Agentic AI could transform UK banks – but only if they stop playing it safe

By Rishi Kapoor, Global Head of Partner Sales Engineering and Solutions, Alteryx

UK banks are stuck in an AI comfort zone.

While deployment of the technology in the financial services and banking (FS&B) sector is widespread – 75% of UK financial services firms are now using AI – the direction of use cases is important. Anthropic has made a buzzy push into agents for financial services this year, but the hard truth is that such cutting-edge use of AI is abstract to the majority of firms.

If we look at why, a key factor is sticking to the status quo. AI is being applied in areas where data is already structured, and analytics exists. Workflows in risk analytics, fraud detection and regulatory reporting are being transformed with AI and easing workloads.

The problem is that such applications of AI, fast-forwarding four years on from ChatGPT’s launch, are now a commodity. The shift to agentic systems and workflows holds exciting potential to go further with applications in customer onboarding, experience personalisation, Anti-Money Laundering (AML) and more. But this requires wrestling with the fragmentation of data across business units and formats that’s commonplace in banks. It’s easy to sit out, but the longer banks do, the longer they’ll lack the means for agentic systems to thrive.

Adding to the hesitancy is risk. Deploying agentic AI over a broken workflow or legacy system could result in hallucinations and inaccurate output that triggers a bad internal decision or customer experience. AI agents move quickly, but without careful work to root agents in a bank’s unique workflows, it can be the wrong way. Banks can’t afford to run that regulatory risk.

The good news for a growing number of banks taking on this undertaking is that there is a path forward.

FS&B tech stacks – advanced but can go further still

There’s no question that FS&B firms are advanced in the use of data and analytics compared to other industries. Data-backed dashboards are commonplace in the management of risk, compliance, fraud claims and customer segmentation. They work well because they use structured data that teams can trust the accuracy of.

There’s also a wealth of context in banks that remains locked away – in legacy systems, spreadsheets or, simply, undocumented human knowledge. Ledger’s policies, approvals, exceptions, local business rules, institutional knowledge. Data that’s the lifeblood of critical workflows but remains scattered and not available for automation.

This problem with this status quo is that it won’t deliver the promise of agentic AI systems. Agents aren’t dashboards. Their impact and potential for embeddedness depend on the depth of high-quality company data that they can use, and a unique set of governance controls that give leaders the confidence to sanction agentic use cases – safe in the knowledge that they’re not deployed over fragmented or poorly governed data.

The last point is important; banks aren’t strangers to governance but need to rethink the foundations for agents. Restrictive access control, manual approvals, segmented teams, and ethical frameworks that don’t sit in digital systems all need to be addressed with an operating system refresh that entails roles-based access to setting up AI agents with human-in-the-loop governance.

How banks can get started

A helpful way to picture that operating model refresh is as a pre-agent layer where governance and business logic sit.

Governance comes in the form of a central data layer and platform that renders agent data inputs visible, understandable, repeatable, and auditable. Platforms that make it easy to set up repeatable workflows for AI systems (consistently arriving at objective answers), are visual in nature (so much so that managers can look at a workflow, quickly grasp its inputs and outputs and sign off with ease) and rich data connector options will make agents operational and embedded into how work gets done more quickly.

Which brings us to business logic. The locked-away context I’ve already described doesn’t sit with IT. It sits with individual business units. Banks need to equip these teams with the means for these teams to identify the agent use cases that will be most impactful for their work and insert the required business context that aligns agents with the way internal processes really work. And while IT teams would love a world where the data required for this sits in one data lakehouse – it’s never that simple in reality. A platform that can unify data is a must.

What’s next

Banks have everything to gain by ditching their AI comfort zone. Applying AI to workflows with mature data foundations was only natural. The pivot to agentic systems requires them to go further. Prioritising visible, understandable, repeatable, and auditable inputs and outputs is the right philosophy to get the task of readying more data for agents right.