Defra opened an AI adviser to the public. Its own tests show why users still need the source pages
The UK department's public beta answers questions from a 40-page AI toolkit and checks quoted rules against their source. Defra also discloses that its model has not yet met every grounding test and retains typed questions and answers for 30 days.
The short version
- What happened
- Ask the toolkit is a public, no-sign-in beta that uses Claude Sonnet 4.6 through Amazon Bedrock in London to answer only from Defra's published AI guidance.
- Who it affects
- Civil servants and other readers can reach a cited rule more quickly, but they still need to open the guidance before relying on an answer and should not enter personal, incident or sensitive details into the prompt.
- What remains unproven
- Public beta
Keep in mind
Public beta
Internal evaluation only
AI judge with limited human agreement checks
Two grounding criteria missed
No representative public sample
No Welsh-language test
No formal fairness assessment
No usage or outcome data at launch
Source analysis · 2 linked sources. Primary analysed source date: 27 October 2026.
Read the full report · 5 min The evidence, methods, implications and limitations
A narrow assistant is now a live public service
Defra has opened Ask the toolkit, a public beta meant to help people navigate about 40 pages of departmental guidance on using AI. A visitor can ask without signing in. The service sends the question, up to four earlier exchanges and the full toolkit to Claude Sonnet 4.6 through Amazon Bedrock in the AWS London region. It has no tools, web access or agent loop and cannot approve a project, grant access or make a decision about anyone.
That constrained design matters for public servants trying to find a rule quickly. Answers link back to the pages used, and a separate check removes any quoted rule that cannot be found on the named page. When the toolkit does not cover a question, the system should say so or point the reader to a person. The public page still warns that AI can make mistakes and that the guidance should be checked before acting.[1][2]
The published tests show useful controls—and unresolved grounding failures
Defra's fixed internal test set is unusually visible. In three runs on 19 October, the service returned the correct status for 91 to 92 of 100 questions, held all seven expected refusals, produced complete answers for 62 to 65 of 71 applicable questions and made no invented quotes. The department counts those as four passed criteria. It missed two stricter criteria: only 16 to 19 of 22 answers quoted the expected rule, and 87 to 90 of 100 stayed grounded in the named pages when every answer had to pass.
Those numbers are not an independent accuracy benchmark. Claude Opus 4.6 judged the automated tests, while two people agreed with 38 and 27 of 40 sampled judge decisions. Defra therefore says the grounded and complete marks should be treated cautiously. Hand-tested prompt attacks all held in one round; 16 of 17 passed in a second, where the failure was an appropriate question wrongly refused. Testing has not covered Welsh or a representative public sample, and no formal fairness assessment has been completed.[1]
The distinction that matters
What the beta checks—and what it cannot prove yet
The public record separates source controls, internal evaluation and real-world outcomes.
Source control
Send the full toolkit and verify quoted rules
Links and exact-quote checks are live
Internal test
Run a fixed question set and prompt attacks
Four of six criteria passed; two grounding criteria missed
Public outcome
Measure correct rule-finding, help requests and harm
No launch usage or outcome data yet
Privacy and scale limits are part of the evidence
The conversation remains in the reader's session for four hours, but the platform logs questions and answers as typed for 30 days so the team can investigate problems or misuse. That means personal information entered despite the warning can remain in a restricted log before a Bedrock guardrail masks some details for the model. A reported problem can also be retained by the team for 90 days and by GOV.UK Notify for seven days. Defra says a full data-protection impact assessment was screened out because the service does not set out to collect personal data.
At launch there was no usage denominator or evidence that people find rules faster, make fewer mistakes or need less staff support. The beta stops after 600 model calls a day, limits a conversation to 20 questions and caps each question at 500 characters. Twenty simultaneous test questions were answered in typical times of 4.6 and 5.2 seconds on two October test days, but these are load checks rather than public-service outcomes. Evidence that would change the assessment includes independent answer audits, real usage and escalation data, Welsh testing, published incident reporting and proof that readers reach the correct rule more reliably than with search alone.[1][2]