TL;DR

The Government Digital Service is testing whether machine learning can help records staff work through the email departments have accumulated over decades. The team’s own assessment is that reviewing it by hand at government scale no longer works. Two components have come out of phase one, and one of them solves the data problem rather than the email problem.

The pile

Knowledge and information management specialists can face hundreds of thousands of individual items, deciding for each what must be kept, what legal and compliance duties apply, and what is sensitive enough to need careful handling. The material runs from personal notes and routine administrative chains through to sensitive exchanges and records of business decisions, with no clean separation between them.

Project Kestrel sits with the AI Data Readiness Incubator inside GDS, developed alongside the government’s records team.

Solving the data problem first

The obstacle was that the team could not safely experiment on the thing they were building for. Real government email carries personal, sensitive and legally protected material, and using it to develop tooling creates exposure.

So they built Tiger Heron, which leans on language models to manufacture plausible fake email resembling what departments actually hold, carrying no genuine content. Development and testing then happen without anyone opening a real inbox.

The second component, Cuckoo, does the intended job. Within a controlled environment it reads at volume and groups messages into clusters by content, presenting them through a visual interface so reviewers see topics and threads rather than an undifferentiated queue. GDS frames the aim as helping people spot patterns and document decisions they can defend, with humans keeping the calls that matter.

Looking forward

Phase one also ran user research with records professionals, took ethics input and produced best-practice material. GDS is explicit that this was an experiment, not a product, and that more testing is needed.

The transferable idea is Tiger Heron, not Cuckoo. Any UK organisation holding sensitive correspondence hits the same wall — you cannot develop a tool on data you are not allowed to expose. Generating a synthetic stand-in is a well-understood technique that few outside government bother with, and it is the part of this pilot most likely to be worth copying.