Can an AI Agent Use Your Bank Account?

A chatbot can explain a bank statement. An AI agent could be connected to a service that reads account information, sorts transactions or prepares a payment. Those are different levels of authority—and the difference matters when money is involved.

The practical question is not whether AI can be useful with money. It is what the software is allowed to see, decide and do, and who must approve a consequential action.

Reading an account is not making a payment

In the UK, regulated account-information services can, with a customer’s permission, read information from a bank account through approved connections. Payment-initiation services are a separate type of service: they can ask a bank to make a payment after the customer authorises it. A permission to read transactions should not quietly become permission to move money.

A useful AI agent might group spending, flag a bill that looks unusual or prepare a list of payments for you to review. If a system can also choose a payee and submit a payment, it has crossed into a more consequential role. The bank, the payment provider and the agent each need clearly defined responsibilities.

The UK is discussing what future delegation could look like

In a June 2026 consultation on the design of future UK retail payments infrastructure, the Bank of England’s Retail Payments Infrastructure Board discussed possible “advanced delegated payments”. The paper described people granting third parties authority to make payments for specific purposes, and noted that this could include an AI agent. It was a consultation about future infrastructure—not an announcement that UK banks now let general-purpose AI agents make payments.

That distinction is important. A proposal can help policymakers and the payments industry think through standards; it does not confirm that a particular bank supports a feature, or that an agent has the right safeguards.

How can an agent go beyond what you meant?

Agents interpret goals and use connected tools. They can misunderstand an instruction, choose an unsuitable option or act on misleading information from a webpage, email or document. Security guidance for agentic AI warns that prompt injection and unpredictable behaviour are reasons to combine careful instructions with technical controls. “I told it not to” is not a substitute for limiting its permissions.

A spending cap helps, but it does not prove that a payment is sensible. A payment can be within budget and still go to the wrong person, use the wrong account or pay for the wrong service.

Safeguards worth looking for

• Use an integration supported by your bank or a regulated provider; never paste online-banking passwords, one-time codes or recovery details into an AI chat.
• Start with read-only access. Add payment authority only if the service and your bank explicitly support it.
• Require your own approval before a payment is sent. Check the recipient, amount, date and purpose on a trusted bank screen.
• Use narrow permissions, sensible limits, an expiry date and notifications for activity.
• Keep a clear activity record and know how to disconnect the agent and revoke its access.

If a provider cannot explain which account data it receives, what it can do with it and how you stop access, do not connect your account. Check the bank’s own guidance rather than assuming that a feature described by an AI company is available through your bank.

The sensible rule

AI may eventually make some financial tasks less tedious. Until a service is supported, clearly bounded and easy to supervise, treat payment authority as something a person keeps—not a convenience to switch on casually.

Read the Bank of England consultation on future retail payments infrastructure →

Read the National Cyber Security Centre’s guidance on managing agentic AI cyber risk →

Read IT Club’s article about giving an AI agent the company credit card →

Read IT Club’s guide to giving AI agents access to work systems →

Have a technology question? Ask IT Club for free.