evidoria

← Back to browse

Good practice Imported

Redbox — the UK Civil Service's Generative-AI Assistant, Built, Scaled and Then Sunset

United Kingdom · London · See the United Kingdom profile · See the London profile

Evidence: Observational / pre–post Top 74% 47/100 · Ask Evidence Copilot about this practice

The UK Cabinet Office's Incubator for AI scaled Redbox, a bespoke generative-AI chatbot for civil servants, to 6,000+ users and 1.3 million messages by 2025 — then sunset it, concluding that Microsoft Copilot delivered similar value once licensed government-wide.

6,000+
Registered users (by 2025)
~150,000
Chats (by 2025)
1.3 million
Messages (by 2025)
Redbox — the UK Civil Service's Generative-AI Assistant, Built, Scaled and Then Sunset

Details

Maturity
Discontinued
Promoter
Incubator for AI (i.AI), Cabinet Office
Period
2023–2025
Keywords
digital government, public sector productivity, generative AI, internal operations, programme evaluation

Context

The Incubator for AI (i.AI), part of the UK Cabinet Office and later the Department for Science, Innovation and Technology, began building Redbox from 2023 as a secure, in-house generative-AI tool to help civil servants summarise policy documents, prepare briefings and search legal and parliamentary material.

Objectives

Redbox became part of the wider "Humphrey" suite of government AI tools unveiled in 2025 alongside Consult, Parlex, Minute and Lex, and its source code was published openly on GitHub so other teams could inspect or reuse the approach.

Activities

i.AI built and scaled Redbox across the civil service, tracking usage as adoption grew.

Results

Adoption reached over 6,000 users, spanning roughly 150,000 chats and 1.3 million messages, according to i.AI's own figures. By 2025–2026, with Microsoft Copilot and Google Gemini licensed across government departments, i.AI's test-and-learn evaluation found Copilot delivered comparable functionality more cheaply for users already inside the Microsoft ecosystem.

Conclusions

i.AI judged that further investment in Redbox was no longer the best use of resources, kept it running only long enough for the Cabinet Office to migrate onto Gemini, and then formally retired it — publishing a public "lessons learned" retrospective rather than quietly discontinuing the tool. It stands as a documented cautionary case: real adoption did not, on its own, justify continuing a bespoke build once comparable commercial tools became available at scale.

Implementation
Indicative cost
Medium (€50k–€500k) — Not quantified in sources reviewed beyond the qualitative finding that Copilot was cheaper for users already inside the Microsoft ecosystem.
Time to results
Short (< 1 year) — Built from 2023; adoption scaled through 2025; retired/sunset by 2025–2026 following migration to Gemini.
Staffing & skills
Incubator for AI (i.AI), UK Cabinet Office / Department for Science, Innovation and Technology

Conditions for success

  • Open-sourcing the codebase on GitHub for reuse by other teams
  • A formal test-and-learn evaluation comparing the in-house build against commercial alternatives (Microsoft Copilot)
  • Willingness to retire the tool once comparable commercial tools matched functionality at lower cost, and to publish a candid retrospective explaining why

Common failure modes

  • Once Microsoft Copilot and Google Gemini were licensed government-wide, the bespoke build no longer offered enough distinct value to justify continued investment, despite genuine adoption (6,000+ users)

Where it fits

Governance type
central government innovation unit (Cabinet Office / DSIT)
Scale
UK civil service, cross-department

Commonly funded by

National / regional programmes

Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.

Do you run this practice? Claim it — verified implementers get a public contact pathway and can propose corrections.

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful