FIS: banks must verify which AI agent authorized the transaction

Payments processor FIS Global has become the first major processor to publicly frame the agent authorization question, according to Forkast News. When an AI agent buys something on a user's behalf, FIS argues, banks must be able to determine which agent is using a credential, what that agent was authorized to do, and whether the transaction matches the original mandate.

The positioning matters because the processor layer sits between merchants and issuers, seeing a transaction from both sides. Troy Mann, SVP of Product Management at FIS Global, noted that while agent-driven payment volume is still small, consumers with stored payment credentials are already entering an ecosystem rebuilt for agentic transactions.

The core problem is that the financial system lacks a universal way to authenticate machine authority. When a human swipes a card, intent is usually clear; when an agent acts, that intent is often buried in code or hidden behind layers of automation. As Mann put it, intent has to be expressed and captured.

The industry has spent years chasing frictionless payments, but FIS is suggesting a rethink of that push for total speed. Mann argues that a little bit of friction is not a bad thing when spending is machine-led.

Instead of letting agents use standard credentials, FIS proposes virtual cards designated specifically for agentic commerce, carrying built-in restrictions such as limits on transaction amounts, specific merchants, or defined time windows.

The hard part is data fragmentation. Signals about intent, authorization and the transactions themselves are scattered across merchants, agent platforms, networks and issuers. Tokenization is a major layer of the security architecture, but Mann points out that it solves only part of the problem; without a unified way to pass intent data through the payment lifecycle, processors are essentially flying blind.

FIS is working with financial institutions to identify agentic transactions, apply differentiated fraud scoring and incorporate new signals across the payment lifecycle, following a crawl, walk, run approach in which humans remain involved in establishing intent and authorization.

There is a practical caveat to this vision. Even if processors build these authorization frameworks, the system depends heavily on the quality of the data supplied by the agent platforms themselves. If an agent cannot clearly communicate its mandate to the processor, or a merchant's system fails to pass that intent signal forward, even sophisticated virtual card controls will struggle to prevent unauthorized or erroneous spending.