Agentic checkout
Agentic checkout is a purchase completed by an AI agent on a buyer's behalf, with the agent handling cart construction and payment authorization while the merchant remains the merchant of record. It is implemented through protocols such as ACP and UCP and through platform-specific integrations, and it depends on a product record accurate enough that the agent's assumptions about price, unit, and availability hold at the moment of the order.
The flow
A buyer states an intent. The agent resolves it to specific products, assembles a cart, confirms details with the buyer, and submits payment using a delegated credential rather than a raw card number. The merchant fulfils, invoices, and owns the customer relationship and any return.
The important structural point is that the agent is a purchasing intermediary, not a reseller. This is what distinguishes agentic checkout from a marketplace: you keep the merchant-of-record position, the margin structure, and the obligation.
What has to be true in your data
Agentic checkout is unusually intolerant of the data problems catalogs tolerate every day, because there is no human at the last step to notice something looks wrong.
- Stable identifiers. OpenAI's product feed specification requires product and variant identifiers to remain stable over time. An identifier that changes between the agent's read and the order is a failed transaction.
- Unambiguous unit of measure. If the price is per case and the quantity is per each and nothing in the record says which, the agent computes the wrong total and orders the wrong amount. This is the single most common agentic failure mode in wholesale data.
- Correct pack and kit structure. A kit SKU that inherits one component's attributes will be selected for the wrong requirement and shipped as the wrong thing.
- Availability the merchant can honour. A national stock figure quoted against a job shipping from one branch makes a promise that branch cannot keep.
Where B2B does not fit yet
Current agentic checkout implementations assume a consumer-shaped transaction, and B2B departs from that in almost every dimension: contract pricing that is not published, payment on terms rather than card, quotes rather than orders, freight quoting on oversized or hazmat items, tax exemption certificates, PO numbers, approval thresholds, and job-site delivery.
The realistic near-term path is not a shopper asking a public assistant to buy 40 high bays. It is a procurement agent operating inside the buyer's own environment, authenticated against a contract, doing what PunchOut does today with fewer clicks. Suppliers whose catalogs answer structured queries completely will be the ones those agents can transact with.
Meanwhile the retrieval side of the problem is live now and worth more. Being the product an agent recommends precedes being the product it buys, and there is no version of the future where a thin record wins either.
Frequently asked questions
What is agentic checkout?
A purchase completed by an AI agent on a buyer's behalf. The agent builds the cart and authorizes payment with a delegated credential, while the merchant remains the merchant of record and handles fulfilment, invoicing, and returns.
Does agentic checkout make the AI company the seller?
No, under the current protocol designs. The agent facilitates the transaction and the merchant stays the merchant of record. That structure is deliberate, since it keeps the customer relationship and the fulfilment obligation with the merchant rather than converting them into a marketplace supplier.
What breaks agentic checkout in a B2B catalog?
Ambiguous unit of measure is the biggest one: price per case against quantity per each produces the wrong order with no human to catch it. Unstable identifiers, kit SKUs carrying the wrong inherited attributes, and national stock figures no branch can honour follow closely.
Should distributors implement agentic checkout now?
For most, retrieval comes first. Contract pricing, terms, quoting, freight, and approvals are not well served by current implementations, while the ability to be found and correctly recommended is available today and is a prerequisite for every later step.