There is a version of the AI conversation that happens at conferences. Governance frameworks. Ethics by design. Responsible adoption. It sounds right, it looks polished, and some of it is genuinely useful. But the most interesting conversations are the ones that get underneath it, where the people actually making consequential decisions about AI talk candidly about what that looks like in practice.
In May, Dr Raoul Gabriel-Urma spoke with two sets of guests doing exactly that. First, Nick Edwards, Group Chief Technology Officer at the AA, and Harj Johal, the AA's Managing Director of Customer Operations. Two leaders in the middle of transforming a 120-year-old British institution using AI from the ground up. Then, Jessica Rusu, Chief Data, Information and Intelligence Officer at the Financial Conduct Authority, one of the most influential figures shaping how AI is governed across UK financial services.
The contexts could hardly be more different. One guest oversees AI governance across 50,000 regulated firms. The other two are deploying ChatGPT enterprise licenses to 1,500 colleagues and launching a Dragon's Den funding process for internal AI ideas. But the questions at the heart of both conversations are the same. What does responsible AI actually look like when you move beyond the framework and into the organisation? And how do you scale something genuinely transformative without losing control of it?
Jessica Rusu's position on AI regulation is deliberate and worth understanding carefully, because it runs counter to what many people in financial services expect from a regulator.
The FCA has not introduced new AI-specific rules, and this is a conscious decision. Its position is that the existing principles-based framework, anchored in the Consumer Duty, is not a weaker starting point than a prescriptive AI rulebook. It is a stronger one. The Consumer Duty requires firms to deliver fair value, meet genuine consumer needs, and ensure that no group of customers is excluded from access to financial services. Those obligations do not have an AI exemption. They apply regardless of whether a decision is made by a human or a model.
As Jessica puts it, there is no one-size-fits-all definition of what responsible AI looks like. A back-office automation tool, a consumer-facing chatbot and a credit decisioning model each carry entirely different risk profiles. Telling firms that ticking ten boxes constitutes responsible AI would not make the system safer. It would give firms a route to compliance that has nothing to do with outcomes.
What the FCA is doing instead is working alongside firms as they build and deploy. The AI Live Testing initiative places FCA teams inside organisations, working in real time as products go to market and asking the questions a regulator should ask. How was the data selected? Why was this model chosen? What guardrails are in place? What happens when something goes wrong, and who is accountable?
That last question matters most. Jessica is direct on this point: firms cannot use AI as a shield. If a model produces a harmful outcome, the fact that it was a black box is not a defence. If it is your firm, you have to understand how it works, and you have to understand the impact on customers, all of the time.
Beyond governance, Jessica's episode covers another area that does not get enough attention in the AI conversation: the role of shared infrastructure in accelerating responsible innovation.
The FCA's digital sandbox, built in partnership with technology providers and now running continuously rather than in fixed cycles, gives fintechs and innovators access to something they cannot easily build themselves. Hundreds of high-quality synthetic data sets covering fraud typologies, transaction data and market data, a technical environment where firms can develop and test products over a three-month sprint, and direct access to a regulator throughout the process.
The impact, as Jessica describes it, is not incremental. Firms regularly report that work which would have taken a year independently has been completed in three months. And the benefits run in both directions. The FCA learns as much from the firms it works with as those firms learn from the process.
The underlying insight here extends well beyond financial services. When there is a centre of gravity, a shared space where organisations can bring problems, contribute data and work alongside others facing the same challenges, collaboration happens that bilateral agreements alone cannot broker. The sandbox is an example of what it looks like when a regulator decides to act as an ecosystem builder rather than just an enforcer.
The AA's episode opens with a scenario that cuts straight to the stakes. A family broken down on the M6 at midnight. You do not know what is wrong with the car. You do not know who is in it. You need to get the right resource, with the right parts, to the right place as quickly as possible.
Nick Edwards uses that image to describe both the opportunity and the risk of getting AI wrong in an operationally critical business. Get it right and the experience is transformed. Get it wrong, and the consequences are real for a real family at a genuinely vulnerable moment.
What makes the AA's story compelling is not the technology. It is the sequence. Before AI entered the picture, the business spent three years on a serious digitisation programme, rebuilding its breakdown reporting capability, reducing unnecessary contact centre volume and creating the kind of digital infrastructure that colleagues and customers could actually use. That foundation mattered. When AI arrived, the organisation was not starting from scratch. It was already on a journey, and the appetite to continue was already there.
Nick is candid about what happened next. When the AA deployed 1,500 ChatGPT Enterprise licenses, the early adoption data showed up in unexpected places. Not the technology function. Finance. HR. Operational teams. Within the first couple of months, colleagues had built 300 custom GPTs without being asked to. A champions' network of 75 colleagues, drawn from across the business at every level, formed organically and started running their own internal sessions to bring others along.
That is not a top-down transformation story. It is a bottom-up one that a well-constructed top-down framework made possible.
Harj Johal's contribution to the episode is a clear articulation of what AI might mean for the future of customer operations, and it is worth dwelling on.
The conventional anxiety about AI in contact centres is that it reduces headcount. Harj's argument is that it does something more interesting. It upgrades the role. His vision for the AA's contact centre colleagues is not a smaller team doing the same job more efficiently. It is a different job entirely: individuals who can access vehicle data, customer history, logistics systems and third-party services simultaneously, owning a customer's situation end-to-end rather than handing it between departments.
Using the same midnight breakdown scenario, he sketches out what that looks like in practice. One person, looking at real-time vehicle data, sees that the car cannot be fixed roadside, arranges a recovery truck rather than a patrol van, identifies a nearby hotel, books it, and manages the whole situation from a single point of contact. What previously required five or six people to each own their piece of a fragmented process becomes one person's responsibility, supported by the right tools.
The AI in that picture is not replacing human judgement. It is giving one person access to enough information and capability to exercise judgement that was previously impossible at the individual level.
Both episodes, despite their differences, return repeatedly to the same quality in the people who make AI transformation work. Not technical expertise. Not seniority. Curiosity.
Jessica talks about it in the context of career development, describing the importance of building both vertical depth in a domain and horizontal breadth across sectors and functions. The people who navigate technological change well, in her experience, are those who keep moving, keep learning and keep adding new strings to their bow rather than waiting for certainty before they act.
Nick and Harj talk about it in the context of organisational culture. The colleagues who built 300 GPTs in the first two months were not the ones who had been told to. They were the ones who were curious enough to try. Harj noted that some of the AA's senior leaders, having gone through a structured AI learning programme, had surprised themselves with how far they had come in just four months, presenting live demonstrations of tools they had built while still doing their day jobs.
That kind of progress does not come from a framework. It comes from an environment where people feel safe enough to experiment, supported enough to develop, and trusted enough to act on what they find.
A financial regulator operating on principles rather than rules. A breakdown services company deploying AI from the operational front line upward. Two very different organisations, but the same underlying conviction running through both conversations.
Responsible AI is not a checklist. It is a culture, a set of ongoing judgements about outcomes, accountability and the impact on people, made by leaders who are close enough to the work to understand what they are actually deciding.
That is a harder thing to build than a governance framework. But it is the thing that lasts.
______________________________________________
If you are a data and AI leader looking to build capability across your organisation, Cambridge Spark works with businesses at every stage of that journey. Find out more at cambridgespark.com.