What Millions of Dollars in Transaction Data Reveal About Agentic Commerce
A data driven analysis of early agentic purchasing behavior.
Much of the discussion around AI-powered purchasing has focused on what could go wrong: unintended purchases, rising fraud, and loss of consumer control. But until now, there has been little real-world evidence to evaluate those concerns.
Cross River analyzed nearly $50 million in agentic purchase volume on its platform over a six-month period. Examining returns, fraud indicators, and authorization outcomes, we set out to answer a question the industry keeps asking: do AI agents make bad purchases?
The data suggests they do not.
Four findings emerged from the data:
- No evidence of systematic unwanted purchases
- Return rates nearly identical to traditional commerce (3.9% vs. 3.4%)
- Fraud indicators comparable to those of human-driven transactions
- Existing card, risk, and authorization controls blocked a substantial share of problematic attempts, including more than 1,000 suspected fraudulent transactions
Across nearly $50 million in transaction volume, agentic purchases behaved remarkably similar to traditional consumer purchases. Whether measured through returns, fraud patterns, or authorization outcomes, the results were consistently aligned with traditional commerce.
Return data indicates strong intent alignment
One of the clearest ways to evaluate whether purchases align with customer intent is to look at returns. If agents were routinely misunderstanding intent, we would expect significantly elevated return rates.
Instead, our data found consumers kept more than 96% of the purchases made through agents.
Travel accounted for a disproportionate share of agentic returns—32.5% versus 25.8% for traditional commerce. Rather than signaling poor purchase decisions, however, the pattern may reflect one of the earliest and strongest use cases for agents: helping consumers compare a broad set of options and optimize travel purchases. Given that travel bookings are frequently modified or canceled regardless of how they are purchased, elevated return activity is not necessarily evidence of agent misalignment.
Existing controls are working
The data also showed that financial controls remain highly effective in an agentic environment.
Excluding expired cards, authorization attempts were declined at the same rate of traditional transactions (12.5% vs. 12.2%) because of existing safeguards, including spending limits, account restrictions, and fraud detection measures.
Rather than bypassing existing controls, agents operate within the same authorization, risk, and compliance frameworks that govern human purchases.
Importantly, fraud indicators remained consistent with traditional commerce as volume increased, suggesting agentic transactions did not create a distinct fraud profile.
Agent capability is not the binding constraint. Infrastructure is.
The data suggests that the risks most frequently associated with agentic commerce are not currently the limiting factor. Existing authorization, fraud, and compliance controls continue to function as intended. The next challenge is not transaction safety but rather building the infrastructure required to support agentic commerce at scale.
As adoption grows, identity, delegated authority, and policy-based controls will become foundational infrastructure. The same logic, risk models, and audit trails that govern human transactions must govern agent-initiated transactions. As my colleagues have said time and time again, modular architecture is what turns early volume into durable volume.
The story isn't about rogue agents. It's about agent-ready infrastructure.



