UK i.AI Redbox — Open-Source AI Document Analysis for Civil Servants

Redbox is a retrieval-augmented generation (RAG) chatbot built by the UK Government's Incubator for Artificial Intelligence (i.AI), a unit within the Cabinet Office. It let civil servants upload letters, briefings, minutes, speech transcripts and policy documents and query them in natural language, receiving summarised answers with direct citations to source documents. The tool was piloted from late 2023 and takes its name from the red briefcases used to carry ministerial papers.

Redbox aimed to help civil servants interrogate large volumes of policy documents more efficiently through a generative AI assistant, while keeping outputs verifiable through direct citation links to source material. This citation feature was explicitly designed to make hallucinations visible and checkable by users, addressing a key trust concern with generative AI in government use.

Redbox was deployed to the Cabinet Office, 10 Downing Street and the Department for Science, Innovation and Technology (DSIT), with algorithmic transparency records subsequently published on GOV.UK for the DSIT and Department for Business and Trade (DBT) deployments. The full codebase was released as open source on GitHub. By February 2025 the platform had 2,000 active civil-service users, growing by approximately 150 per week, and in that month alone it processed 7,000 documents including grant evaluations, invoices and policy papers.

Over its full lifecycle Redbox served more than 6,000 civil servants, supporting 150,000 chats and processing 1.3 million messages. A user survey (n=282) found that 80% felt Redbox improved the quality of their work and 89% said it saved them time, with a median estimated saving of 2 hours per week; the resulting Net Promoter Score of 56 was above the 'excellent' threshold of 50.

In 2025, i.AI announced Redbox's discontinuation after the Cabinet Office consolidated onto commercial enterprise AI services (Microsoft Copilot and Google Gemini) offered across government. i.AI published a public retrospective identifying five lessons from building and sunsetting the tool, including the need to align sovereign AI tooling with departmental strategies before scaling — a rare candid post-mortem of a government AI project.