Redbox Reflections: 5 Key Lessons from Building (and Sunsetting) Our Government AI Chatbot

Innovation in government is a journey of bold steps, continuous learning, and sometimes, strategic goodbyes. Our Redbox project, an AI-powered chat and document analysis tool, recently reached the end of one such journey. From a hackathon concept to a tool used by over 6,000 civil servants, Redbox taught us a phenomenal amount. Today, we're distilling that experience into five key lessons that we believe are crucial for anyone working on AI in the public sector, exemplifying the government's commitment to a 'test and learn' culture. 

Lesson 1: Solve a Real, Pressing Need (But Verify It’s the Right Need)

Redbox started as a hackathon project with a clear goal: help ministerial private offices manage vast amounts of information using LLMs. This was a genuine pain point. We built a RAG system, then added functionality for full document analysis, citation features for trust, and connections to data sources like gov.uk. These were all logical extensions.

The Insight:

However, usage data delivered a surprise. Around 70% of interactions were users simply wanting to "talk" to an LLM for general tasks – drafting emails, brainstorming, etc. – because Redbox was their only secure work-based option. While we were building for complex document analysis, the most pressing, unmet need was often basic, secure LLM access. 

Takeaway:

Continuously validate that the problem you're solving is the most critical one for your users. Sometimes the simplest version of your product meets the biggest need.

Lesson 2: The Power of "Good Enough" and Rapid Scaling

Our initial focus on sophisticated features for a niche audience (private offices) delayed wider rollout. When we realised the overwhelming demand for basic LLM chat, we pivoted. We stripped back Redbox to its core chat and document interaction capabilities, prioritised a clean user experience, and focused on scaling.

The Insight:

This shift led to explosive growth, with Redbox quickly becoming the primary LLM access point for the Cabinet Office. In one month we doubled our unique monthly users from 1,000 to 2,000 and suddenly 30,000 messages were being exchanged every week! 

Takeaway:

Don't let the perfect be the enemy of the good, especially when there's a clear, widespread need. Sometimes, getting a core, valuable capability to the most people quickly is the highest impact strategy.

Lesson 3: The AI Market Moves at Warp Speed – Stay Agile

Redbox was developed when secure, in-house LLM solutions were scarce. Our Azure-deployed models offered a significant privacy advantage. However, the AI landscape is incredibly dynamic. The rise of enterprise solutions like Microsoft Copilot, especially once copilot chat became freely available to many government departments, fundamentally altered Redbox's value proposition.

The Insight:

The test and learn programme in DSIT showed Copilot offered similar functionality. What was once a unique offering from Redbox became one of several options, and often not the most integrated one for users already in the Microsoft ecosystem. 

Takeaway:

Constantly scan the horizon. What's unique today might be commonplace tomorrow. Be prepared for your product's strategic fit to change and adapt your strategy accordingly. This includes being ready to sunset a product.

Lesson 4: User Trust is Non-Negotiable, Especially in Government

From early user research, it was clear that for civil servants to use an LLM for their work, especially with sensitive documents, trust was paramount. They needed to be confident that responses were grounded in the provided information and that their data was secure.

The Insight:

This led us to prioritise experimenting with features like robust citation capabilities, linking AI-generated statements directly back to source documents. Our secure model deployments, and the ability for the user to select the model they wanted to talk with was also a cornerstone of this trust. 

Takeaway:

For any AI tool in a sensitive environment like Government, building and maintaining user trust through transparency, reliability, and security is key.

Lesson 5: Know When to Declare Victory (and Move On)

Redbox achieved significant impact. It filled a critical gap for LLM access, supported thousands of users, and provided us with invaluable experience. The code has also been forked and built on by others to create successful products. However, with the changing market and the availability of tools like Copilot, investing further in a broad rollout of Redbox no longer represented the best use of resources. The decision was made for a controlled shutdown. We are going to keep Redbox running until the end of the year to allow time for the Cabinet Office to transition onto their alternative enterprise LLM option - gemini. After that it will be shut down, enabling us to prioritise higher impact work.

The Insight:

This wasn't a failure, but a recognition that Redbox had served its primary purpose and the strategic landscape had evolved. The open-source codebase remains an asset, and the lessons learned are fuelling future projects. 

Takeaway:

Success isn't just about perpetual growth. It's also about achieving a mission and knowing when to strategically pivot or conclude a project, freeing up resources for the next challenge. This is a vital capability in the fast-paced world of AI.

Conclusion

The Redbox journey took many turns. We're proud of what we built and the value it delivered. More importantly, we're taking these lessons forward as we continue to explore how AI can best serve the public good.