparison, and a team that has benchmarked alternatives once can, in principle, do it again. The record doesn't describe a tested fallback or exit plan, though. It also doesn't mention data residency in the UK, so treat Assist as a single documented case, not proof of how every tool is built.
Hosting versus control
The headline blurs two separate things:
- Where data is processed. This is about residency, privacy and data protection, and the Assist record describes specific safeguards.
- Who controls the model. This is about pricing, terms, version retirement, access and availability.
UK hosting addresses the first and not the second. A foreign model provider does not, on its own, mean UK government data flows to that provider. The architecture and the contract decide that. Concentration risk is a continuity and bargaining-power issue, not automatically a data-leak issue.
Minute is a useful contrast. Its service page says it transcribes meetings and drafts minutes, and is suitable up to OFFICIAL SENSITIVE. It says access is expanding across departments. The page doesn't name a model supplier or hosting arrangement, so nothing should be inferred about it from the overall vendor count.
What IT teams can take from this
You don't need to run a government department to apply the lesson. Practical steps for admins and architects:
- Inventory AI dependencies by model, not by tool. Several tools can sit on one provider, so ten products may be one point of failure.
- Separate the interface from the model. Where a retrieval layer, prompts and user workflow live in your own code, swapping the model is a configuration task. Where they are fused to one vendor's service, it is a migration.
- Test an alternative before you need it. Run a benchmark on your own tasks against at least one other provider or an open-weight model, and keep the results.
- Read the exit terms. Check notice periods for model deprecation, pricing change clauses and data return provisions.
- Keep the human check in place. Government guidance to civil servants warns that generative AI output is prone to bias and misinformation and must be checked. Its earlier guidance also says not to enter classified or sensitive information into public tools. That guidance has been superseded by the Generative AI Framework for HMG, so use it as background, not as current policy.
The bottom line
The reporting supports a narrow claim: by Aventail's FOI count, nearly all Cabinet Office AI tools run on US-developed models. Neither the reporting nor the government records show an outage, a supplier withdrawal or a failed contingency plan. What's missing is visibility, because the continuity documents exist but are withheld.
As Lewis noted, the government may have negotiated well, but "we don't know" is a weak foundation for essential services. For Windows and Azure administrators, it is a familiar lesson. Convenience and capability tend to arrive first, and the exit plan gets written later, if at all.