Caseworker documentation & copilots

A generative documentation assistant spreading through UK social care — the live test of human-in-the-loop at scale.

A state-built AI meeting scribe piloted through a governed council cohort and scaled nationally - an exemplary published governance record sitting on top of no published accuracy evaluation.

An in-house probation copilot that transcribes supervision sessions into the case record - scaled to every probation officer in England and Wales before any accuracy evaluation was published, and sitting at the head of a pipeline whose downstream node scores reoffending risk from the records it helps write.

The largest published government copilot evaluation anywhere — about 20,000 UK civil servants across 12 organisations — found a self-reported 26 minutes saved per working day and most users unwilling to go back, while a companion departmental evaluation found no robust evidence the time savings improved productivity and about a fifth of its users caught the tool hallucinating.

A language-model scanner reads roughly 25,000 paper letters to the UK welfare agency every day and decides which claimants surface on the potentially-vulnerable shortlist. Its transparency record names precision, recall and F1-score as its metrics but discloses no values, and the claimants were never told it exists — the impact assessment said letter writers do not need to know.

The Home Office piloted and then rolled out two AI copilots for asylum casework — one that compresses the interview transcript into a summary and one that summarises country policy. In the Home Office's own pilot the summariser saved a measured 23 minutes a case, 9% of its summaries were removed as inaccurate before any caseworker saw them, and 23% of users were not fully confident in the rest; the summaries carried no source references, no requirement to verify them exists, and claimants are never told AI touched their case.

The largest trial court in the United States is piloting an AI clerk that drafts rulings in each judge's own writing style, under a contract with a roadmap into criminal, family, and probate cases — and because California's disclosure rule is triggered only by documents written entirely by AI, the litigants whose cases it touches are never told.

A federal disability-adjudication copilot that runs the other way — it verifies human-drafted decisions against 43 automated quality flags instead of drafting them — and the atlas's clearest example of a governed, precision-gated checker, along with the one failure mode a checker has.

Rules-based automation grants veterans' survivor-benefit claims end to end with no human when the rules match — and a federal watchdog on a multi-year clock was the only correction channel that ever changed anything.

The first Swedish municipality to fully automate social-assistance decisions — a rules-based robot that decides about one in three recurring reapplications with no human at all, and re-inserts a caseworker only when a single routing rule fires — and the atlas's clearest example of automation replacing discretion by design rather than overriding it in practice.

The fairness-first welfare-fraud screener a city built with nearly every recommended safeguard — bias audits, debiasing, human-rights and data-protection assessments, a citizen panel, public transparency registers — and then discontinued itself after a live pilot showed the bias had re-emerged inverted, now wrongly flagging the majority group, women, and parents, with no gain over caseworkers. The atlas's governed-exit case.