UK Home Office
UK Home Office pilots AI summaries of asylum interviews and measures 23 minutes saved per case
The UK Home Office piloted a large language model tool that summarises asylum interview transcripts for decision-makers, with a comparison group of colleagues working the usual way. Test users reviewed transcripts 23 minutes faster on average, but 9% of summaries were judged inaccurate or incomplete and pulled from the pilot.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
23
minutes
Average time saved reviewing an interview transcript, test group versus comparison group (pilot of the summarisation tool)1
32%
Share of transcript review time saved, test group versus comparison group1
9%
Summaries judged inaccurate or incomplete by technical specialists and removed from the pilot1
23%
Users of the summaries who reported they were not fully confident in the summary information1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Asylum decision-makers spend a large part of each case reading interview transcripts, which can run to 50 pages or more. Re-familiarising themselves with a case adds to a heavy mental load, and the Home Office wanted to see whether an AI tool could shorten that step without weakening the decision.
What they built
The Asylum Case Summarisation tool uses a large language model, an AI system that reads and writes text, to turn an existing interview transcript into a short summary. It is designed as an aid. A decision-maker cannot use it to decide a case, and the full transcript stays available. A second tool, Asylum Policy Search, a chat interface for country policy notes, was piloted separately, in other units; this case covers the summarisation tool only.
How they did it
The Home Office ran the summarisation pilot in two decision-making units, with a test group and a comparison group working as usual. Phase one lasted eight weeks, in May and June 2024, and compared 60 decision-makers using the tool with 15 working as usual. Phase two lasted eight weeks, in September and October 2024, and compared 45 users with 15. Everyone logged each case, giving 334 test cases and 95 comparison cases. Technical specialists checked every summary before use. The team also ran surveys and interviews and sampled decisions for quality.
Results
Decision-makers using the tool got through transcripts 23 minutes faster on average than the comparison group, a 32% saving. A dip sample of decisions showed a similar share of sustained decisions in both groups, so the tool neither helped nor hurt quality on that check. Technical specialists judged 9% of summaries inaccurate or incomplete, and those cases went through the normal process. Among users of the summaries, 23% were not fully confident in them.
What went wrong
The main weaknesses were accuracy and traceability. The 9% of summaries that failed the specialist check were the main accuracy problem, and users also wanted references back to the transcript. The report says these limits should be fixed before any full roll-out.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Average time saved reviewing an interview transcript, test group versus comparison group (pilot of the summarisation tool) | 23 minutes1 | |
| Share of transcript review time saved, test group versus comparison group | 32%1 | |
| Summaries judged inaccurate or incomplete by technical specialists and removed from the pilot | 9%1 | |
| Users of the summaries who reported they were not fully confident in the summary information | 23%1 | |
| Survey respondents who felt the summary helped them quickly understand the case | 77%1 |
Lessons for your project
Test against a comparison group. The Home Office logged cases for both a test and a comparison group, so the time saving is a difference between groups and not only an opinion.
Check every output before use while the tool is new. Technical specialists reviewed all summaries before decision-makers saw them, which is how the weak ones were caught.
Keep the person in charge of the decision. The tool was built as an aid that cannot decide a case alone, and users said it did not influence their decisions.
Design for traceability. Some interviewees saw little time saving because summaries did not point back to the place in the transcript.
Caveats
- The Home Office evaluated its own tools and published the evaluation. The report calls its findings an indication from small-scale pilots and says a full roll-out would need further evaluation.
- All the evaluation methods rely on self-reporting, and pilot units were chosen by availability, not to be representative.
- The report does not name the model. The Open Rights Group, a digital rights campaign, says the tool uses OpenAI's GPT-4 through Microsoft Azure, but that comes from the campaign, not the Home Office page.
- The Open Rights Group strongly criticises the tools and disputes their reliability; its page is dated 14 January 2026. It repeats the Home Office figures and adds none of its own, so the results rest on the Home Office report alone. We did not rely on its arguments.
- The date is the start of the first pilot phase in May 2024. The report gives pilot phases of May to June and September to October 2024; the tool was not in routine use at the time.
- The second-phase survey had only 6 responses (13%) and was left out of the report's analysis. The quality check showed similar rates of sustained decisions in both groups on a dip sample.
Sources
- Evaluation of AI trials in the asylum decision making process
GOV.UK (Home Office) · 2025-04-29 · Research paper · gov.uk
- Automating the hostile environment: AI in the asylum decision making process
Open Rights Group · 2026-01-14 · News · openrightsgroup.org
Summarized in our own words from the sources above. Spotted an error? See corrections.