How GOV.UK Chat turns government guidance into a conversational answer

GOV.UK Chat gives people a way to ask about government services without first knowing which department or page contains the answer. Officially launched in the GOV.UK app on 14 May 2026, following a soft release in March, it draws on roughly 80,000 pages of guidance. Early demand concentrated on tax, benefits, driving and transport.[1]

That is a practical use of conversational AI: a person can describe a situation in ordinary language, then follow a more focused route through the information. The useful result is an understandable answer connected to the appropriate official page, including when the question spans more than one subject.

How the answer is assembled

The system combines search with generation. Before searching, GDS filters the available material and divides pages into meaningful sections while preserving their heading structure. Search matches the meaning of a question, then reranks results using information about the documents, helping current, relevant guidance rise above a merely similar passage.[3]

A routing step decides how to respond: answer from guidance, ask for clarification, or handle a request that falls outside the service’s scope. The model receives the selected material and produces a structured response. GDS also checks questions before generation and checks answers afterwards. This is how the engineering team connects a broad language model to a narrower public-service task.[3]

The privacy notice makes that scope concrete. Chat has no access to somebody’s passport application or Universal Credit account. It searches filtered GOV.UK pages, excluding external government websites and separate departmental service domains. Claude models run through Amazon Bedrock within GDS-controlled accounts; the notice says Anthropic cannot access the processed conversations.[5]

Reading the pilot results

GDS tested a web version with 10,136 users and an app version with 641 users. Its March 2026 report separated answer quality from whether people found the experience useful. Accuracy was assessed through subject experts and automated tools, with completeness included in the standard.[2]

In practical terms, a service can answer a question yet omit a condition that matters; it can also decline a question it might have answered after clarification. The separate measures help identify which part of that experience needs improvement. The survey describes the respondents’ views, rather than establishing satisfaction across the population.

Designing for a phone

The August design account explains several changes for smaller screens. GDS removed the chatbot avatar to recover space, distinguished messages through colour and contrast, and supported enlarged text. VoiceOver and TalkBack can announce the input character limit. These details affect whether a person can comfortably ask a complicated question and read the result.[4]

The introductory screens explain the service’s scope, source checking and use of conversation data. Source links sit with the answer, so a reader can move directly from the explanation to the guidance. The team was also working on helping people formulate their first question, especially those unfamiliar with AI.[4]

Conversation history lets users continue across sessions. The privacy notice says clearing it from the device does not immediately erase the server copy: conversations and associated data are kept for 12 months. It also provides a route to request deletion. That is useful information before typing identifying details about a personal case.[5]

From finding information to completing the next step

For now, the documented service is an information assistant. GDS’s launch account describes further work on helping users continue to the appropriate department or support channel. It does not announce that Chat can submit applications or make decisions on someone’s eligibility.[1]

Its immediate value is easier navigation through a large body of official material. A good answer should leave the reader knowing which service applies, what information to prepare and where to continue. That offers a practical route to better access even before an assistant can carry out a transaction.

Report an error or suggest a correction