A summary of Mike Hill (Head of Technology, UKI at Atos) and Mark Thompson's (CDIO at the Ministry of Justice) session at Modernising Criminal Justice 2026.
Most people know ChatGPT or Copilot. You ask it something, it generates an answer. That's reactive. Agentic AI is different. Mike Hill (Atos) and Mark Thompson (Ministry of Justice) explored what that means, and why it matters for how criminal justice actually works.
Understanding Agentic AI
Mike defined it simply: agentic AI has a goal and figures out how to achieve it. Instead of just generating content, it plans, makes decisions, and completes complex multi-step tasks. Think of it as a digital worker, not a smart assistant.
The power is in integration. An agent can pull data from multiple sources - gov.uk, case files, different systems - without you having to manually collect it. That matters in justice because the system is full of fragmented data sources. An agent could theoretically orchestrate information across them.
But here's the catch: agentic AI is non-deterministic. It takes different approaches to achieve the same outcome each time. That means observing, controlling, and governing it is absolutely critical. You can't just set it loose and hope for the best. That's the fundamental challenge and opportunity.
The Trust Line
Mark was asked a provocative question: if an agent could significantly improve speed, consistency, and accessibility of justice, but citizens would no longer fully understand how certain decisions are made, would you accept that trade-off? Where should society draw the line?
Mark's answer was unequivocal: the Ministry of Justice is taking a very cautious approach. Yes, they're ambitious about exploring AI opportunities. But they're working in a domain that's incredibly complicated, with legal protections and people who think carefully about change. The approach is to foremost think about trust - being clear about what they're doing and transparent. Accountability comes first.
The accountability boundary doesn't shift, Mark argued. If AI contributes to a decision that impacts someone's life - a sentence recommendation, probation prioritisation, legal support - accountability remains with the institution and the civil servants using it. That's where frameworks come in. The Ministry of Justice has the data science ethics framework (principles: safety, sustainability, accountability, fairness, explainability, data responsibility) and the AI governance framework that provides guardrails.
Where We Are and Where We're Heading
Today, the Ministry of Justice is focused on transcription and summarisation. Justice Transcribe - which probation officers use to convert interview recordings into structured case notes - has been scaled beyond pilots. (The exact numbers are being announced by Lord Timpson and the Deputy Prime Minister in the coming days.)
There are trials on document review and workflow automation. And they're thinking purposefully about connecting large language models to data sources. The justice system has roughly 900 systems and 2,000 terabytes of data in SharePoint alone. The challenge isn't data volume - it's accessing it, using it effectively, breaking out of silos, and ensuring quality.
The vision is different from just automating existing processes. Mike emphasised that this isn't about shoehorning suboptimal processes into new platforms. It's about understanding how data interconnects, how it flows across organisations, what outcomes are required - then designing an agentic landscape around orchestration, not tasks. Linear workflows give way to agents dynamically determining the best path to an outcome.
The Real Substance Versus Hype
Mike offered practical tests to distinguish real capability from market noise:
Does it complete work and not just generate content? Does it operate across systems, not just within a chat window? Can it be trusted and governed in a production environment?
If yes to all three, you've got real capability. If not, it's probably still hype. Anything that's essentially a rebranded chatbot - requiring step-by-step prompting and stopping at generating answers rather than taking action - hasn't crossed the line into agentics.
The real shift isn't just better AI. It's better and more evolved operating models. Organisations need to move from people using tools to humans orchestrating digital workers. That requires changing processes and governance as much as it requires new technology.
Building the Foundations
Mark highlighted that success requires thinking end-to-end. The Ministry is thinking systematically about the value stream for AI. They're building an integration capability, working out foundations to use AI sensibly and smoothly. Within that, AI plays are considered in the right circumstance, for the right use case.
The key insight: data quality, data architecture, and digital development matter as much as the AI itself. You can't layer agentic AI effectively onto fragmented systems. You need foundational work first - getting data organised, systems talking to each other, ensuring quality.
Mike compared it to the NHS, which has made real productivity gains with patient-doctor transcription services. The criminal justice system has similar opportunities. But they have to be pursued cautiously, transparently, and with clear accountability.
The technology is here. The frameworks are being built. But the transformation is organisational as much as technical. That's where the real work lies.
Jessica Kimbell, GovNet
