Few parts of the public sector are under more pressure – or more scrutiny – than policing, justice and security. These services meet people at some of their most vulnerable moments, handle some of the state’s most sensitive information, and deliver some of the most fundamental roles that a government performs for its citizens. They are regularly the target of cyber threats due to the high‑value data they hold, and operational challenges are clear when you look at the front line: court backlogs stretching into years, probation caseloads rising without matching increases in staff, and prisons running close to capacity. This environment makes artificial intelligence (AI) a highly attractive tool to improve efficiency and outcomes, but reaping the benefits of AI capability presents specific challenges and risks.
Generative AI tools can process unstructured information at scale, from body‑worn video footage and interview recordings to sprawling case files and handwritten notes, handling material that older systems struggled with. The government has recognised this potential through initiatives such as the AI Opportunities Action Plan, but adoption is harder than almost anywhere else in government. Every deployment must comply with strict due‑process requirements, preserve clear records of how decisions were reached, and demonstrate to a sceptical public that the benefits are fair to everyone.
Senior officials from across the sector came together at a recent CSW roundtable, held in partnership with NTT DATA and under the Chatham House Rule, to explore both the potential and constraints. Below are the key takeaways.
Lead with purpose, not technology
Participants cautioned against starting from the technology, noting the temptation to find uses for tools that are now available rather than to ask what the system needs to do better. The starting point, they argued, should be the underlying policy objective, expressed in terms a frontline officer would recognise: reducing reoffending, resolving cases more quickly, improving support for victims, or managing higher caseloads safely in the community. As one attendee put it, the question is not “what can AI do for us?” but “what are we trying to achieve, and could AI help?” That clarity also makes it possible to measure whether you are getting there. Participants also highlighted the need for more systematic, sector‑wide diagnostics on where the real bottlenecks are, with clear lines of accountability for acting on what they reveal.
The same focus also helps unlock ideas from the people closest to the problems, and the Civil Service AI & Data Challenge, supported by NTT DATA, is a testament to the real appetite for change. Ideas emerging through that programme include using AI to kick-start case investigations, cut unnecessary admin, and help officials make sense of much larger evidence sets. In HMRC and DWP, for example, AI is already being explored as a way to suggest where to start on a case and where the key leads are, while leaving decisions with the people best placed to make them.
Fix the foundations before scaling the tools
Data is fundamental to any AI‑enabled justice and policing system, and here participants saw a structural problem that technology alone cannot overcome: the underlying digital infrastructure is fragmented. In youth justice, for example, local services use three main case management systems with different data schemas and upgrade cycles, making it hard to get a consistent national picture. Police forces often procure their own systems, which may not communicate well even with neighbouring forces. And legacy infrastructure is hard to retire, so organisations end up running green-screen systems long past their useful life along with paper‑heavy processes, which in turn constrains what AI can plug into.
These issues are amplified by the nature of justice as a system. The flow of information across police, prosecutors, courts, prisons and probation means an improvement in one place can displace pressure to another. Faster police case preparation, for instance, can increase workloads at the Crown Prosecution Service and in courts. Reducing the courts backlog will, in time, increase the prison population unless community capacity keeps pace. As participants noted, squeeze the balloon and it simply pops up somewhere else.
Participants repeatedly framed these obstacles as systemic rather than purely technical, and the AI & Data Challenge shows it, with roughly half of submissions involving cross‑government data sharing between organisations. They also recognised the importance of getting this right. AI is a force multiplier of both good and weak practices, so if it is layered on top of poor data, fragmented systems and manual workarounds, it will only accelerate those flaws.
Govern the tools, not just the data
The market for AI tools feels, in the words of one attendee, like “a bit of a wild west”, with multiple commercial models, such as ChatGPT, Gemini and Claude, making it increasingly hard for a civil servant to know which to use or how to compare them. Better governance should provide that clarity, but participants noted that existing frameworks were not designed with AI in mind, and at least one justice organisation represented at the table had to build new AI governance from scratch – defining who can deploy which models, on what data, and with what level of sign‑off. But guidance does exist – notably the government’s AI Playbook, which several participants cited. Awareness, it seems, is uneven and real‑world experience in applying it to live deployments is still emerging.
Participants disagreed about whether AI is inherently agile. While vendors can release new models rapidly, once a particular model is embedded in business processes – for instance, as the engine behind a transcription service or case‑summarisation tool – replacing it can be costly and disruptive. Government already has a long track record of vendor lock‑in for databases and line‑of‑business systems, so the same risks apply to AI platforms if interfaces and contracts are not designed carefully.
Invest in people as much as in platforms
Driving change is never just about technology, so the discussion turned to capabilities and the culture needed to adopt AI successfully, with several participants suggesting that building AI literacy is the right starting point. Addressing the skills gap, they argued, would help people understand how to use AI well in their own roles, from writing effective prompts to get a usable first draft, to judging when an output can be trusted against policy and case law, and knowing when to escalate to human review. This is not about everyone becoming a data scientist, but with the right guardrails built into tools and simple interfaces embedded in familiar systems – such as office productivity suites or case management tools – most staff could self-regulate their use of AI effectively.
Some participants suggested that the design of the tool matters as much as the training around it, but one concern emerged about what they called “the AI slop”. Generative AI makes it easy to produce large volumes of text quickly, and several officials have already seen 20-page AI-generated letters from the public, which are instantly recognisable as such but still have to be read and checked against case references. Internally, the same risk applies if staff equate value with volume. Helping people understand what good looks like – concise, accurate, purposeful use of AI – and that value is not measured in pages, is as important as any technical training.
Scale with care, not speed
Several participants highlighted the difference between redesigning processes around AI and retrofitting AI onto existing ones. The latter often creates friction and can struggle to deliver a return on investment unless the process itself is redesigned.
Justice Transcribe was held up as an example of what good looks like. Earlier transcription systems, trained on ideal studio conditions such as TED‑style talks, often fell apart in police interview rooms, especially when faced with strong regional accents and background noise. By contrast, modern AI transcription has proved far more useful in those real‑world conditions, opening up new use cases for probation interviews, body‑worn video and other recorded evidence that previously took hours of manual review.
The key is to start small and build trust through visible wins, like some of the most promising applications discussed at the event. There’s HMRC’s voice‑authentication to reduce friction on helplines; personal‑productivity tools that help officials draft clearer letters faster; and back‑office uses such as expenses and report‑writing. Participants agreed that one or two lighthouse use cases that staff, leaders and the public can point to could shift attitudes more effectively than a large portfolio of inconclusive experiments.
A second, underused route to scale is adopting tools already proven elsewhere. AI‑assisted legal services are already operating in civil and immigration law, where supervised tools prepare standard case documents and correspondence for a relatively low fee, with a qualified solicitor ultimately responsible. Participants suggested that taking such approaches into criminal and administrative justice could reduce risk for senior decision‑makers and accelerate the move from pilot to deployment. They also recognised that scaling demands a whole‑system view, so when a pilot speeds up one part of the justice chain it must also be assessed for its effects on everything downstream.
Build for accountability and design for everyone
Participants were clear that AI in this sector should inform and assist human decision‑makers, and that it’s not there to replace them. The question is how to design tools so that accountability remains with people. This design challenge is not a reason to avoid AI, but it means building tools with transparency about what data they draw on and how outputs are generated; rigorous governance of the data and models they rely on; and clear processes for staff, victims and defendants to challenge and review AI‑assisted decisions.
They also raised the risk of an AI monoculture, drawing on recent research that heavy reliance on the same models can narrow the range of ideas. Applied to justice, the concern is that if everyone across the sector is using identical tools, there is a danger of convergence in thinking precisely where independent judgement and diversity of thought are essential. Participants also noted that AI can reflect and scale existing biases in its training data, so using it to surface patterns of disproportionality, and benchmarking outcomes across forces, regions and demographic groups, would be critical.
There is also the question of who benefits from this transformation. For many people, civil justice is already out of reach, especially those who do not qualify for legal aid. Participants pointed to real opportunities here, such as AI-supported chatbots that help people understand options short of going to court, AI-assisted victim communications that are clearer and more empathetic, and tools that explain procedures and rights in plain English. The goal, they argued, should not be just to protect those with limited digital access from being left behind, but to use AI to actively improve access to justice for the people who need it most.
The roundtable closed on a note of cautious optimism. The challenges are real, from fragmented systems and uneven capability to public trust and political scrutiny, but none are insurmountable. The sector has working examples to build on, a growing body of guidance, and staff who are engaged and willing to change, as evidenced by the hundreds of ideas submitted to the AI & Data Challenge. What is needed now is clear direction about the outcomes that matter, governance that gives people the confidence to act, and the discipline to build carefully, starting with specific, high‑value problems, rather than racing to adopt AI for its own sake.
To discuss any of the issues addressed in this article, contact NTT DATA’s global head of public sector AI via the email bill.m.wilson@nttdata.com