In September 2025, OpenAI launched Instant Checkout, letting people buy inside ChatGPT. By February 2026 it had extended to over a million Shopify merchants. In March 2026, OpenAI dropped plans to operate checkout inside ChatGPT and moved to a discovery-first model instead: the assistant handles research and intent, and the merchant keeps the checkout under the Agentic Commerce Protocol.
That reversal is the most important thing to have happened in agentic commerce, and it barely registered as news. It settles a question that had been genuinely open — who owns the transaction — and the answer is you.
Which means the answer to a second question follows automatically. When an AI agent buys the wrong size, who edits the order?
This guide covers:
What agentic commerce actually looks like in 2026, by the numbers
Why the discovery-first model puts post-purchase back on you
Why agent-assisted orders are more likely to need editing
The problem with edit flows that assume a browser session
What to get right now, while volumes are still small
Where Agentic Commerce Actually Is
The gap between AI-assisted shopping and AI-completed purchasing is the whole story, and most coverage collapses them.
Behaviour | Adoption |
|---|---|
Used AI to shop in the past six months | ~78% |
Using AI for shopping weekly or more | ~43% |
Willing to let AI purchase on their behalf | ~34% |
Actually letting AI complete a purchase | ~8% |
So the overwhelming majority of AI shopping is research, and a small minority is delegated buying. Nearly one in three people say they will not let AI spend on their behalf at all.
But the traffic is real and it converts. AI agents were credited with driving around 20% of global orders during the 2025 holiday season, and in March 2026 AI-referred traffic to US retail sites converted materially better than non-AI traffic — a reversal from the year before, when it converted worse.
The trajectory matters more than today's share. Around 60% of shoppers expect to use AI agents within a year, and ChatGPT alone reached roughly 900 million weekly users by February 2026.
Discovery-First Means Post-Purchase Is Yours
Under the model OpenAI briefly pursued, the assistant owned checkout. Post-purchase support would plausibly have followed — the customer's relationship being with the interface they bought through, not the store.
Under discovery-first, that inverts. The agent finds the product and forms the intent; your checkout takes the order; your confirmation email lands in the customer's inbox; your order status page is where they go when something is wrong. You own the entire post-purchase surface, including every mistake made during a conversation you never saw.
This is good news commercially and inconvenient operationally, because agent-assisted orders arrive with characteristics that make editing more likely, not less.
Why Agent Orders Need More Editing
The Buyer Did Not See Your Product Page
A customer who reached checkout through your PDP saw your size guide, your fit notes, your variant imagery and your shipping copy. A customer who arrived via an agent's summary saw a description of your product, generated from whatever the agent could retrieve. Any gap between that summary and reality surfaces after payment.
Specification Drift
Agents compress. "Navy, medium, cotton blend" is a reasonable summary that loses the distinction between two navies and three fits. The buyer confirms a description that matches their intent and not necessarily the SKU that gets ordered.
Address and Contact Data Come From Somewhere Else
In a delegated purchase, the shipping address may come from a stored profile rather than being typed with attention. Stale addresses are already the largest controllable cause of failed deliveries, and an agent filling a form from memory is not a mechanism that improves address freshness.
Realisation Is Displaced in Time
With delegated purchasing, the buyer often reviews what was bought after it was bought — when the confirmation arrives. That is precisely the pattern that makes a post-checkout correction window valuable, and it is more pronounced here than in ordinary self-directed shopping.
When There Is No Browser Session
Here is the practical problem, and it is a design problem rather than a volume problem.
Most post-purchase edit surfaces assume a logged-in browser session: the customer returns to your site, is recognised, and edits from an account page. Agent-mediated purchases break that assumption in several ways at once. There may be no account. The session belonged to the agent, not the buyer. The buyer may never visit your storefront at all — their entire relationship with you runs through email.
Which means, concretely:
The confirmation email has to carry the edit path. If editing lives behind a login the buyer does not have, it does not exist for them. A tokenised link from the email is the reliable route.
The order status page has to work for guests. Not "create an account to manage your order" — that is a wall in front of a correction you want them to make.
The edit surface has to render without a rich client context and be legible on a phone, since the confirmation email is usually read there.
Payment on an upward edit cannot assume a stored card. A delegated purchase may have used a payment method the buyer cannot re-authorise on your site.
Shopify's order status page is the natural home for this precisely because it is reachable from the confirmation email without a login.
What to Do Now
At roughly 8% of buyers actually delegating purchases, this is not yet a volume problem. It is the right moment to get the architecture right cheaply, before it is.
Make your product data unambiguous. Agents summarise structured data. Precise variant naming, explicit fit and dimension attributes, and clean specifications reduce the drift that causes post-purchase corrections. This is the same work that gets you cited by AI answers in the first place.
Put the edit path in the confirmation email. Not behind an account. This single change covers most agent-mediated cases and helps guest checkout generally.
Assume the address needs confirming. Validated post-checkout address correction matters more when the address was filled from a profile rather than typed.
Tag the traffic source. You cannot analyse what you do not label. Segment orders by AI referrer so you can compare edit rates, return rates and AOV against ordinary traffic.
Keep fulfillment holds honest. None of the above helps if the warehouse pulls the order before the buyer has read the confirmation. Same sequencing constraint as everything else post-purchase.
Conclusion: You Kept the Customer, So You Kept the Problem
The agentic commerce discussion has focused almost entirely on acquisition — how to be discovered, cited, recommended. That is the visible half.
The half nobody is planning for is what happens after an agent-assisted purchase goes slightly wrong, which it will do more often than a self-directed one, because the buyer never saw your product page and may not have typed their own address. When OpenAI handed checkout back to merchants, it handed back the post-purchase relationship too. The support ticket, the wrong variant, the stale address and the refund are all yours.
The good news is that the fix is not agent-specific. A correction window reachable from the confirmation email, addresses validated after checkout, and fulfillment that waits — these serve every customer. Agentic traffic just raises the cost of not having them.
Ready for Orders You Did Not Watch Being Placed?
If an AI agent puts the wrong variant in a customer's order, the correction has to happen on your storefront, from your email, before your warehouse picks it.
Account Editor puts self-serve editing on the order status page reachable from the confirmation email — no account required — with validated address correction, payment protection on upward edits, and fulfillment holds that keep the window real.
See how Account Editor handles post-purchase for agentic orders.
