What Are The Risks Of Agentic AI, According To UK Banks?

Banks experimenting with agentic AI payments say the payment function is the most sensitive area of banking. Cécile De Sousa, Apac chief operating officer at Natixis, is wary. She says: “For me, the payment function of the bank is the most critical level of bank functions. Deciding that we will rely completely on agents or even supervise an army of agents seems to be difficult for me to comprehend.”

Her view is about handing execution power to software. Agentic systems can act on instructions and complete purchases without a human pressing confirm each time. In trials reported in The Banker, customers set parameters and the agent completes the transaction, sometimes with preapproval. That creates efficiency, but it also hands operational control to a model that may misinterpret instructions.

There is also the risk of errors generated by large language models. Sandeep Malhotra, executive vice president for core payments, Asia Pacific at Mastercard, says early pilots are needed to identify issues and “explore dispute resolution if there are hallucinations from the agents.” If an AI misreads a price cap or date, a payment could go through incorrectly.

Liability adds to the risk, David Tilbury of Pinsent Masons says: “There is no clear answer at the moment as to how liability is apportioned if [the agent] gets that wrong, if it makes a purchase that goes above a spending threshold, or if it doesn’t quite align with set parameters.” Until that question is settled, banks carry legal uncertainty each time an agent executes a transaction.

Can Existing Payment Rails Cope With Machine Driven Commerce?

Marc Boiron, chief executive of Polygon Labs, believes infrastructure could become a bottleneck. He says: “Traditional payment infrastructure was built for human-initiated payments, not continuous machine-driven commerce.” If AI agents begin making transactions throughout the day without human pacing, volumes could increase sharply.

At present, agentic payments run on existing card networks. According to The Banker, transactions are tokenised in a similar way to digital wallets such as Apple Pay, meaning the card number is not directly shared. That protects data, but it does not change the underlying rails.

De Sousa questions the value of agentic AI in payments under current conditions. “It is only valuable if it goes beyond agentic, and is coupled with technologies like stablecoins and blockchain,” she says. Without changes to settlement systems, she believes the benefits may be limited.

Banks also face operational strain if transaction volumes increase faster than fraud controls and monitoring tools adapt. As Abhijeet Ramesh, head of innovation, growth products and partnerships, Asia Pacific at Visa, says, banks may need to adapt their processes around risk monitoring and customer experience as agentic transactions become more common. That implies added cost and redesign of internal systems.

Are Governance And Return On Investment Also At Risk?

Operational risk is only one side of the issue. McKinsey’s 2025 global banking report, discussed on the McKinsey Talks Operations podcast, found that nearly 80% of financial institutions it works with in Asia report using some version of AI led applications. A similar proportion globally report no significant impact on their bottom line.

Abhilash Sridharan, a partner at McKinsey & Company, says creating value from AI “won’t be a cakewalk, and it’s going to take time.” He explains that institutions need to rewire operations, frontline distribution, technology, data science and risk management with AI at the core. When AI projects are run in silos with unclear linkage to financial value, returns can disappoint.

He also speaks against deploying limited use cases. Building a chatbot or a credit memo writer may deliver early wins, but banks can plateau if they do not redesign processes end to end. That creates investment risk, where spending on pilots does not translate into measurable gains.

David Deninzon, a senior partner at McKinsey & Company, notes that between 50 and 60% of bank full time equivalents are tied to operations. That scale means agentic AI has large potential impact, but it also means mistakes can have large operational consequences. Changing how work gets done across such a large workforce carries execution risk.

Banks such as Banco Santander have begun controlled trials. Santander and Mastercard recently completed Europe’s first live end to end payment executed by an AI agent within a regulated framework.

Matías Sánchez, global head of cards and digital solutions at Santander, says: “At Santander, we see AI as a transformative force in the evolution of payments. Our role is not only to adopt innovation, but to shape it responsibly, embedding security, governance and customer protection by design. As AI agents become part of everyday commerce, building trusted, scalable frameworks will be essential to unlocking their full potential.”