Summary

• A freedom of information response obtained by British company Aventail identifies US-developed models in ten of eleven AI tools deployed or purchased by the Cabinet Office.
• Google, OpenAI, Microsoft, and Anthropic feature among the providers; the reported exception is a service using technology from UK company PolyAI.
• The Cabinet Office says relevant continuity information exists but has withheld it on crime prevention and detection grounds.

Ten of eleven artificial intelligence tools deployed or purchased by the Cabinet Office depend on models developed by US technology companies, according to a freedom of information response that reveals the concentration of suppliers behind parts of the government’s expanding AI infrastructure.

The findings, obtained by British AI company Aventail, identify technology from Google, OpenAI, Microsoft, and Anthropic among the models used directly or through cloud platforms. The only reported exception is a system using technology developed by UK company PolyAI for shared services provider SSCL.

Google Gemini is the Cabinet Office’s primary authorised enterprise generative AI tool. Other identified applications include Hansard AI for parliamentary research, Minute for preparing meeting notes, and GCS Assist for government communications.

The findings concern a defined set of eleven tools associated with the Cabinet Office, rather than every AI application deployed across central government. They nevertheless provide evidence of a substantial reliance on a small group of external model developers at a time when ministers are investing in domestic AI capabilities and greater control over critical technology.

Government applications depend on external models

Departments can develop their own interfaces, workflows, and information systems while obtaining the underlying AI capabilities through models maintained by commercial providers. Such arrangements can reduce the cost and time required to deploy applications, particularly where a government organisation does not need to train a large model itself.

The resulting services can appear to be government applications even when their main processing functions depend on external technology. That distinction affects commercial control because the organisation operating the application may not determine how the underlying model is developed, priced, or made available.

Hosting arrangements add another layer. The FOI-based reporting states that the identified services use UK-hosted infrastructure, but locating data processing within the country does not necessarily give the customer control over the company developing the model or the contractual terms governing its use.

A model provider may introduce new versions, retire older interfaces, change usage charges, or modify available functions. Application owners must decide whether those changes affect accuracy, security, operational processes, or the cost of providing their service.

Dependence is not necessarily inappropriate where a supplier offers suitable functionality, reliability, and contractual safeguards. The difficulty arises when organisations cannot establish how much work would be required to replace a service or continue operating if access to it became unavailable.

For applications supporting routine administrative work, alternative tools may be relatively straightforward to introduce. More specialised systems can be harder to replace if they depend on particular model behaviour, supporting databases, or carefully tested workflows.

Assist illustrates the underlying architecture

The government’s published Algorithmic Transparency Recording Standard entry for Assist provides a more detailed example of how one Cabinet Office application operates. Developed for the government communications profession, Assist supports drafting, reviewing, and researching material while leaving responsibility for the final output with the employee using it.

Its July 2026 transparency update identifies Anthropic’s Claude Sonnet 5 as the principal model, with Haiku 4.5 supporting some background tasks. The application also uses retrieval-augmented generation, drawing on approved guidance and documents to provide information relevant to the user’s request.

The system was developed and maintained by government teams, but the underlying language models are supplied through an external provider. That arrangement allows the service to benefit from model developments without requiring government developers to maintain equivalent AI infrastructure themselves.

Changing providers would still involve engineering and evaluation work. Different models may interpret instructions differently, cite information in different ways, or produce outputs that vary in structure and reliability, even when supplied with identical supporting documents.

The application would therefore require renewed testing before another model could be treated as interchangeable. Government teams would need to confirm that it continued to retrieve the intended material, follow relevant instructions, and produce acceptable results for the tasks involved.

Assist is not designed to make automated decisions, according to its transparency record. Users remain responsible for reviewing its outputs, providing a degree of human oversight that differs from applications capable of executing transactions or initiating operational changes.

Continuity arrangements remain undisclosed

Aventail’s information request also sought details of contingency plans, risk assessments, and risk register entries concerning the loss of access to major AI suppliers.

The Cabinet Office acknowledged holding relevant information but declined to disclose it, arguing that publication could prejudice the prevention or detection of crime. The response leaves the public without the details needed to assess those arrangements, although it does not establish that contingency planning is absent.

Continuity arrangements for AI applications can differ from conventional disaster recovery. Restoring servers and databases may be insufficient if the application requires a model service that has become unavailable or cannot be accessed under its existing terms.

One possible approach is to support several model providers through an application architecture that reduces dependence on a single interface. However, alternative models must be evaluated for the actual tasks they are expected to perform, rather than assumed to be equivalent because they support similar inputs and outputs.

Open-weight models can provide another option where organisations have suitable computing resources and technical expertise. Operating those models independently may offer greater control over deployment and upgrades, but it also transfers responsibility for maintenance, performance, security, and the underlying infrastructure.

Maintaining several commercial suppliers can improve flexibility, although the arrangements may still share infrastructure dependencies. Contracts with different model developers do not necessarily guarantee independence if their services rely on the same cloud region or supporting platform.

Procurement determines future flexibility

Supplier dependence can be influenced long before an application becomes operational. Procurement specifications can require information about model portability, supported interfaces, data handling, continuity arrangements, and the consequences of changing a provider.

Application architecture is equally important. Systems that separate their business processes and information repositories from the underlying model may be easier to adapt than services built around functions available from only one supplier.

Such flexibility carries costs. Supporting alternative models involves additional testing, integration work, and operational management, which must be weighed against the potential consequences of prolonged dependence on a single provider.

Lawyer Mark Lewis of Stephenson Harwood noted in the reporting that the government might have negotiated contractual protections reducing its exposure. Without access to the relevant agreements, the extent of those safeguards cannot be established from the supplier list alone.

The FOI findings consequently provide evidence of concentration rather than a complete assessment of procurement quality, system security, or resilience. The eleven applications may also differ substantially in importance, meaning a simple count does not measure the operational consequences of losing an individual service.

For central government, domestic AI investment and international supplier relationships are not mutually exclusive. Departments can continue using established commercial models while investing in UK research, computing capacity, and alternatives for workloads where greater technical control is justified.

The Cabinet Office’s position illustrates the difference between being able to deploy AI quickly and retaining long-term flexibility over how that capability is provided. The extent of the government’s preparedness for changing suppliers remains difficult to assess without the continuity evidence it has declined to release.