Agent loops, chargeback blind spots, and the 1,000 eggs problem: agentic commerce's failure modes

Everyone's talking about agentic commerce as the next trillion-dollar shift in how people buy things. But in a series of interviews we recorded with senior payments and merchant leaders, the conversation shifted away from the hype and into the mechanics of agentic commerce - and its potential pitfalls.

Volodymyr Sytnikov, CEO and Managing Partner at Brainstack, gave an example of an agent stuck negotiating with another agent over a purchase. Round and round it goes, looping through the same exchange with no resolution. "You just spent like 300 bucks on tokens on the best negotiation things," he said, and you still haven't bought your bag.

That's the exact kind of operational risk that payments teams need to plan for now - not when agentic commerce has become more widely adopted. Here are five failure modes surfaced in interviews with payments and merchant leaders, and what they mean for the infrastructure behind agentic commerce.

Sytnikov's story highlights the issue of bots communicating with one another: two agents negotiating a purchase can get stuck in a cycle, each one re-evaluating the other's terms without ever converging. The transaction never completes, but the token spend does. For merchants and platforms, that's a new kind of cost that has nothing to do with fraud or chargebacks. It's just an agent burning compute on its own indecision. Left unmanaged, agent loops are a resource drain with no offsetting purchase at the end of it.

Cristiano Betta, CEO and Founder of Gr4vy, explains the issue with bots taking instructions literally with no guardrails in place to prevent purchases: "Just because somebody is on their website, in their chatbot asking to buy a thousand eggs doesn't mean they probably want to buy a thousand eggs." An agent given a loose instruction will execute it literally. A human shopper who accidentally enters the wrong amount in a quantity field catches the mistake before checkout. An agent doesn't, unless the merchant has built in the guardrails to catch it first. That's the gap: agents follow instructions, not intent, and merchants need controls that sit between the two.

Daniel Kornitzer, Head of Global Partnerships at Ebanx, raised a different problem. In agentic commerce, there's no human at the checkout screen to catch something suspicious. It's an agent talking directly to a database. If that agent has been hijacked, or was fraudulent from the start, the transaction still goes through, and it still ends up as a chargeback the merchant doesn't want to honour. Kornitzer's question cuts to the point: when a consumer says an agent made a purchase they didn't authorise, who takes the loss? The merchant? The human who owns the agent? He called it "a very valid question", and it doesn't have a settled answer yet.

Rita Narkevičienė, Payments Operations Manager at Kiloverse, raised the issue of chargeback liability: "No one is talking about how chargebacks, when AI is purchasing you the item, are going to be treated." Existing chargeback frameworks assume a human weighed up a purchase decision and can be held to it. When an agent makes that decision on a human's behalf, the framework doesn't have a clean answer for who bears the risk, or whether customers will use "my agent did it" as a new route to fraudulent disputes. Narkevičienė's worry is straightforward: if that question stays open, merchants lose either way.

Andre Moeller, Head of Payments, Risk & Fraud at Volkswagen Group Charging, pushed the problem one step further. Say an agent orders something on a human's behalf and gets it wrong: the wrong product, or the right product sent to the wrong address. Whose risk is that, the agent's or the person who briefed it? Now stretch the chain: agent A can't find what it needs, so it hands off to agent B, which hands off to agent C. Something goes wrong three steps down the chain. Moeller was candid that nobody has the answer yet, but he was clear on the direction: liability is shifting, and the industry needs to get ahead of it rather than discover the answer after the fact.

Each of these failure modes points to a specific structural gap, and payments infrastructure needs to close it before agentic volume makes it expensive.

Start with authentication. Andrzej Tomaszewski, Head of Business Products at Zing Coach, framed it as a new category alongside customer-initiated and merchant-initiated transactions: agent-initiated transactions need their own verification path, something like a "know your agent" (KYA) process sitting next to know your customer (KYC).

Without it, he warned, merchants risk becoming the industry's unpaid beta testers while payment service providers and orchestrators figure out the standards later. That KYA layer is the direct answer to the fraudulent agent problem: verify what's transacting before the money moves, not after.

The 1,000 eggs problem is a merchant-side fix: order-level controls and anomaly detection that catch agentic orders that fall outside sane bounds, using the same principles that already flag a sudden spending spike on a card. Here it's applied to agent behaviour and quantities.

The chargeback blind spot and the liability chain both come down to the same open question, and it's still being worked on rather than solved. Marco Conte, VP Product AI Data and Insights at Ixopay, was direct about where things stand - liability rules and chargeback management processes both need to change, and the protocols for doing that are still being drafted. Real-time fraud detection built for agent behaviour, and dispute processes that can actually attribute responsibility across a chain of agents, are the two pieces that need to exist before agentic volume reaches the level these executives expect.

These five stories are from interviews we've recorded with payments and merchant leaders. For the primary data behind them, our white paper "Before They Check Out" combines a survey of more than 1,750 European consumers with in-depth interviews from 16 merchant and industry executives on what payments infrastructure needs to look like next.