Agentic commerce isn’t some future scenario—it’s already happening. Merchants offering agentic shopping experiences, and processors are facilitating payments through bots. What remains unclear is whether a shopping bot has been used to commit fraud, But given how quickly agents are evolving, it’s a question of when, not if.
The problem is that banks, in particular, are not prepared. Many don’t yet understand how to combat or even detect agents. A new report from Javelin, Good Bot, Bad Bot: How Agentic AI Changes Fraud Detection, looks at the risks ahead and what’s being done to anticipate them.
“When AI came out, we focused solely on the analog and digital fraud is that was already in front of us,” said Jennifer Pitt, Senior Analyst of Fraud Management at Javelin Strategy & Research. “That’s exactly what’s happening now. Banks in the U.S. are ignoring agentic fraud, and when that type of fraud is here, they’re going to be reactive.”
Finding Agentic Fraud
Most people have heard the story of how an OpenAI agent broke out of its sandbox and gained access to something it shouldn’t have. In other words, there are already examples of how these agents can go rogue. Combine that with the potential for fraud, and it’s an issue financial institutions need to address.
Part of the challenge is that agentic fraud isn’t yet being categorized as its own type of crime. The Internet Crime Complaint Center, the FBI’s cybercrime reporting center, generally reports crimes related to AI. That means consumers are largely responsible for identifying and reporting scams or fraud involving agents—but they may not now an agent was involved.
If someone is working with a legitimate, authorized agent but the payment method is compromised, for example, they may view it simply as an account takeover rather than fraud involving an agent.
Several vendors now claim they can detect agentic fraud before it happens, but the industry as a whole is still figuring out what that threat looks like. Some proactive vendors have already developed solutions designed to identify it. The challenge will be getting financial institutions to adopt those tools.
Assigning Responsibility for Mistakes
One of the first steps for financial institutions will be to clearly define what constitutes an agent and establish where liability lies. There is already controversy over who should be responsible for unauthorized purchases.
“Right now, if an agent is used for any of purchases at Target, then essentially it’s considered a valid purchase,” Pitt said. “I can definitely see pushback on that and lawsuits coming along. We need to explain the liability ahead of time and explain how agents can be manipulated or taken over so that people can watch their own access, much like we tell them to watch their own accounts for changing account profile information and things like that. Just being transparent is the biggest thing.”
A major concern is first-party fraud. Say a consumer tells an agent they want a red sweater, but the agent orders a maroon one. Who is responsible—the consumer or the agent? Assuming the customer is at fault could not only alienate consumers but also lead to more chargebacks.
“We can’t say that everything is the consumer’s responsibility,” said Pitt. “We’re going to have to do investigation like we do in other cases of fraud. We’re going to have to look at intent and behavioral signals, just like we do with traditional first party fraud.”
Bots Are Becoming More Human
Another concern is that agents will increasingly behave like humans. The differences may come down to subtle nuances that traditional fraud technology doesn’t catch, such as how and when an account is accessed.
These agents may not look fully human, but they also won’t necessarily behave like traditional bots carrying out large numbers of transactions at extremely high speeds.
If bots can act—and even type—like humans, how can FIs distinguish legitimate agents from malicious ones? That’s where context and intent become critical, along with continuous authentication and the assessment of risk signals.
“If we take agents out of the picture, current bot detection with financial institutions says ‘bot bad,’ right?” said Pitt. “The bot is bad, automation is bad. But the sophisticated bots are able to manipulate behavior where they can make it look more human.”
Banks can layer agentic-specific solutions on top of the fraud detection systems they already have, but traditional bot detection technology may not be able to distinguish between a legitimate agent and a malicious one. Agentic fraud will likely require continuous identity verification throughout the transaction lifecycle, much as financial institutions already do with other forms of fraud.
“The truth is that we can’t rely on one signal,” said Pitt. “We can’t just rely on behavior. We can’t just rely on typing speed or the initial verification of the agent and call it good. It’s going to have to be continuous. It’s going to have to be a lot of different signals and then look at the context and intent.”
Know Your Agent
Know Your Agent processes are already being developed, much like the Know Your Customer or Know Your Business processes that are widely used today. Technologies such as tokens and cryptographic keys can help verify that a particular agent is authorized to act on a customer’s behalf.
The good news is that agents can also be used to help identify fraud. Incorporating agents into fraud workflows could help financial institutions not only detect fraud but also manage costs by shifting resources to where they are most needed.
“Agentic commerce is here,” Pitt said. “Agentic fraud will be here very soon. We have to get ahead of it. Think about the possible solutions, get with your vendors.”
“We have to address it now or else we’re going to be very behind the curve,” she said. “This is going to come very fast, I think. We need to be prepared.”
