By Phoebe Seers and ‌Kirstin Ridley

LONDON, Dec 17 (Reuters) - A race among banks to adopt agentic AI, which can make decisions and take autonomous action, runs new risks for retail customers, said ‍Britain's financial ‌watchdog, which is pledging to ensure their interests are not sidelined.

AI 'agents' could revolutionise how people budget, save and invest, for example by moving idle cash into higher-yield accounts or ⁠adjusting portfolios in response to market swings.

Unlike generative AI, which produces text, code or ‌images based on human prompts, agentic AI is touted as the next big opportunity for companies because it can plan, make decisions, execute tasks, learn and adapt based on set goals.

Britain's NatWest, Lloyds and Starling told Reuters they are working with the FCA as they prepare for retail-customer trials, in a major shift from banks' existing back-office usage.

HITTING THE MARKET IN 2026

Financial Conduct Authority (FCA) Chief Data Officer Jessica ⁠Rusu expects early consumer-facing applications to hit the market "in earnest" early next year.

"Everyone recognises that agentic AI introduces new risks, primarily because of ... the ability for something to be done at pace," Rusu told Reuters.

The autonomy and ​speed at which the AI acts and interacts with other agents magnify risks relating to financial stability and ‌governance. 

The FCA will apply rules such as the senior managers regime and the consumer duty ⁠to hold bosses responsible for wrongdoing and ensure customer interests are put first, Rusu said.

BRITISH BANKS LEAD PILOTS

Research firm Gartner predicts that 40% of all financial services firms will be using AI agents by the end of 2026.

However, it also expects that more than 40% of agentic AI projects across industries will be axed by the end ​of 2027 as a result of escalating costs and unclear business value.

British banks lead European rivals in trialling AI agents for customer-facing tasks, partly because of the FCA's approach which includes an AI sandbox for firms in the experimental phase and a recently announced live testing initiative, for those ready to roll out new AI applications to UK financial markets. 

Meanwhile, the European Union's AI Act, which aims to foster responsible development and deployment in the bloc, has created uncertainty about how agentic AI in finance will be treated.

U.S. banks ​including JP Morgan ‍are using agentic AI for back-office work. JPMorgan declined ​to comment on plans to deploy virtual banking agents for customer-focused actions. 

NatWest said it is testing the technology to accelerate complaints handling by investigating and analysing cases - a feature it will announce in early 2026.

Lloyds last month announced a pilot for employees aimed at helping customers to better manage their money.

Savings could one day be automatically invested into tax-free ISA accounts if customers agree to it beforehand, Lloyds' chief data and analytics officer Ranil Boteju told Reuters.

Starling's chief information officer Harriet Rees said customers will soon be able to ask for personalised budgets and set predictive spending caps based on past habits, with an automated tool "setting up all the ... standing orders and spending trackers to make this happen".

MULTIPLE AGENTS CREATE SYSTEMIC ⁠RISKS

While agentic AI performs well on simple, short tasks, it fails on more complex ones, said Ram Gopal, professor of information systems at Warwick Business School. 

And bigger challenges could lie ahead.  

The systemic risk lies not with one AI agent, but with many interacting ​and executing actions simultaneously, said Suchitra Nair, head of Deloitte's EMEA Centre for Regulatory Strategy.

"These AI agents could react to identical market signals, rapidly shifting deposits or funds between accounts, dramatically accelerating the probability and pace of a bank run, for example," said Nair. 

Reliability is also a concern as AI systems can "hallucinate", suggesting responses that sound plausible but are untrue.

"How well do they know the person they're 'talking' to? That's a real problem in an advisory context," said Taylor Wessing lawyer ‌Martin Dowdall, who also questions how well bank executives understand the technology.

"These systems are designed by the cleverest MIT computer scientists ... Does the average banker really understand what it takes into account or what it doesn't? Probably not," he said.  

(Reporting by Phoebe Seers, Kirstin Ridley; Additional reporting by Jesus Aguado in Madrid; Editing by Tommy Reggiori Wilkes and Alexander Smith)