Agentic CommerceWho’s Responsible When AI Gets it Wrong? Who Covers the Chargeback Costs?

Agentic Commerce is Coming. But Who Owns the Consequences?

Earlier this year, Visa announced plans to integrate its payment network into AI-driven shopping, enabling users to make purchases through ChatGPT-style assistants. 

Headlines focused mostly on what the technology could do. Admittedly, it’s a big leap forward for the payments space. It also demonstrates how quickly AI agents are evolving. We’ve moved beyond AI systems that simply answer questions or generate recommendations. More and more, these tools are being positioned as digital assistants capable of acting on a consumer’s behalf. That includes making unsupervised purchases based on predefined parameters.

Agentic Commerce Offers Clear Benefits & a Value Proposition

The value proposition for consumers is obvious. Forget spending thirty minutes comparing products, evaluating reviews, and navigating checkout pages. Users can simply describe what they want and let an AI agent handle the rest.

For the payments industry, it means more commerce flowing through the ecosystem. That benefits everyone: networks, issuers, and merchants, as well as fintech providers building fraud prevention and risk management solutions.

OpenAI and other brands in the AI space have been moving in this direction since day one. Agentic commerce has been a primary aim for a long time. So, now that it’s here, the real story isn’t that OpenAI is positioning AI agents to make purchases. What’s notable is that now Visa’s involved in building the supporting infrastructure.

Payment card networks exist, in part, to establish trust, standardize transactions, and assign liability. They also maintain the systems used to resolve disputes when transactions go wrong. But, if we’re moving toward allowing computers to make autonomous purchases, responsibility issues stop being theoretical concerns.

The conversation is focused on convenience, efficiency, and adoption. A much bigger question is possibly being overlooked, though: what happens when the system doesn’t work?

What happens when an AI agent evaluates the available information, follows the given criteria, and reaches a conclusion that is technically correct — but still wrong? And, who is liable for it? It’s a simple question, but it may expose a liability gap that the payments ecosystem isn’t currently equipped to address.

The Chargeback System Wasn’t Designed for This

The chargeback system was created to address a fairly predictable set of problems. Unauthorized purchases made with stolen credentials; merchants failing to deliver what was promised; operational errors like unprocessed refunds or mishandled transactions.

In each case, the process is generally the same. You determine where the transaction broke down and assign responsibility accordingly. Agentic commerce complicates that equation by inserting another link into the chain.

Here’s a scenario: a consumer asks an AI assistant to purchase a television based on a few criteria, such as size and budget. The AI evaluates multiple options, selects one, and completes the transaction.

The merchant ships exactly what was ordered. The product arrives on time. The transaction was authorized, and the description was accurate. But a week later, the customer decides that AI made the wrong choice.

Technically, nothing went wrong. The merchant fulfilled the order correctly, the issuer made no error, and no fraud occurred. The transaction worked exactly as intended. But you still have an unhappy customer who doesn’t feel fully responsible because the purchasing decision was delegated rather than made directly. So, the customer decides to dispute the transaction, arguing that they should not be liable for an unsatisfactory purchase that was not first-hand authorized.

Does the customer have a valid claim here? The ambiguity is the problem.

The Rise of Delegated Authority

For decades, liability decisions have largely centered on a familiar set of questions. Did the cardholder authorize the transaction? Did the merchant fulfill its obligations? Did anybody break the rules? Answer those questions, and assigning liability is usually pretty obvious. 

Once consumers start handing off purchase decisions to software, though, you’re adding another query: Did the agent act within its given authority?

At first glance, that might not seem to change the game much. After all, a person still asked for the purchase. What’s different is that AI is the one making the key decisions about what actually gets bought. 

That complicates the traditional liability model. Now, the real question isn’t fraud or merchant failure. Rather, it’s whether the agent interpreted the request the way the consumer expected. 

Maybe a spending limit was respected, but the model wasn’t what the customer expected. Maybe the AI leaned too heavily on price or ratings and missed something the customer would have prioritized. Or, maybe it technically followed the instructions, but the outcome simply doesn’t feel right.

None of these situations fit neatly into the usual buckets like unauthorized transactions, merchant error, or non-delivery. They all stem from decisions made on someone else’s behalf.

In each case, everyone involved can make a pretty reasonable argument that nothing actually went wrong.

The Next Liability Battle

Our current dispute system was developed in the 1970s; long before autonomous purchasing was even a consideration. Chargeback reason codes weren’t designed for a world where transactions can be authorized and correctly fulfilled, but still end up being disputed.

Agentic commerce introduces that exact scenario. Who’s responsible when an AI agent misunderstands instructions, oversteps its authority, or simply makes a poor purchasing decision? Does liability stay entirely with the consumer who authorized the agent, or will merchants have to absorb these losses, despite fulfilling the order correctly? Alternatively, is there a point where AI providers become part of the accountability chain? None of that has been clearly defined yet.

Traditionally, payment networks have drawn those lines, but as AI agents begin directly participating in commerce, that will almost certainly need to change. By whom and how much is still up in the air.

The uncertainty shows up on the seller side, too. Even if new rules eventually emerge, merchants will still need ways to defend themselves against disputes tied to agent-driven decisions. Ironically, this is where agentic commerce may actually help.

AI-initiated transactions could, by default, produce far richer records than we’ve ever had before. They can capture customer instructions, purchase constraints, approval history, and even the logic behind AI’s decision. That level of detail could become a powerful evidentiary tool in transaction representment. It could also raise the bar for what constitutes compelling evidence in the first place.

Of course, that’s assuming the existing chargeback system will work at all for agentic commerce. We may find that the payments space needs to overhaul the chargeback process from the ground up.

Can Agentic Commerce Help Clarify Dispute Liability?

At the same time all these questions are being posed, I should also point out here that the technology is uniquely positioned to meaningfully improve how payments work, if leveraged properly.

For example, one of the biggest challenges in dispute management is the lack of transaction context. When a cardholder disputes a purchase, the transaction record typically shows little more than merchant name, timestamp, and amount. It tells the merchant almost nothing about how the decision was made.

AI-assisted transactions could change that. Data-rich records could include more signals as to customer intent. Purchase instructions, evaluated options, recommendation logic, and approval history. That added context could help reduce disputes stemming from forgotten purchases, unclear descriptors, and so on. It could also give merchants stronger evidence when responding to invalid claims.

I’m not dismissing any of these benefits at all. What I’m saying is that they’re only part of the story. And, they need to be rolled out in a measured manner to make sure we’re covering critical points of liability.

The Time to Develop a New Framework is Before Agentic AI Rolls Out

The Visa/OpenAI announcement is more than a product update. It’s also an early signal of a shift toward a new category of liability; one where intent, delegation, and responsibility aren’t cleanly aligned.

The payments industry is already building the infrastructure for agentic commerce. The governance layer, however, isn’t keeping up. We’re developing systems to enable AI-driven purchasing at scale, but there’s still no clear standards for who owns responsibility when delegated decisions go wrong.

That gap won’t stay theoretical for long. We need a realistic framework for how liability should be assigned in agent-driven transactions. Unfortunately, that’s going to require coordination across networks, issuers, merchants, fintech providers, regulators, and AI developers. 

It’s likely to be one big logistical labyrinth, which is why we need to get started now. We can’t wait until AI agents are routinely making purchasing decisions at scale. The payments industry needs to move now to clarify liability, while we have the luxury of creating a new framework.