AI 101 — What is agentic commerce?
Agentic commerce is shopping in which an AI agent does the browsing, the comparing and — the part that makes it commerce — the paying, inside limits you set in advance.
That definition sounds like a small extension of a shopping chatbot. It is not, and the difference is the last two steps: an agent that hands you a shortlist has done research; an agent that completes checkout with your money has done commerce. Everything hard about the category sits in that gap — who authorised the purchase, who holds the card details, who carries the loss when the order is wrong, and who gets paid for the referral.
Why it matters right now
Two of the largest consumer surfaces on earth shipped the missing pieces within four months of each other.
OpenAI launched ChatGPT Wallet on Friday for US ChatGPT Pro subscribers: add a card once, then its "dots" agent can find, fill and pay for an order, with every purchase approved by Face ID or fingerprint and paid with a network token rather than your card number, per OpenAI's announcement. We covered the launch and its zero take rate in ChatGPT Wallet launches: agents pay, OpenAI takes no cut.
Google went the standards route. On January 11 it announced the Universal Commerce Protocol, an open standard co-developed with Shopify, Etsy, Wayfair, Target and Walmart, which lets an agent run a merchant's real checkout — discount codes, loyalty credentials, subscription cadence, final-sale terms — from inside AI Mode in Search or the Gemini app, per Google's announcement. The merchant stays the seller of record.
Before either, in September 2025, OpenAI and Stripe published the Agentic Commerce Protocol, an open standard for the same conversation between an agent and a store, per Stripe. Three arrivals in fourteen months is the tell: this stopped being a demo and became plumbing.
The mental model: four jobs, not one
A purchase is four separate jobs, and it helps to keep them apart because each has its own protocol and its own failure mode.
Discovery is finding products that fit — the job chatbots already do well, reading a catalog instead of a page. Decision is weighing price, reviews, stock and delivery against what you actually said. Checkout is the transaction itself: cart, tax, shipping, terms. Payment authority is proof that a human allowed this specific agent to spend this specific amount.
A shopping assistant does the first two and links you out. Agentic commerce does all four. That is why the industry needed a stack rather than a feature: the store conversation and the payment authorisation are different problems, solved by different companies, and neither works if the merchant can't tell a legitimate agent from a scraper.
The analogy: the company card
Think of a new hire with a corporate card and a written spend policy. The card is not the new hire's own money; it carries a limit, a category, sometimes a time window. Finance can see every charge. If the hire buys the wrong thing, the policy — not the card — decides who eats it.
That is the shape of an agentic purchase. You set the limits (budget, merchant, delivery date, sometimes a category), the agent transacts inside them, and a scoped credential stands in for the card. Google's Agent Payments Protocol, announced in September 2025, is exactly this idea formalised: the human signs a mandate describing what the agent may buy, and the merchant checks the mandate rather than trusting the agent's word.
The analogy also shows where it breaks. A company card is issued to a person who can be fired and sued. An agent is not a legal person, so the liability question has no clean answer yet — which is why most launches ship with a biometric approval on every purchase.
Common misconceptions
"It's just a shopping chatbot with a checkout button." Authority is the dividing line, not the interface. Approving every purchase with your face is a checkout shortcut, not autonomy — useful, but a long way from "reorder the printer paper under forty dollars when it runs out."
"The agent holds your card number." Good implementations never expose one. The merchant receives a payment token scoped to that transaction, the way Apple Pay has worked since 2014, and card rewards and dispute rights are meant to survive. Whether they do — the fine print says "eligible" disputes — is the open question.
"It's one protocol." It is a stack. The store conversation runs on ACP or UCP; payment authority runs on AP2-style mandates; agent identity runs on signed requests such as Web Bot Auth, which OpenAI's merchant documentation asks stores to accept; the general plumbing for connecting an agent to tools is MCP. They overlap less than the marketing suggests.
"Merchants are delighted." Many are not. Amazon won a court order against Perplexity's shopping agent in March and still blocks agent purchases, and United said this month it will block third-party agents from booking and managing reservations, citing credential sharing and the rule that a human must attest at check-in. The reasons are rational: an agent that compares twelve stores compresses the shelf space a retailer paid to occupy, and a bad agent purchase turns into a chargeback the merchant eats first.
"Price comparison is the whole value." For a consumer, the interesting part is the constraint, not the discount — "under one hundred and fifty dollars, delivered by Friday" is a search no human performs well across forty tabs. For a merchant, the interesting part is that being machine-readable becomes a distribution channel: a product feed with current stock now does the work a landing page used to.
Where to learn more
The specifications are public and readable without an engineering background in the interesting parts: the Agentic Commerce Protocol's site documents how a merchant exposes a feed and accepts an agent checkout, Google's developer write-up of the Universal Commerce Protocol walks the same flow from the merchant's side, and OpenAI publishes a commerce playbook aimed at retailers rather than developers.
Should an agent's mistaken purchase be the buyer's problem, the platform's, or the store's? Write to us via our contact page.
Related reading: AI 101 — What is an AI agent? · AI 101 — What is MCP?
