Earlier this month, we took our roundtable format to Leeds for the first time. Around the table were around fifteen senior leaders exploring the foundations of digital products: directors, CEOs, a COO, heads of product, data leaders and a portfolio non-exec. They came from banking, savings, insurance, energy, retail, pharmaceuticals, broadcasting, health tech and automotive software.
We framed the evening around the following three questions:
1. Where is AI actually working?
2. Who owns the customer?
3. How does change really happen?
The session ran under Chatham House rules, so I won’t be attributing anything to anyone. But the themes that emerged were as follows:
The best AI stories didn’t start with AI
The examples that landed hardest all had one thing in common: nobody led with the technology. They led with a problem.
In primary care, AI scribing is picking up the small details patients mention in passing, the throwaway comment about a pain that “probably isn’t anything”, and coding them properly. The benefit isn’t time saved. It’s better information, which leads to better diagnosis. In automotive manufacturing, modelling thousands of material variations for a single component found the right specification and saved millions in testing. In automotive retail, propensity models are helping salespeople in dealerships understand customers across every channel.
Contrast that with the stories of organisations told to “implement AI” with no outcome attached. One person described reviewing 120 proposed AI use cases at a single company. Only six genuinely needed AI. The rest could have been solved with existing technology, and pursuing them all would have burned a small fortune for no reason.
One leader put it brilliantly. Imagine simply asking how your balance looks, whether the bills are covered, and whether you can save a bit more this month, and getting a straight answer.
That’s AI underneath, but nobody needs to say the word. Every conversation at the moment starts with the tech. The ones that will win start with the customer.
Guardrails are part of the product
If one word came up more than any other, it was guardrails.
The cautionary tale of the evening involved an engineering team that gave an agent the job of fixing a bug and building a feature at the same time, with nothing in between to check its work. By the next morning it had spun up hundreds of agents, each creating new bugs for the next to fix, and the last two had decided to rewrite the entire codebase. Somebody came in and had to restore everything.
The counter-example was the same technology with a different shape: a checkpoint with a subject matter expert before anything was built, agents doing the heavy engineering lift, and a further gate before regression testing. Faster, cheaper, and it delivered something that actually meant something. Human in the loop at every stage.
In regulated sectors, this goes beyond good practice. Life sciences regulators are asking for audit trails that go right down to the individual machine where something was manufactured. One person described building a credit engine for young borrowers in Southern Europe who are underserved by traditional banks. The regulator was willing to experiment, on one condition: it had to be explainable from start to finish. If anything, they asked for more explainability than they would have from a human. Early losses were higher than planned, but improving customers’ financial literacy brought them down dramatically within months, and the data has become an asset in its own right.
There was also talk of regulators beginning to expect firms to know their agents as well as they know their suppliers. That shifts the focus from guardrails to observability: not just what you’ve allowed agents to do, but what they’re actually doing day to day.
Governance needs a rebrand
This was my favourite exchange of the night. Someone pointed out that we’d been saying “governance” a lot, and that it’s a really boring and negative word.
The suggestion was to talk about enablement and quality of outcome instead. Nobody gets excited about rules. Everybody understands the difference between buying something cheap that falls apart and buying something built properly. Make good governance desirable rather than something people feel they’re being told to follow.
That matters, because the alternative is already happening. Several people described teams working around slow processes because they can’t wait six months for a data integration. One strategy team had stuck an extract of commercially sensitive data into an AI tool, built a dashboard and demoed it to a client, without realising the data was sitting in the code. Shadow IT has been around forever. LLMs have just made it much easier.
On ownership, the room landed somewhere sensible. It isn’t tech owning AI or the business owning AI. It’s a matrix: tech owning the framework, the business owning customer outcomes, legal and risk at the table, and ideally one accountable owner across it. Underneath all of that sits the data. If you don’t trust your data before you put it into a model, you’ll trust it even less afterwards.
And then there’s education. Some organisations are piping sensitive data through consumer AI tools without any idea of the risk. Others have people who won’t touch AI because they’re worried it’ll replace them. Neither is a technology problem.
Who owns the customer when nobody opens the app?
This was the question that kept pulling the conversation back in. Someone who worked in banking through 2008 described it as the moment loyalty died. People stayed with the bank that gave them their student overdraft for years. Then they didn’t
Today, customers move money to the best rate, switch energy suppliers on price and hold several bank accounts on their phone for convenience rather than loyalty. Ask a consumer AI tool for the best fixed rate cash ISA right now and it’ll tell you. The wealth managers in the room admitted that’s exactly how easy it
has become.
The next step is agentic. In a couple of years, someone standing in their kitchen will ask an assistant to switch their tariff or move their savings, and simply authorise it. Many large enterprises spent years and billions catching up with challengers on their apps. The uncomfortable thought is that the next battleground might not involve the app at all. One person described it as AI optimisation replacing SEO, and the question of how you make sure your product is the one the agent recommends.
It’s already changing commercial models. One software business is moving from user-based licensing to pricing based on the outcomes its customers achieve. That won’t work everywhere: in financial services, consumer duty means you can’t be seen to incentivise sales at the expense of good outcomes, but it shows how far the ground is shifting.
What I took away
If there was a single thread running through the evening, it was this: the foundations of digital products haven’t changed, they’ve just become more important.
Start with the customer problem. Know your data. Build guardrails into the product rather than bolting them on afterwards. Keep humans in the loop where it matters. Organisations that skipped these steps in the digital era could often get away with it. With AI, the cost of skipping them shows up faster, and in front of more people.
The other thing that struck me was how much the different sectors had to learn from each other. Pharma and broadcasting were wrestling with the same trust questions as banking. Energy and savings were facing the same agentic future. That’s exactly why we wanted a broad room in Leeds, and why we’ll be back.
If any of this resonates, or if you disagree with all of it, I’d love to hear from you. These conversations are better with more perspectives in the room.
Red Badger is a digital product consultancy that works with enterprise organisations to build software that matters. We regularly host closed-door conversations for senior leaders. If you’d like to join a future session, get in touch.