Inside Barclays: turning AI experiments into real-world impact

Francois Buet-Golfouse's team at Barclays has reviewed over a thousand AI ideas in less than two years.

The ones that made it to production weren't the flashiest. They were the ones that solved real problems, fit into actual workflows, and earned trust from the people using them. AI teams aren't short on ideas. The challenge is knowing which ones deserve to become systems that people rely on.

Francois is Global Head of AI and Machine Learning for Global Markets at Barclays. He works at the intersection of quantitative finance, machine learning, and trading infrastructure. His job is to figure out which AI ideas should become real systems in one of the most complex and regulated environments in finance.

We spoke to him on the DevLab Podcast about what it actually takes to move AI from experiment to production, how Barclays builds responsibly without slowing down, and where AI in finance is heading beyond the noise.

Starting with the problem, not the model

Francois is clear about where successful AI work begins: before anyone picks a framework.

"The idea is to always start from the business problem and the strategy as opposed to the technology itself."

It sounds obvious, but it's where a lot of AI projects drift. A new model drops, a team runs a proof of concept, and only later does everyone ask where it actually fits.

At Barclays, the entry point is different. Francois' team works with people across sales, trading, operations, and technology to understand what problem needs solving. Which decision could be sharper? Which workflow is too manual? Where could colleagues or clients have a better experience?

A thousand ideas, and the skill is choosing

The scale of opportunity inside an organisation like Barclays is huge. Francois told us his team has reviewed "a thousand ideas or more" in under two years.

AI leadership isn't measured by how many ideas you generate. It's measured by how well you choose.

Francois filters on three things. Impact: will this genuinely help colleagues, clients, or stakeholders? Reusability: could the same pattern help other teams or asset classes? Effort: how difficult will it be to build, test, and maintain?

It creates room to explore without treating every idea as equally valuable.

It also means stopping work early when it's not panning out. If a proof of concept shows a use case isn't worth pursuing, the process has done its job. The team learned something before committing too much time, budget, or risk.

Production is where it gets real

A demo can look impressive in a controlled environment. Production asks harder questions.

Will people use it? Does it work with messy data? What happens when the input is unexpected? How is it monitored? What controls sit around it? Does it improve the workflow enough to justify the change?

Barclays approaches this with a staged process. Early ideas can be explored, tested with pilot users, and moved towards production if they prove useful. Teams get space to experiment while keeping a clear route to live use.

One example from the conversation: a bot built to read client trade requests. On the surface, that sounds like a focused workflow improvement. But once teams started working with it, they realised clients could ask for more through the same channel.

The AI work didn't just improve one task. It helped the team rethink the client experience and what that interaction could become.

This is also where Barclays' scale becomes an advantage. A pattern that works in one part of global markets might be useful somewhere else. If a solution can be abstracted and reused, the value doesn't stay locked inside one use case.

Moving fast without chasing everything

AI changes quickly. A better model or tool can appear with little warning, and teams need to be able to move when something genuinely improves what they can build.

But there's a trap in speed: chasing every new thing makes it harder to finish anything.

Francois describes the balance well. Teams need to be agile enough to use better approaches when they arrive, but careful enough not to rebuild around every new framework. That means designing systems with flexibility in mind. If a better large language model or agent framework shows up, teams should be able to swap components without starting from scratch.

That kind of engineering judgement is easy to overlook in broader AI conversations. The flashy part is the model, but the durable part is the architecture around it.

For Barclays, this means working with external solutions where they're useful, developing capabilities in-house where that makes sense, and avoiding unnecessary dependence on any single approach.

Responsible AI has to be built in

In financial services, trust carries weight.

Barclays has been around for centuries, and that history creates a clear expectation: new technology needs to be handled carefully. AI systems have to be useful, but they also need to be explainable, controlled, and appropriate for the context they operate in.

Francois puts it directly: "We want to embed a lot of guardrails, a lot of controls of what we do."

Those controls aren't treated as a final box to tick, they're part of how the system is built.

That involves input from legal, compliance, model risk, cybersecurity, and other specialist teams. Each group brings a different lens to the same question: what needs to be true for this system to be safe and trusted?

For teams outside financial services, that might sound heavy. In practice, it can be an advantage. It forces better thinking earlier. What should the system be allowed to do? Where does a human need to remain involved? How should uncertainty be handled? What would make an output explainable enough to use?

These questions become more pressing as AI systems become more capable and autonomous.

LLMs are only part of the story

Large language models have shifted what people expect from AI. They make it easier to interact with tools, gather information, summarise content, and trigger workflows through natural language.

Francois sees that value clearly, but he's careful not to reduce AI to LLMs alone.

In finance, many problems still need traditional statistics, machine learning, deep learning, and subject matter expertise. In some cases, the most useful setup is a combination: an LLM interprets intent or calls tools, while a more specialised model handles a specific task like forecasting, pricing, risk, or analysing structured data.

That should be reassuring for engineers, data scientists, and ML practitioners who've spent years building technical depth. The future of AI will still need people who understand models, uncertainty, architecture, data, and systems.

Francois' point about coding assistants makes this tangible. Some bugs are loud. They break the code and are easy to spot. Others are quieter. The code runs, the output looks plausible, and the issue only appears if you understand the logic well enough to question it.

"When you look a little bit more closely at those outputs, you realise there's something that's a little bit off. This is what we'd call some extent of silent bug. Not something that's going to break your model running, but simply that actually there's a small thing in there that's not quite exactly what you wanted. You've got to be able to detect that."

AI tools can help people move faster, but speed is only useful when someone can still judge the quality of the result.

Agentic AI needs ambition and boundaries

Agentic AI is one of the most talked-about areas in tech right now. The promise is compelling: systems that can take a goal, plan steps, call tools, gather data, analyse results, and return something useful.

In finance, that could support workflows like preparing reports, pulling together internal and external data, running analysis, comparing outputs, or drafting information in a format people can use.

Francois sees real potential here. He also points to the work needed around guardrails, controls, alignment, uncertainty, and delegation. The more autonomy a system has, the more clearly teams need to define what it can do, when it should stop, and when it should hand back to a human.

Finance has been working with model risk, controls, and governance for a long time. AI introduces new challenges, but it doesn't arrive in an organisation with no existing discipline around risk.

The Barclays perspective shows how agentic AI can be explored with genuine ambition while respecting the boundaries that matter in a high-stakes environment.

What AI professionals should focus on now

For people working in tech, the pace of AI can feel exciting and exhausting. New tools arrive quickly. Old assumptions shift. Skills that once felt specialist are becoming part of everyday engineering and data work.

Francois' advice is grounded: critical thinking, strong foundations, and adaptability.

Understanding models properly. Knowing enough statistics, machine learning, software engineering, and architecture to ask better questions. Using AI tools without blindly trusting them.

As AI generates more code, analysis, and content, the ability to evaluate quality becomes more valuable. Can you spot the silent bug? Can you understand why an answer looks right but isn't? Can you design systems where the pieces work together, rather than simply producing more pieces?

These skills will become more valuable as AI gets embedded in how teams build.

The human side of AI

One of the most interesting threads in the conversation: AI might create more room for human work.

If AI handles more repetitive, time-consuming, or data-heavy tasks, people can spend more time on judgement, relationships, and higher-value decisions. In markets, where client relationships and trust matter deeply, that's a meaningful shift.

Francois is careful not to make bold predictions about where AI will be in five years. The future is moving quickly, and serious teams are honest about what's still uncertain.

What feels clear is the direction. AI systems will become more embedded in daily work. The organisations that get the human-AI relationship right will have an advantage.

Francois described it as one of the defining challenges ahead: how to manage the collaboration between humans and AI in a way that lets people focus more on the relationship aspect of their work, whether at work or outside it.

Barclays' approach is grounded in that thinking: explore widely, choose carefully, start with the business problem, build in controls, keep humans involved where judgement matters, and make sure good ideas have a real path to production.

---

Watch the full conversation

Want to hear more from Francois on AI in global markets, agentic systems, responsible AI, and what skills matter for the next generation of AI professionals? Watch the full DevLab episode with Francois Buet-Golfouse, Global Head of AI and Machine Learning for Global Markets at Barclays.

[We will embed the YouTube video here]