Agentic commerce: half a revolution (for now)
A year ago it looked like a done deal: we would ask ChatGPT to find us a pair of running shoes and it would buy them, payment included. In March, OpenAI backed off. Agentic commerce isn’t over; it has stopped halfway. AI agents increasingly decide what we buy, but the cart, for now, stays in the store.
What agentic commerce is
Agentic commerce is when the shopping is no longer done directly by a person, but by an AI agent working on their behalf. The agent works out what’s needed, searches across stores, compares prices, reviews and delivery times and, if authorized, buys and pays.
One basic thing changes. An online store’s customer is no longer just whoever browses the website: it’s also a piece of software that reads catalogs, prices and availability and compares them with competitors in seconds. Software isn’t won over by a nice homepage. It weighs the data.
Not every AI-assisted purchase is the same. It helps to separate three levels of delegation.
How much do we hand over to the agent?
- RecommendsThe AI searches, compares and suggests; the person buys, on the store’s websiteWidespread
- Buys with confirmationThe agent prepares the order and pays once the user says yes, without visiting the siteBeing tested
- Buys on its ownThe agent purchases autonomously, within rules and budgets set in advancePilot projects
Today almost everything happens at the first level. The past year’s battle was fought over the second. The third is still in the lab.
Who is writing the rules
For an agent to buy from a store it doesn’t know, it needs shared rules: how to read the catalog, how to build a cart, how to pay and how to prove the user agreed. Plenty have arrived over the past year, almost all from the big names in tech and payments.
| Protocol | Backed by | What it does |
|---|---|---|
| Agentic Commerce Protocol (ACP) | OpenAI and Stripe, September 2025 | Connects the assistant to the store’s checkout; built for purchases inside ChatGPT |
| Agent Payments Protocol (AP2) | Google, September 2025 | Proves what the user authorized, with digitally signed “mandates” |
| Universal Commerce Protocol (UCP) | Google with Shopify, Etsy, Wayfair, Target and Walmart, January 2026 | Lets agents, stores and payments talk to each other across the whole purchase |
| Visa Intelligent Commerce and Mastercard Agent Pay | The card networks, April 2025 | Give agents dedicated payment credentials, with limits and controls |
For now this is less a standard than a race to set one. For sellers the message is clear: it pays to be on platforms that adopt these protocols, without chasing each one.
Checkout in ChatGPT: what happened
In September 2025 OpenAI launched Instant Checkout. In the United States you could buy from Etsy sellers without leaving the chat, and the announcement promised to extend it to more than a million Shopify stores. Walmart joined soon after. It looked like the start of conversational commerce.
Six months later, in March 2026, OpenAI changed course. ChatGPT now focuses on product search and comparison; the purchase goes back to the store, on its website or in its own app inside ChatGPT. Walmart, for example, brought its own assistant, Sparky, into ChatGPT and Gemini.
The reasons, as reported, are very practical:
- Few live stores. By some estimates, only a few dozen Shopify merchants were actually live.
- Low conversion. According to Walmart, purchases inside ChatGPT converted at a third of the rate of its own website.
- Data that’s hard to standardize. Prices, variants, availability, shipping, taxes and returns differ from store to store. Making them all work in one flow is harder than it looks.
- Who handles what comes after? Returns, customer service, fraud and customer data: for a store, giving up checkout means giving up part of the relationship with the buyer.
Where the revolution stopped
The half that has arrived is product discovery. A McKinsey study of around 21,000 consumers in 11 countries found that more than half already use AI for shopping, and that for 29% of them it was the biggest influence on their latest purchase. Morgan Stanley estimates that by 2030 AI agents could drive between $190 billion and $385 billion of US online spending.
The missing half is the transaction. Beyond data and trust, clear rules are missing. The best-known case is Amazon’s lawsuit against Perplexity, whose Comet browser shopped on Amazon on users’ behalf. In March a judge blocked the agent; in August the Ninth Circuit Court of Appeals lifted the block, reasoning that it’s the user, not the agent’s developer, who accesses the site. The case goes on, though, and the underlying question remains open: can a store decide which agents to let in?
And in Europe?
Almost everything started in the United States. In Europe the first signs are showing, mostly in payments. In March 2026 Santander and Mastercard completed Europe’s first live end-to-end payment executed by an AI agent, in a controlled setting but on real banking infrastructure. Around the same time Nexi and Google Cloud signed a preliminary agreement to bring UCP and AP2 to Europe.
There’s also a distinctly European issue. Under PSD2, online payments require strong customer authentication, and an agent that pays on its own has to deal with that rule. It’s one of the reasons we expect autonomous buying to arrive later here.
What an online store can do today
Waiting for the revolution to finish doesn’t pay. Agents are already choosing, and they choose the stores they can read best. With our clients we start here:
- Complete, structured product data. Clear titles, attributes, variants, GTIN codes, up-to-date availability and prices, delivery times and return terms: that’s what an agent uses to compare.
- Feeds and structured data. Up-to-date feeds in Google Merchant Center and schema.org data on product pages: that’s where assistants draw from today.
- APIs before the theme. A headless architecture exposes catalog, cart and orders through APIs, and it’s the foundation for connecting to the protocols when the time comes. We covered this in our MedusaJS vs Shopify comparison.
- A policy on agents. Blocking every AI bot means disappearing from assistants’ answers; letting all of them in invites scraping and abuse. It takes a deliberate choice, reviewed over time.
- A frictionless checkout. If the purchase stays on the site, whoever arrives from an assistant has already decided: guest checkout, digital wallets and few steps make the difference.
- Measure. Traffic from ChatGPT, Gemini and Perplexity already shows up in analytics, and it’s the first sign of how much agentic commerce matters for your store.
For now
Checkout inside assistants will come back, probably in different forms and with clearer rules. When it does, it will reward the stores that already have tidy data, ready APIs and a clear idea of which agents to let in. The work is now, on the half that’s already here: getting found and getting chosen. If you want to know where to start, here’s how we work on e-commerce.