The Checkout Problem: Agentic Commerce Needs a Chain of Authorization
Photo: autonomous shopping agent completing checkout on smartphone with credit card, via www.wikihow.com
Autonomous shopping agents have moved from novelty demos to checkout flows. In 2026, a software agent can compare prices, apply coupons, choose a delivery window, and complete a purchase with a stored payment credential. The technical demonstration is no longer the hard part. The hard part is the receipt.
Agentic commerce asks merchants, banks, card networks, and AI vendors to treat a software process as a purchasing principal. Existing rules were built for a human cardholder who taps, clicks, or signs. They assume a person can be authenticated, authorized, and held accountable. An agent complicates all three.
A demo that cannot pass underwriting
A shopping agent can look impressive in a controlled demo and still fail a risk review. Underwriting asks who is liable if the agent buys the wrong item, duplicates an order, triggers a subscription renewal, or falls for a prompt injection that redirects a payment. It asks how a merchant can distinguish a legitimate agent from a credential-stuffing bot. It asks what evidence a bank needs when a customer disputes a charge.
Those questions do not have clean answers under current US card-network rules and consumer-protection frameworks. The rules are not silent on fraud, but they are not designed for delegated purchasing. A card-not-present transaction already carries risk. An agent-initiated transaction adds a new layer: the cardholder may have authorized the agent in general but not the specific purchase.
Authorization is not authentication
Authentication asks whether the person is who they claim to be. Authorization asks what the agent may do. Most agentic commerce pilots focus on the first and underinvest in the second. A stored card credential can prove that a payment method exists. It cannot prove that the agent had permission to spend two hundred dollars on a specific item at a specific merchant.
A workable authorization model needs scoped credentials. The agent should carry a mandate that specifies merchant category, spend ceiling, time window, item quantity, and return policy. It should distinguish one-time purchases from recurring ones. It should record the user instruction that produced the transaction. Without those constraints, every agent purchase looks like a card-not-present transaction with an unusually vague cardholder.
Who bears the loss when the agent errs
Consider four common scenarios. In the first, a user tells an agent to buy a replacement filter and the agent orders the wrong size. In the second, a merchant lists a product deceptively and the agent completes the purchase. In the third, a prompt injection hidden in a product review redirects the agent to a fraudulent seller. In the fourth, an agent renews a subscription the user meant to cancel.
Consumers will expect protections that resemble existing dispute rights. Merchants will see transactions that were technically authorized by a credential. Banks will see disputes that do not fit existing chargeback reason codes. AI vendors will point to their terms of service. The result is a liability gap that no party wants to own.
US regulators have signaled interest in automated decision systems and unfair or deceptive practices, but agentic payments sit at the intersection of payments law, contract law, and agency law. The FTC Act gives enforcers tools against unfair or deceptive conduct. State consumer-protection statutes add another layer. Card-network rules and issuer agreements govern much of the operational reality. None of these frameworks was written with autonomous purchasing agents in mind.
Receipts as infrastructure
The most useful near-term innovation may be boring: a machine-readable receipt. An agent-initiated transaction should produce a structured record that identifies the principal, the agent, the instruction, the spend limit, the merchant, the item, the price, the consent event, and the return window. That record should travel with the transaction and be available to the issuer, merchant, and customer in a dispute.
A receipt is not just a log. It is an accountability artifact. It lets a merchant prove that the listing matched the agent instruction. It lets an issuer evaluate whether the customer authorized the agent scope. It lets the customer correct a preference without canceling the payment method. It also creates privacy obligations: the receipt should not become a surveillance dossier.
What merchants, banks, and vendors should settle in 2026
Merchants need clear terms for agent-initiated orders, including cancellation, returns, and liability caps. Banks need dispute workflows that can ingest agent logs and evaluate scoped authorization. Card networks need pilot rules that distinguish agent traffic from bot fraud. AI vendors need to expose agent identity, permission scope, and audit trails to their customers.
The commercial questions are as important as the legal ones. Who pays for a wrong purchase. Who pays for a fraudulent merchant. Who pays for a prompt-injection loss. The answers will vary by contract, but they should be written down before scale, not litigated after an incident.
Agentic commerce will not scale on better browsing. It will scale on credentials, limits, logs, and dispute rules. The agent economy needs a receipt because trust at checkout depends on more than a convincing demo. It depends on knowing who authorized what, and who pays when the answer is wrong.