The Return Counter Nobody Built

By Masumi Network | August 2026

Everyone spent 18 months building the checkout for AI agents. Nobody built the return counter.

Everybody spent 2025 and early 2026 solving the same problem: how do AI agents pay for things?

That’s the problem that got all the attention. x402. Agent Payments Protocol. Cloudflare Wallets. Mastercard’s AP4M. Visa’s Agentic Ready program. American Express’s developer kit. Pick your acronym. Every major payment network in the world spent the last eighteen months building checkout rails for autonomous agents.

They solved it. Mostly. Agents can now transact at machine speed, across card networks, bank accounts, and stablecoins, with spending caps and cryptographic credentials and delegated authority from their human principals.

What nobody solved is what happens next.

When an AI agent buys the wrong thing, the fraud system flags it as a bot attack. When the purchase goes through anyway, the human who owns the agent disputes it as unauthorized. And when everyone ends up in a room together — the merchant, the card network, the AI provider, the consumer — nobody can find a legal framework that says who pays.

This is where agentic commerce actually is in August 2026. The checkout works. The return counter doesn’t exist.

The False Decline Problem

Fraud detection systems were built to stop malicious bots. AI shopping agents look exactly like malicious bots. Same behavioral fingerprint: no mouse movement, no dwell time, no scroll pattern, just a direct hit to the checkout API.

Chargebacks911 flagged this in May 2026. They found false declines cost merchants roughly thirteen times more than actual fraud, and only sixty-four percent of merchants even track the rate. Agentic AI traffic rose nearly eight thousand percent year-over-year in 2025, according to Human Security’s State of AI Traffic benchmark. Most fraud systems haven’t adapted. Legitimate purchases are being rejected at scale, silently, with no recourse.

The Chargeback Nobody Planned For

The chargeback problem is newer and stranger. Consumers are now disputing purchases their own AI agents made. The psychological mechanism is simple: if you didn’t click “buy,” the purchase doesn’t feel authorized, even if your agent was acting on explicit delegation.

No regulation clarifies whether that constitutes an unauthorized transaction. The Electronic Fund Transfer Act assumes a human initiated the transaction. The Fair Credit Billing Act assumes a human disputes it. Neither covers an autonomous agent acting on delegated authority without moment-of-purchase confirmation.

American Express made the first move. In April 2026 they committed to covering erroneous purchases made by registered AI agents on their network, as part of their ACE developer program. It’s a sensible policy. It also shifts enormous liability to Amex and does nothing for Visa, Mastercard, or bank-to-bank rails.

The First Legislative Attempt

Senator Warner’s AI AGENT Act discussion draft, released June 29, 2026, takes a different approach. It creates a federal registry of “Custodial User Agents” with fiduciary-like duties to their principals — they cannot manage a user’s online interactions in a way that benefits the agent provider at the user’s detriment. If your agent is registered and acts within its authorized scope, the liability chain becomes traceable. If it isn’t, you’re in the same fog everyone is in now.

This Is Not a New Pattern

None of this is surprising in hindsight. Every new payment rail in history solved the transaction first and the dispute second. Credit cards launched in 1950. The Fair Credit Billing Act passed in 1974. Twenty-four years of chargebacks with no consumer protection.

The gap this time is the same gap. We built the road. We’ll argue about guardrails for a decade.

The practical implication for businesses building on agentic commerce today is straightforward: log everything. Every permission granted to an agent, every action the agent took, every moment the authorization chain can be traced. Not because regulators require it yet. Because when the dispute arrives — and it will arrive — the merchant with the cleanest audit trail wins.

Sources

• Chargebacks911 warns AI agents are creating a new era of dispute risk
• AI Shopping Agents Trigger False Decline Crisis — The Fintech Times
• Senator Warner makes a first foray into agentic AI regulation — TechPolicy Press
• DLA Piper: Senator Warner Discussion Draft — Top Points
• When AI agents make incorrect purchases, who’s responsible? — The Financial Brand
• AI Agent Chargeback Liability — Chargeflow
• Agentic Commerce is Coming, but Regulation Meant for Humans Will Slow It Down — Center for Data Innovation