Why we are building a way to ask the toolkit
Why we are adding a way to ask the toolkit questions, and how it builds on what GDS learned from GOV.UK Chat.
I lead design for the AI Capability and Enablement team, including the AI digital toolkit. In this post I explain the problem Ask the toolkit solves, how it builds on GOV.UK Chat, and how we keep people in charge of the answers that artificial intelligence (AI) writes.
Search needs the right words
The question we hear most is "Can I put this data into an AI tool?" The answer is on a page called Using data with AI. But you have to know that page exists, and what it is called.
We searched our own site for "can I use copilot with personal data" and got 20 results. None of them was the right page. So people email us instead, and wait.
Asking keeps the detail that matters
When you ask a colleague, you tell them what you are doing. "I'm a supplier." "It's test data with real names in it." That context changes the answer. A search box throws it away. A conversation keeps it, and lets you ask a follow-up.
Conversation is a new way to use government
In May 2026, the Government Digital Service (GDS) launched GOV.UK Chat in the GOV.UK app. People can ask about government in everyday words and get one joined-up answer. GDS calls it the biggest change to how people use government content since GOV.UK launched in 2012.
Chat started in the app, and GDS intends to make it available on the GOV.UK website too. Conversation could become a way to use government wherever people are. Designing it well is a new skill, and every department will need it.
Ask the toolkit builds on what GDS has shared about designing Chat. It is a small, specialist service on the web, for Defra's digital teams. That means we can try ideas quickly, and share what works across government.
AI can be wrong and still sound sure
That is the risk with a conversation. AI writes fluently, so a wrong answer can sound as convincing as a right one. We gave each job to whatever does it best.
- You ask in your own words
- AI drafts from the guidance
- Code checks quotes and links
- You decide with the sources
The guidance holds the rules. When an answer quotes a rule, code checks the words are really on the page. If they are not, the quote is removed. People write the guidance, and answer what the AI should not.
Ask is also honest about what it is. It has no name or face, and we designed it never to say 'I'. Every answer ends with 'AI can make mistakes. Check the guidance before you act.' And a person is always one step away, even when the AI is down.
We build on how GDS tests GOV.UK Chat
GDS checks Chat's answers against clear measures, such as whether an answer comes from GOV.UK, and whether it is accurate and complete. AI marks the answers at scale, and experts check them by hand.
We do the same. We have more than 100 test questions, each with an answer we would accept. An AI marks every answer, and people check a sample. We also test for the risks in our own service:
- whole conversations, to check Ask keeps hold of what you told it earlier
- questions only Defra teams ask, such as which tools they can use with which data
- the marker itself, to check it marks fairly
That last one mattered most. People found the AI marker was too kind. It passed 13 of 21 answers that quoted a rule from the wrong page. So we made the rule stricter, and the scores went down. Ask had not got worse. The marking had got honest.
What guides our work
AI can help you find an answer faster. The guidance, and the people who write it, stay in charge of what the answer is.
Before we tested Ask with people, we tested it with AI. Sharing our AI skills for user-centred design explains how, and shares the skills we built to do it.
Want to help us test Ask the toolkit before it launches?
Email AICapabilityAndEnablement@defra.gov.uk