Keynote: Scaling artificial intelligence in highly regulated environments
International keynote — 5 August 2026, 9:10am Dan James, Advisor to the UK Cabinet Office and Director of AI, UK Ministry of Justice.
The Day 2 international keynote came from Dan James, who leads AI at the UK Ministry of Justice. The topic: what it takes to scale AI in one of the most highly regulated environments in government.
The human gap
After telling us a bit about his background, Dan opened not with technology but with what he called the human gap: trust, participation, social engagement, vision, incentives, skills, confidence and change appetite.
He'd been conducting a programme of interviews, and described how they changed almost overnight, when ChatGPT was released in December 2022. The shift was particularly stark among the technologists he was speaking to, who moved to talking about doing much more with less, and about not needing software engineers.
Dan became interested in looking at both the non-technical themes coming out of those interviews and the exciting technology that had emerged.
Saying yes to the job
The story of how he took the Director of AI role was fascinating. He was in Berkeley when the phone rang: it was the UK Ministry of Justice, asking him to join the office as Director of AI. He turned it down. He'd been through the challenges in government before, and wasn't convinced the conditions would be in place for him to succeed.
But they kept ringing, and in each conversation James would raise another objection about why not to join. Looking back, he realised he was effectively writing a job spec as his objections were answered one by one.
Did he have the budget to do this? Yes.
Autonomy to run the team as he wanted? Yes.
Could he recruit his own team and lead them his way? Yes.
Could he deviate from traditional government ways of working to try new things? Yes.
In the end he had very few objections left, and he joined the Ministry of Justice.
The challenges he inherited
Dan was candid that the challenges when he joined were severe:
- A court backlog of around two years
- Prisons at 99% capacity
- Probation officers working at 110% capacity — into the night, up to 1am, doing case work
- People had also been released by mistake in 2024, because identifiers for individuals are split across many different systems (courts and others), the ID isn't consistent across those systems, and the legacy systems weren't joined up.
The AI Action Plan
The response was an AI Action Plan built on three strategic pillars:
- Strengthen the foundations
- Embed AI
- Invest in people and partners
Scan, pilot, scale
To embed AI, the team followed a "scan, pilot, scale" approach:
- Scan the market for challenges and the solutions out there
- Pilot rapidly
- Scale up across staff
Dan noted that scaling is the hardest part.
Five tips for scaling AI
He offered five tips for scaling AI:
- Lay strong foundations — data protection and ethics from day one, an AI and data science ethics framework, strong evaluation practices, and union engagement as the key to trust. The goal is to combine governance with fast iteration.
- Small teams, big outcomes — small teams that live with the product all the way to scale, rather than handing it over to a different business-as-usual team. He used the Justice Transcribe case study as an example of this strategy (see below).
- Get onsite — they use a forward-deployed approach, with engineers going out into the field. Dan showed a photo of someone on their fourth day holding keys to a prison, free to walk around and interact with prison officers to understand the pain points on the ground. His point was that this is very different to the traditional approach inside government, and that what you understand from behind a desk is very different to what you understand onsite. Putting technical expertise on the frontline builds strong relationships with frontline staff, and produces more innovative solutions to problems the team didn't know existed.
- Make it voluntary — nobody had to use their products. Dan said this mattered because there is real fear about AI (how it might displace jobs or negatively affect individuals) and being able to say "you don't need to use it" was helpful. It also meant their products had to prove themselves by people wanting to use them, rather than through a mandate.
- Work with startups and scaleups — government needs to increase appetite for working with smaller enterprises.
A new approach to digital governance
The Ministry also moved to a new approach for digital governance, with the AI team and the Ministry of Justice introducing what James calls proportionate governance, based on where the product is at. The team builds evidence as they pilot, so when it’s time to scale and start going to governance meetings they have all the info and data they need to address governance questions and issues.
Human in the loop is not a safeguard on its own
Dan also pointed out that a human in the loop isn’t always enough because generative AI can produce confident prose that humans can read and agree with, so it’s essential to mitigate against this.
Case study: Justice Transcribe
One of the key case studies Dan referred to was Justice . This AI product was originally built by just two people to address a key pain point: the UK has around 12,000 probation officers who spend a lot of time taking notes, looking down at paper or a laptop rather than looking the person in the eye and building a strong relationship. They also spent many hours a day writing up case notes.
Justice Transcribe transcribes meetings and drafts case notes. Dan reports that it's now been used in over one million meetings, has saved 166,000 hours so far, and from 70,000 in-app reviews has an average rating from probation officers of 4.7 out of 5. The product team is still only five people, and they've used AI to help with coding, customer support and more.
This case study demonstrates the “small teams, big outcomes” tip for scaling AI while also being an example of starting with a pilot and then scaling up.
Tech is advancing rapidly — lean in
Dan closed on the pace of change. The interviews he'd been doing at Berkley originally took six months, conducted in-person at three meetings a day. Since then, JobGraph has developed an AI harness that can interview people at scale. It gathers all the transcripts, analyses them, and turns them into a report similar to what you might get from McKinsey. The report captures the key points people raised, the challenges and the opportunities. It's being used to carry out social-science interviews, and can also form part of the government discovery process. In government, discovery with user testing might take six months, and this can cut it down to two weeks.