What has been the most significant challenge in embedding AI as a foundational layer within a national framework for the nuclear sector?  

The most significant challenge has been building a foundational layer around safety and regulatory compliance—ensuring we adhere to every letter of the law whilst creating space for innovation. 

Everything must be wrapped in compliance and assurance. You could argue that's a constraint on innovation, but the key is giving teams the ability to explore curiosity within controlled parameters. That balance allows us to discover new techniques safely, without compromising the rigorous standards the sector demands. 

What's emerged from this approach has been genuinely transformative. It's opened fantastic opportunities to change how we work, particularly around digital data and AI. The recommendations coming out of the Nuclear Regulatory Taskforce have been excellent, and we're thinking long-term here—not just years, but decades and centuries. 

The challenge isn't just technical—it's cultural. It's about proving that AI can be both innovative and safe in an environment where the stakes couldn't be higher. 

How have the four proof-of-concept AI projects that you have pioneered for the Nuclear Sector, including Pattern Recognition for predicting risk exposure, Financial Risk Models for budget management, and the "Risk Assistant" for real-time project analysis, collectively transformed nuclear risk management from reactive to real-time decision-making?   

There are those four that are in the public domain, but in reality, we've built over 30 different active proof-of-concept models. The fundamental principle behind all of them is driving curiosity—creating an environment where we can test, learn, and iterate rapidly. 

What we've found is that compiling data around risk aversion and using AI to analyse it has delivered some genuinely valuable, insight-driven outputs. We're not replacing experts—we're augmenting them. The models provide high-quality insights that enhance workflows and processes, allowing our teams to make faster, more informed decisions. 

The key has been getting these models up and running quickly. Failure is part of the process—it's how we understand what works and what doesn't. We're even using AI to help capture lessons learnt, so we're constantly improving. 

The shift from reactive to real-time decision-making hasn't come from one or two projects—it's come from building a culture of experimentation, learning fast, and embedding AI where it genuinely adds value. 

How has your approach to managing "acceptable risk" in AI evolved, particularly in critical sectors like nuclear decommissioning?  

In the nuclear space, risk management is absolutely fundamental—more so than in any other sector. Risk touches every workflow, every decision, every process. Risk managers here aren't just managing one type of risk—they're managing risk across the entire operation, and the stakes are extraordinarily high. 

What AI has allowed us to do is augment those risk management insights and achieve higher-quality data exploitation at pace. We're creating complex matrices of data that we can analyse in real time, which gives us genuine opportunity to make better decisions—even when dealing with deep legacy systems and decades of historical data. 

The evolution in my approach has been recognising that acceptable risk in AI isn't about eliminating uncertainty—it's about understanding it better and faster. It's about giving risk managers the tools to see patterns, predict exposure, and act with confidence, even in an environment where the margin for error is virtually zero. 

AI doesn't change what acceptable risk means in nuclear decommissioning—it changes how quickly and accurately we can assess it. 

What role does ethical AI regulation play in your work, and how do you balance innovation with compliance in government projects?  

Ethical and responsible AI is an absolute priority—it's front and centre in everything we do. We have robust policies in place, and we work closely with AI teams and leadership to ensure those principles are embedded across all use cases, both internal and external. It's mission-critical and one of the foundational pillars of our work. 

When you marry that with compliance, you've got a hard stop that cannot be compromised. Compliance and responsible usage go hand in hand, and for many organisations—particularly in the nuclear sector—this has been a critical topic for decades. 

The balance between innovation and compliance isn't about choosing one over the other. It's about designing systems where compliance enables innovation rather than constraining it. In high-stakes environments like ours, you simply cannot innovate responsibly without rigorous ethical frameworks and regulatory adherence built in from day one. It's not a trade-off—it's a requirement. 

How have your "AI Playbooks" helped standardise AI adoption across government departments, and what impact have they had on the workforce?  

As with any technology rollout, AI represents a fundamental shift in how we work—it's a new way of working, not just another application on the desktop. It brings new challenges around explainability, trust, and human culture that we need to address head-on. 

The playbooks have played an important part in navigating those challenges, but for me, this is about the journey we're on. AI isn't going away, and this isn't a vanity programme of change. It's industry-relevant, mission-critical, and designed to be fully accessible to all users across the organisation. 

The formats, styles, and approaches within the playbooks are there to find what's fit for purpose—giving teams practical, standardised guidance that actually works in their day-to-day roles. It's about making AI adoption real, not theoretical. 

What strategies do you use to improve AI literacy among civil servants and ensure they are equipped to work effectively with AI tools?  

The pace of change in AI—whether we're talking about funding at enterprise level, product development, or what vendors like OpenAI and other leading players are releasing—is relentless. To keep pace, we've built a three-tier approach to AI literacy that meets people where they are, giving them the AI tools to work effectively. 

There are three levels to our strategic approach to this: 

Base-level awareness is about making AI accessible to everyone. We run regular briefings, maintain open-screen and open-door policies, and ensure basic accessibility across all teams. The mission here is straightforward: bring awareness, help users understand what AI is, and dilute the fear and doubt that often surrounds the hype. 

Formal Training is the next level—formal sessions on how to use AI tools effectively and get the best outcomes. This is available to all, and we also run functional sessions tailored to specific teams—finance, operations, risk—so the training addresses the real needs of experts in those areas. 

Expert training is for those who need deep technical capability—certifications, deep-dive training, coding, building models, blending models. On top of that, we collaborate with academia to bring in some of the most powerful, grounded AI expertise in the world. 

I'd argue the best is yet to come, and we're building the capability now to be ready for it. 

How has your experience building custom GPTs for enterprise risk management shaped your perspective on the future of generative AI in government?

The pace of improvement in the large AI vendor space is fascinating. We're in an interesting transition period—initially, users were empowered to build mini custom GPTs for different types of work. That's now being replaced by more sophisticated agent-building capabilities. 

Tools like Copilot Studio are enabling standard desktop users to codify workflows at the highest level without needing deep technical expertise. The ability to allow real users to dip into code and build functional solutions is a significant step forward—it's changing the mindset around who can create with AI. 

What excites me most is how quickly we can bring people into the thinking of articulating the problem versus jumping straight to the solution. If you can clearly articulate the issue, you can create a fully functional app to start solutioning in a hands-on, physical way. It's about pivoting into new problem spaces—wrapped with all the standard controls—and really defining the problem before designing the solution. 

That's where the real value lies not just in the technology itself, but in how it forces you to think differently about problem definition and solution design. For government, that shift is transformative. 

How do you ensure that AI policies developed are practical and implementable by technical teams on the ground? 

Policy adherence is non-negotiable—it's something you need to do. The challenge is that the pace of change in AI is so rapid that real compliance and regulatory compliance become critically important. You can't afford to develop policies in isolation from the reality of implementation. 

That's why we take an open-screen, open-door approach to innovation policy. We encourage teams to bring ideas forward, test them, and see what level of viability they have. It's about creating a feedback loop between policy development and practical application—so policies aren't just theoretical documents, but tools that technical teams can actually work with on the ground. 

We involve technical teams early in the policy development process, test policies in real-world scenarios, and iterate based on what works. That way, compliance isn't a barrier to innovation—it's built into the process from the start. 

What do you see as the next major milestone for AI in the nuclear sector, and how do you envision its role evolving over the next decade? 

Over the next 12–24 months. we're going to see more examples of the bad use cases of AI—deepfakes, generation of software viruses, creation of misleading images in socially and culturally sensitive positions. The fundamental question becomes: just because you can, should you? 

We're facing cultural challenges around AI use that echo the debates we've had with social media—issues around data ingestion, digital ethics, and how we govern these technologies responsibly. In the nuclear sector, where safety and security are paramount, these risks are amplified. 

But there's also the positives. We're beginning to surface genuinely transformative discoveries—cancer diagnosis, pharmaceutical drug development, breakthroughs in materials science. These are fantastic innovations that simply wouldn't exist without AI, and there's much more to come. 

For the nuclear sector specifically, I see AI playing a critical role in predictive maintenance, safety monitoring, regulatory compliance, and operational efficiency. The milestone will be when we can deploy AI at scale in safety-critical environments with full transparency, explainability, and trust—where the technology enhances human decision-making rather than replacing it. 

The next decade will be about proving AI can be both powerful and safe in the most demanding environments. That's the challenge, and the opportunity.