As AI agents begin making purchases for customers, the financial crime controls built around human behaviour may be looking in the wrong place. According to Vivox AI, a transaction can sit comfortably within every spending limit and still stem from a manipulated decision made long before checkout.
Vivox AI’s analysis draws on a panel held at QUBE Events’ sixth Financial Innovation Forum: Payments & RegTech. Bank of London deputy director of financial crime Zainab Shode moderated the session. She was joined by Manish Kumar, former head of payment acceptance products at Starling Bank; Mastercard vice president of product management Priyanshi Mathur; PXP group CEO Kamran Hedjri; and CGI vice president of payments solutions Ainsley Ward.
Consider a simple instruction to book a hotel. An agent might pick a convincing fake website and pay a fraudulent merchant with the customer’s full authority. In a more extreme scenario, it could cancel another guest’s booking to secure the room. In both cases, Vivox AI notes, the final payment reveals only part of the story.
The panel framed consent as a set of specific, checkable permissions rather than a single approval. Payment providers need to know how much an agent may spend, which merchants or categories it may use, when its authority expires and how a customer can amend or withdraw it. These rules will vary between cards, digital wallets and account-to-account transfers, each of which handles authorisation, disputes and liability differently.
Identity also becomes more complex. Firms must confirm not only who the customer is, but which agent is acting and whether it is the one the customer authorised. Vivox AI highlights the value of records that link customer identity, agent identity and the scope of authority, allowing investigators to reconstruct what was requested, what the agent encountered and what it did. Revoking an agent’s access in an app is only meaningful if that change is recognised across the entire payment journey.
Fraud may now happen upstream. Panellists pointed to fake merchant sites, fabricated catalogues and malicious instructions hidden in content agents read, as well as interference when permissions are first set. Once attackers find a weakness, automated agents could exploit it repeatedly and at speed.
Liability remains unresolved. Participants disagreed on how responsibility should be shared between customers and the firms supplying agents, payments and safeguards, and on how existing reimbursement rules would apply. What they agreed on was the need for evidence detailed enough to pinpoint where a failure occurred.
For Vivox AI, the takeaway is clear: firms should test now whether their fraud, money laundering and sanctions controls can separate legitimate automated activity from misuse, rather than waiting for a finished monitoring model.
For more, read the full report here.
Copyright © 2026 FinTech Global
