GOV.UK is adding an AI chatbot. Here's what government teams need to know
The Government Digital Service (GDS) plans to add an AI chatbot to the GOV.UK app in early 2026, then expand it across the main GOV.UK website used by most departments and services. The goal: let people ask plain-language questions like "I've just had a baby - what help can I get?" and receive one clear answer that draws on guidance from HMRC, DWP, DfE, and others.
GDS is also exploring whether the chatbot could handle simple transactions in time. Ukraine's Diia has already taken steps in this direction, generating income certificates on request and aiming for 24/7 automated assistance.
Where it starts
The first deployment will be inside the GOV.UK app, which has been in public beta since July and reached nearly 260,000 downloads by November 24. More than 80 percent of users have customized their app homepage, including adding their local authority - a good sign people will adopt features that save time.
How it works under the hood
The chatbot is built on OpenAI technology with retrieval-augmented generation (RAG). It pulls from GOV.UK content with personal data stripped out, so it can reference official guidance rather than inventing answers.
GDS says it has run extensive testing, including red-teaming by government colleagues trying to break or corrupt it. The system has been in development for more than two years.
What went wrong in the 2023 pilot - and what changed
A private pilot with 1,000 users in late 2023 showed that people liked using the chatbot, but accuracy fell short. It also produced some outright mistakes.
Since then, GDS has added filters and rules to block certain questions or responses. In other words, the bot will sometimes refuse to answer rather than guess - a trade-off most service owners will prefer to a wrong result.
What this means for departments and agencies
If your guidance lives on GOV.UK, your content quality now directly affects what the chatbot says. Out-of-date or fragmented pages will surface as confusing answers or refusals.
Expect a shift in how users arrive at your services. More queries will start as open questions instead of clicks through navigation. Your best defense is clear, current, structured guidance.
Practical steps to prepare
- Audit your GOV.UK content for accuracy, clarity, and duplication. Fix contradictions across pages.
- Front-load pages with the essential facts and eligibility rules the bot will need to cite.
- Add plain-language headings and question-style phrasing that map to what people actually ask.
- Define "safe to answer" vs "must refuse" topics for your policy area. Share these with GDS.
- Set up a rapid correction loop: when the bot gets something wrong, who updates the source content and how fast?
- Coordinate with press/comms and service owners on escalation paths for sensitive topics.
- Train frontline teams to spot chatbot-induced confusion and feed issues back to content owners.
Risks to manage from day one
- Accuracy and citations: users will expect grounded responses. Push for clear references to official guidance.
- Over-blocking vs. bad answers: filters reduce risk, but refusals can frustrate users. Monitor refusal rates by topic.
- Content drift: policy updates must propagate quickly or the bot will lag. Time-box content reviews after changes.
- Edge cases: the bot may struggle with complex eligibility or multi-department scenarios. Provide hand-offs to human channels.
Transactions on the horizon
GDS is exploring simple transactions after launch, taking cues from services like Ukraine's Diia. If this proceeds, expect deeper integration with departmental systems and stricter governance on identity, consent, and audit trails.
What success looks like
- Fewer clicks to the right answer, higher task completion, lower avoidable contact.
- Reliable refusals for unsafe questions, paired with clear next steps.
- Fast content corrections when issues surface, measured in hours not weeks.
- Transparent reporting: accuracy, refusal, and escalation metrics visible to service owners.
For policy leads, content designers, and service owners, the work starts now: clean up guidance, define guardrails, and agree update processes. That's how you make the chatbot an asset rather than another support channel to triage.
For ongoing updates and implementation detail, watch the GDS blog. If you need to upskill teams on prompts, RAG-aware content, or red-teaming practices, see our courses by job.
Your membership also unlocks:
