Banks vs Bots: Agentic AI Opens Up a New Fear for the Finance Industry

For years, the financial sector’s relationship with artificial intelligence has been one of cautious enthusiasm. Banks have deployed machine learning to spot fraudulent card transactions, screen loan applications and automate the drudgery of compliance paperwork. Now a newer breed of software u2014 so-called agentic AI, which can take actions on a user’s behalf rather than simply answer questions u2014 is prompting a far more uncomfortable conversation in boardrooms and regulators’ offices alike.

The worry is simple to state and difficult to solve. If an AI agent can log into an account, compare products, move money and execute a trade without a human pressing the button at each step, then the bank on the other side of the transaction is no longer dealing with a customer in any conventional sense. It is dealing with a bot acting for a customer u2014 and the rules that govern banking were not written with that in mind.

Who is the customer?

Much of modern financial regulation rests on assumptions about human behaviour: that a person consents to a payment, that a customer understands the risks of a product, that identity checks verify a living individual. Agentic systems blur all of this. When an autonomous program initiates a payment, questions of authorisation, liability and consent become genuinely novel. If an agent misreads an instruction and empties a savings account into a bad investment, who is responsible u2014 the customer, the bank, or the company that built the model?

Anti-money-laundering frameworks face a similar strain. Know-your-customer rules are designed to tie accounts to real, identifiable people. An ecosystem in which software agents open relationships, route funds and interact with other agents at machine speed is a harder environment to police, and potentially an attractive one for bad actors.

Speed is the systemic risk

Beyond individual transactions lies a broader supervisory anxiety: herd behaviour. Banking crises have always been partly a story of panic, but panic used to move at the speed of queues outside branches. Recent years have already shown how quickly deposits can flee when customers can tap a smartphone. Agents acting automatically on price signals or news could compress that timeline further u2014 and if many agents are built on similar underlying models, they may react in the same direction at the same moment, amplifying swings rather than smoothing them.

That prospect is the sort of thing that keeps financial stability watchdogs awake. Concentration risk is also in play: a handful of AI providers could end up underpinning decisions across large parts of the market, creating a dependency that resembles the sector’s existing reliance on a few cloud computing giants.

Defence and opportunity

Banks are not purely on the receiving end. The same technology promises genuine gains u2014 faster customer service, sharper fraud detection, cheaper back-office operations u2014 and institutions that ignore it risk being outpaced by nimbler fintech rivals. Fraudsters, meanwhile, are expected to use agents of their own, meaning banks may need automated defences simply to keep up with automated attacks.

The likely result is a period of rule-writing and improvisation. Expect debates over whether AI agents should be registered or identifiable when they interact with financial systems, over how banks verify that an agent genuinely speaks for its owner, and over where liability sits when things go wrong.

For an industry built on trust, the hardest question may be the most basic one: how do you trust a customer you can never meet? Read More