How agentic checkout works
An agentic checkout is a purchase finished by software acting for a person. The shopper tells an assistant what they want, the assistant finds and compares candidates, and then completes the order — the part that is new — without the shopper ever loading your checkout page.
Three things have to be readable by a machine for that to happen. The product: what it is, the price including tax and delivery, whether it is in stock, when it arrives, how it can be returned. The offer: a purchasable item at a stated price, not a page a human has to interpret. And the payment: some way for your shop to know the assistant genuinely holds authority to spend a stated amount for a named person, and for the shopper’s card or wallet to honour that.
The first two are familiar work — the same product structured data and merchant feeds that already drive shopping results. The payment and authority piece is where the standards work sits, and it is not settled. Payment companies, storefront platforms and AI providers have each put forward an approach. Most merchants will meet agentic checkout through whatever their store platform adopts rather than by building anything themselves.
Why agentic checkout matters
If a slice of buying moves inside assistants, what wins the sale changes. There is no hero image in the middle of it, no banner, no persuasive layout. What the agent can read decides whether you make the shortlist: accurate stock, an honest delivery estimate, a plain returns policy, an unambiguous total.
So it is a data quality problem long before it is a technology problem. That is the useful part — the same tidying makes ordinary shopping listings and visibility in AI search results better today, whichever standard eventually wins.
Where agentic checkout goes wrong
The first mistake is treating it as a channel you can buy into. There is no auction to enter. Anyone offering to place your products inside assistants should be asked precisely what is being placed and where it appears.
The second is blocking the very traffic you want. Bot rules written to stop scrapers do not reliably tell a shopping agent from a nuisance, and a blanket block will turn both away at the door.
The third is assuming the order is somehow safer because software placed it. An order with no human session behind it still has to pass your fraud checks, address validation and refund process, and a dispute still lands on you. There is also a hard local limit worth knowing: where much of ecommerce still settles by cash on delivery or a domestic wallet, as it does across Nepal, an agent that can only pay by card cannot finish most orders at all.
What to do about it
Fix the readable layer first, because it pays regardless. Make sure every product carries correct structured data, that the feed matches the page, that stock and delivery times are true rather than optimistic, and that the returns policy is written where a machine can find it, not buried in an image.
Then ask your platform, rather than a vendor, what it supports and when. Storefront software is where this will arrive for most shops, and a plan that depends on custom development is usually premature.
Finally, learn to see it in your reporting. Assistant referrals and agent traffic look like nothing much in a standard report, so agree now how you will label them — otherwise the first orders of this kind will be filed under direct traffic and nobody will notice the shift.