The UK Ministry of Justice established the Justice AI Unit to address critical inefficiencies in the prison, probation, and court systems that were causing operational failures including erroneous prisoner releases. The unit adopted a forward-deployed engineering model where a lean team of approximately 40 engineers work directly on-site with prison officers, probation officers, and court staff to rapidly build and deploy AI-powered tools. By spending 2-3 days per week embedded with frontline staff, the team ships production-ready AI products in days to weeks rather than the typical years-long government procurement cycles. Key solutions include transcription services, CCTV analytics, automated assistance agents, and offline-capable tools designed to work within the constraints of legacy infrastructure, poor connectivity, and fragmented systems across England and Wales.
The UK Ministry of Justice’s Justice AI Unit represents a distinctive approach to deploying AI and LLM-based systems in production government contexts. Founded approximately 18 months prior to this presentation by director Dan James, the unit has scaled from a single person to roughly 40 team members, operating with output that would typically require 300 people in traditional government structures. The unit operates within the Ministry of Justice, which employs 95,000 people including approximately 30,000 operational staff such as prison officers and probation officers.
The catalyst for the unit’s creation was a crisis situation where prisoners were being erroneously released due to systemic failures across disjointed systems. Different identification numbers existed for prisoners across courts, probation services, and prisons themselves, combined with poor organizational communication, problematic legacy contracts, fragmented APIs, and outdated governance structures. The Justice AI Unit was created explicitly to address these failures through rapid AI deployment rather than traditional lengthy government procurement and development processes.
The core operational philosophy of the Justice AI Unit centers on forward-deployed engineering, a model borrowed from private sector companies like Palantir. Engineers spend significant time physically embedded in prisons, probation offices, and courts rather than working remotely or from centralized offices. Engineers receive security clearances and keys to access prison facilities freely, allowing them to observe workflows firsthand, sit with officers during their daily work, and understand pain points that would never surface in requirements documents or stakeholder meetings.
New engineers join prison sites as early as their second day of employment, with one engineer reporting being in a probation office on day two and inside a prison by day four. This immersion approach fundamentally changes the development process from theoretical requirements gathering to observational problem identification. The team explicitly rejects traditional roadmap-driven quarterly planning in favor of identifying immediate pain points through direct observation.
The forward-deployed model is particularly effective in the justice context because the documented procedures and policy documents describing how prisons and probation services should operate often diverge significantly from actual operational reality. Officers develop workarounds and pragmatic solutions to get their jobs done that bear little resemblance to idealized policy descriptions. By observing actual work rather than documented work, the engineering team builds tools that integrate with real workflows rather than theoretical ones.
The Justice AI Unit operates on timescales measured in days and weeks rather than months or years. The team builds MVPs in days, conducts on-site demonstrations with officers, gathers immediate feedback, and iterates rapidly. This represents a dramatic departure from typical government technology projects that often take years from conception to deployment.
The team uses feature flagging extensively to enable safe experimentation in production. New capabilities are built with users on-site, feature-flagged for controlled testing, and either scaled up based on positive feedback or quickly deprecated if they don’t provide value. This “fail fast” philosophy is explicitly embraced, with the team willing to kill code that doesn’t work rather than committing to predetermined deliverables.
One probation-focused product progressed from MVP to pilot to national rollout across all probation offices in England and Wales within a few months, built primarily by just two engineers. This contrasts sharply with traditional government digital transformation projects that typically involve teams of 40+ people in various specialized roles and take years to reach national deployment.
The team maintains constant communication channels with frontline users. Probation officers and prison staff directly message engineers throughout the day reporting issues, requesting features, or asking for help. Engineers provide immediate support, sometimes functioning as informal IT support while also gathering product feedback. This direct connection eliminates layers of intermediaries typical in government technology projects.
While the presentation doesn’t provide exhaustive technical implementation details, several specific AI and LLM application areas are highlighted as production deployments:
Transcription Services: Described as “low-hanging fruit” for government AI deployment, transcription capabilities are being deployed to reduce administrative burden on probation officers and prison staff. The goal is to free officers from desk work and paperwork to spend more time on their core responsibilities of working face-to-face with prisoners and people on probation. This likely involves speech-to-text models processing interviews, meetings, and case notes.
CCTV Analytics and Predictive Intelligence: One engineer named David is specifically working on computer vision applications for analyzing CCTV footage in prisons. Currently CCTV serves primarily reactive purposes, reviewed after incidents occur. The AI unit is developing analytics tools to make CCTV analysis more proactive and intelligence-driven, potentially identifying patterns or concerning behaviors before incidents escalate.
Assistance Agents: The team is deploying AI agents to replace or augment paperwork processes that are described as “outdated and messy.” These agents likely use LLMs to help officers complete administrative tasks, fill out forms, generate reports, or navigate complex case management systems.
Data Integration: AI is being applied to the prisoner release crisis that motivated the unit’s creation. The challenge involves reconciling data across fragmented systems where prisoners have different identifiers in different databases and information doesn’t flow cleanly between courts, prisons, and probation services.
Education and Learning: The team is developing tools to help staff across the Ministry of Justice understand and effectively use AI capabilities, moving beyond what they describe as “boring AI seminars” to practical hands-on education.
The Justice AI Unit faces significant infrastructure constraints that heavily influence their LLMOps approach. Many prison facilities lack Wi-Fi entirely, either due to security policies preventing prisoner access to wireless networks or because Victorian-era buildings with thick stone walls don’t support wireless signal propagation. Other facilities have inconsistent connectivity that one engineer colorfully described as requiring “the Lion King impression waiting for the Wi-Fi to connect.”
These constraints drive specific technical solutions. The team builds offline-first capabilities into their applications, using browser caching and local storage to enable functionality without consistent network connectivity. Applications synchronize data when connectivity becomes available rather than requiring constant connection. This offline-first architecture is essential for production viability but represents additional engineering complexity.
The technology landscape across the Ministry of Justice is highly fragmented with inconsistent infrastructure across England and Wales. Some facilities have modern connectivity while others operate with severely limited capabilities. Rather than waiting for infrastructure improvements that might take years, the team adapts applications to work within existing constraints, treating infrastructure limitations as fixed parameters to design around rather than blockers.
Legacy systems and contracts create additional complexity. CCTV systems in particular involve multiple software layers from different vendors serving the same cameras, negotiated through poorly structured contracts that the AI unit doesn’t control. Hardware procurement happens through prison-level or regional bodies rather than centrally, creating heterogeneity that complicates standardized solutions. The team acknowledges these hardware and contractual issues as longer-term challenges while focusing initially on software-layer improvements that can be deployed despite underlying infrastructure problems.
Despite the rapid development pace, the Justice AI Unit maintains rigorous attention to security, data protection, and governance requirements. The team includes platform engineers whose role is explicitly to “bat away any questions or concerns from MOJ” headquarters about security and data handling. All technology is built on what the team describes as “really strong foundations” with data security “up to scratch.”
The team proactively engages with multiple stakeholder groups including unions, ethics committees, and policy teams from the beginning of product development. Rather than treating governance as a gate at the end of development, they integrate governance considerations throughout the process. This approach is described as maintaining “basic best practice” around data handling and processing as core requirements rather than afterthoughts.
The security posture is particularly critical given the sensitive nature of justice system data including prisoner information, case details, and CCTV footage. The team cannot adopt a purely “move fast and break things” mentality but must balance speed with safety. They describe this as recognizing “you can’t be too scrappy in the context of doing something dangerous.”
A significant portion of the presentation addresses challenges in navigating government governance structures that were designed for policy development rather than product development. Traditional governance processes involve lengthy committee reviews by people who may have never visited a prison but make decisions about tools intended for prison officers. These processes can take months and involve stakeholders far removed from the operational context where products will be deployed.
The Justice AI Unit’s strategy for navigating bureaucracy involves demonstrating working products rather than discussing theoretical proposals. Instead of requesting approval for concepts, they build prototypes rapidly, test them with real users, gather observable behavioral data and feedback, then present concrete results to governance bodies. This approach provides “clear signal” rather than speculation, making governance discussions more productive and accelerating approvals.
The team also explicitly trusts frontline staff to make good decisions about tool usage. Their philosophy holds that if prison officers are trusted to operate in high-stakes prison environments, they should be trusted to decide which tools best serve their needs. This user-trust philosophy pushes back against governance approaches that treat frontline staff as needing extensive protection from technology choices.
Resistance and skepticism within Ministry of Justice headquarters represents an ongoing challenge. Many staff members have added “AI” or “innovation” to job titles despite limited hands-on experience with AI technologies. There exists an embedded mistrust of new technology and a significant knowledge gap between headquarters staff and operational realities. The forward-deployed model explicitly aims to bridge this gap by having engineers who understand both the technology and the operational context serve as informed intermediaries.
The Justice AI Unit deliberately recruits from diverse backgrounds outside traditional government channels. Team members come from Y Combinator startups, private sector companies like Palantir, and various technology firms, bringing startup mindsets and private sector velocity to government. The team includes a professor of sociology focused on long-term futures thinking about what justice systems might look like in 10-20 years, balancing rapid execution with strategic vision.
The team operates with roughly 40 people total, predominantly engineers, which is described as a 40:300 ratio compared to the headcount traditional government projects would require for similar output. The unit is actively hiring additional forward-deployed engineers to expand capacity.
Engineers are encouraged to “fall in love with the problem, not the product,” prioritizing impact over specific technical implementations. When infrastructure constraints make planned solutions infeasible, the team rapidly pivots to alternative approaches rather than treating obstacles as blockers. This pragmatic problem-solving culture emphasizes deployment and user value over technical elegance or predetermined deliverables.
The culture emphasizes visible presence and accessibility. Engineers are known by name to hundreds of frontline users who contact them directly with issues and requests. This represents a radical departure from typical government technology where development teams are invisible to end users and communication flows through multiple organizational layers.
The Justice AI Unit’s approach to user experience and communication is notably pragmatic. All language used with frontline staff avoids technical jargon. Engineers wouldn’t explain “caching locally” but instead would simply say a feature will “work offline.” This translation of technical implementation into user-meaningful language ensures staff understand capabilities without needing technical knowledge.
The team heavily uses video demonstrations and visual communication rather than relying on written documentation or email announcements. When users struggle to understand features, engineers jump on quick calls or provide on-site demonstrations. This multimedia, synchronous communication approach accommodates varying technical literacy levels among users.
Product design philosophy emphasizes making interfaces “two times simpler to make it 10 times better.” The team recognizes that frontline staff need to focus on their core mission of working with prisoners and people on probation, not on learning complex software interfaces. Simplicity is treated as a multiplier of effectiveness rather than a compromise.
Engineers sometimes find themselves providing basic IT support, helping users with phone connectivity or VPN configuration. Rather than viewing this as beneath their role, the team sees it as essential relationship-building and trust-establishment that enables better product development through deeper user understanding.
While the Justice AI Unit demonstrates several mature LLMOps practices including rapid iteration, feature flagging, user feedback integration, and production deployment at national scale, the presentation reveals limited information about specific LLM technical architectures, model selection processes, evaluation frameworks, prompt engineering approaches, or monitoring systems.
The focus remains primarily on organizational model and deployment velocity rather than detailed technical implementation. This may reflect the early-stage nature of some deployments, the sensitive nature of government systems precluding detailed public disclosure, or simply the presentation’s focus on the forward-deployed model rather than technical architecture.
The absence of discussion around model fine-tuning, RAG architectures, embedding strategies, or specific LLM providers suggests either that implementations use relatively straightforward applications of existing models and APIs, or that technical details are intentionally not disclosed publicly. The emphasis on “low-hanging fruit” for transcription applications implies initial deployments may focus on well-established speech-to-text capabilities rather than cutting-edge LLM applications.
The Justice AI Unit positions itself as forward-looking, explicitly thinking about what justice systems will look like in 10-20 years. The vision centers on empowering operational staff to focus on their core mission rather than administrative burden. Prison officers should be on wings with prisoners, not at desks. Probation officers should provide face-to-face support to people trying to avoid reoffending, not complete paperwork.
This human-centered automation philosophy views AI as enabling human expertise rather than replacing it. The goal isn’t to reduce headcount but to redirect human attention to high-value interpersonal work that requires judgment, empathy, and expertise while automating routine administrative tasks.
The team’s existence and expansion suggests institutional support within the Ministry of Justice despite bureaucratic resistance. Growing from one person to 40 in 18 months, achieving national rollouts, and actively hiring indicates the model has demonstrated sufficient value to secure continued investment and expansion.
While the presentation effectively communicates the forward-deployed model and rapid deployment achievements, several important caveats warrant consideration. The presentation lacks quantitative metrics around actual impact. There are no statistics on time saved, error rates reduced, prisoner outcome improvements, or officer satisfaction scores. The focus remains on deployment velocity and user engagement rather than measured outcomes.
The sustainability of the high-touch forward-deployed model at larger scale remains unclear. With 95,000 Ministry of Justice employees and 40 engineers, the team can only directly support a small fraction of potential users. Scaling to broader coverage may require transitioning from direct engineer-user relationships to more traditional support structures, potentially losing the velocity advantages the model provides.
The tension between rapid iteration and government governance requirements may not be fully resolved despite the team’s success navigating bureaucracy thus far. As AI applications move from supporting administrative tasks to potentially influencing sentencing, release decisions, or risk assessments, governance scrutiny will intensify. The presentation acknowledges ongoing friction but doesn’t detail how governance scales as applications become more consequential.
Technical debt accumulated through rapid development may create future challenges. Building offline-capable applications with browser caching as solutions to infrastructure gaps may create maintenance burden and limit functionality compared to solutions built on reliable connectivity. The long-term architectural implications of these pragmatic workarounds remain to be seen.
Finally, the presentation comes from team members actively recruiting, which may introduce optimistic bias in how challenges and successes are framed. Independent evaluation of outcomes, user satisfaction, and comparative performance against traditional government delivery approaches would provide valuable context currently absent from this self-reported account.
Snorkel developed a specialized benchmark dataset for evaluating AI agents in insurance underwriting, leveraging their expert network of Chartered Property and Casualty Underwriters (CPCUs). The benchmark simulates an AI copilot that assists junior underwriters by reasoning over proprietary knowledge, using multiple tools including databases and underwriting guidelines, and engaging in multi-turn conversations. The evaluation revealed significant performance variations across frontier models (single digits to ~80% accuracy), with notable error modes including tool use failures (36% of conversations) and hallucinations from pretrained domain knowledge, particularly from OpenAI models which hallucinated non-existent insurance products 15-45% of the time.
Digits, a company providing automated accounting services for startups and small businesses, implemented production-scale LLM agents to handle complex workflows including vendor hydration, client onboarding, and natural language queries about financial books. The company evolved from a simple 200-line agent implementation to a sophisticated production system incorporating LLM proxies, memory services, guardrails, observability tooling (Phoenix from Arize), and API-based tool integration using Kotlin and Golang backends. Their agents achieve a 96% acceptance rate on classification tasks with only 3% requiring human review, handling approximately 90% of requests asynchronously and 10% synchronously through a chat interface.
Predibase, a fine-tuning and model serving platform, announced its acquisition by Rubrik, a data security and governance company, with the goal of combining Predibase's generative AI capabilities with Rubrik's secure data infrastructure. The integration aims to address the critical challenge that over 50% of AI pilots never reach production due to issues with security, model quality, latency, and cost. By combining Predibase's post-training and inference capabilities with Rubrik's data security posture management, the merged platform seeks to provide an end-to-end solution that enables enterprises to deploy generative AI applications securely and efficiently at scale.
