Shipping Production AI Inside Government — William Tarr, Ministry of Justice (DO NOT PUBLISH)
The UK Ministry of Justice Justice AI Unit operates as a startup within government, deploying engineers directly into prisons and probation offices. This 'forward deployed' model enables rapid, user-centric AI product development, overcoming bureaucratic hurdles by prioritizing real-world feedback and agile implementation to empower frontline staff and solve critical operational challenges.
The UK Ministry of Justice’s (MoJ) Justice AI Unit, a lean “startup within government” of 40 people achieving the output of 300, is fundamentally changing how technology is developed and deployed in the public sector. William Tarr, a “forward deployed engineer,” highlights their mission to ship AI products that solve real-world problems, contrasting this with traditional government approaches characterized by “pilots and slides.”
The Problem with Traditional Government Technology
The existing landscape of government technology is plagued by several issues:
- Disjointed Systems: Different identification numbers across prisons, courts, and probation lead to inefficiencies.
- Poor Organization: Lack of communication and understanding of end-to-end processes among personnel.
- Bad Contracts & Legacy Tech: Layers of historical tech procurement result in “dodgy connections” and “weird APIs” with convoluted governance.
- Centralized Detachment: Technology built in a centralized manner often lacks connection to frontline users (prison officers, probation officers), leading to ineffective solutions.
These issues contribute to critical failures, such as the erroneous release of prisoners, which served as a stark introduction to the unit’s challenges.
The “Forward Deployed” Model: A Solution
The Justice AI Unit’s differentiating factor is its “forward deployed model.” Engineers spend two to three days a week directly embedded with users in prisons, probation offices, and courts. This approach involves:
- Direct Observation: Sitting with officers, observing workflows, and understanding unspoken processes not captured in policy documents. William Tarr recounts holding prison keys and walking wings to understand operations firsthand.
- Rapid MVP Development: Building Minimum Viable Products (MVPs) in “days” or “weeks,” testing them on-site, and showing users “what we can do with AI.”
- Problem-First Approach: Louis Hogarth, another AI product builder, emphasizes “falling in love with the problem, not the product.” Their probation officer tool, for instance, went from MVP to national rollout in months with just two engineers, aiming to ease the pain points of 12,000 probation officers.
- Diverse Talent: The team comprises individuals from various backgrounds (startup, Palantir, YC) and includes roles like a professor of sociology to consider the long-term future of justice, fostering a “forward looking” mindset.
This direct engagement helps them build solutions that align with how frontline staff actually operate, rather than idealized versions of workflows.
Navigating Government Challenges
Operating as a “startup in government” comes with inherent friction:
- Outdated Governance: Decision-making processes are designed for policy, not agile product development. Committees composed of individuals with no frontline experience often dictate product needs.
- HQ Hesitancy and Mistrust: A lack of technical understanding and embedded mistrust of new technologies within headquarters.
- Bridging the Gap with “Signal”: Instead of theoretical discussions, the unit brings tangible “signal” – working prototypes and real user feedback – to governance meetings. This “safe to fail piloting” approach allows them to navigate bureaucracy with “ruthless efficiency.”
- Safety and Security: Despite the agile approach, they prioritize data security and ethical considerations from day one, collaborating with unions, ethics colleagues, and policy teams to ensure responsible deployment. Strong platform engineering foundations are crucial for addressing MoJ concerns.
Future Applications of AI in Justice
The unit envisions AI empowering frontline colleagues by automating tedious tasks, allowing them to focus on core duties:
- Transcription and Paperwork Automation: “Low-hanging fruit” like transcription saves many hours, allowing officers to engage more with prisoners and probationers.
- CCTV Analytics and Predictive Intelligence: Moving from reactive to proactive use of CCTV in prisons.
- Assistance Agents: To streamline outdated and messy paperwork.
- Better Data Integration: Addressing issues like the “releasing error problem.”
- AI Education and Learning: Empowering colleagues to effectively use AI tools.
Addressing Foundational Infrastructure and User Education
During the Q&A, challenges related to physical infrastructure and user education were discussed:
- Legacy Infrastructure: Many prisons lack Wi-Fi due to “Victorian buildings with kind of 6-7 ft walls” or security bans. The unit’s solution involves developing offline modes and adapting solutions for on-premise contexts, leveraging direct user feedback to identify and circumvent these limitations.
- Hardware Constraints: Acknowledged, especially with CCTV, where outdated contracts for hardware are a long-term problem. The unit currently prioritizes software/AI “low-hanging fruit” for immediate impact while recognizing the need for longer-term solutions.
- User Education and Adoption: They simplify language, avoid jargon, use visual aids (videos), and provide on-site support to ensure staff understand and can effectively use new tools, effectively acting as “IT support” when needed.
The Justice AI Unit is actively seeking “forward deployed engineers” to join their mission, emphasizing the tangible impact of their user-centric, agile development model within a large government organization.