Muse, Instinct and dots are learning to place one order for one person. Gifting, rewards and procurement platforms need a thousand a day with nobody watching. Same protocols, different business.
Two motions that look alike and are not
Two things are both being called agentic checkout right now, and they are different businesses.
The first is a personal agent placing one order for one person. Muse, Instinct, Grok and OpenAI's dots do this, and they do it with the human in the loop: the agent finds the options, the human picks, the agent runs the checkout and the human confirms the size, the address and the payment. I wrote about that split in Human in the Loop, or Not. Across a day, that motion is a thousand different people each placing one order, with a thousand different humans watching their own.
The second is a platform placing a thousand orders a day through you, with no human watching any of them. A gifting platform, a rewards or loyalty program, a procurement tool, a vertical AI app with a buy button. The decision was made up front; everything after it is programmatic.
Same protocols, same merchants, same card networks. Completely different operating model. The first is best effort, and best effort is fine for it. The second is not, and the rest of this post is about why.
One order each for a thousand people
A shopper asks for running shoes under $150, the agent finds them, the agent buys them. It is impressive, and it is also the easiest version of the problem, because the human is still in the loop as the error handler. If the agent picks the wrong size, the shopper notices. If checkout fails, the shopper retries or opens a browser. If the merchant's fraud engine blocks the card, the shopper gets an email and sorts it out. The cost of a failure is a few minutes of one person's time.
Nobody promised a success rate and nobody is measuring one. A 60% completion rate feels fine because the four failures in ten each land on a person who can absorb them, and the six successes feel like magic. Volume is spread across thousands of unrelated shoppers, so no single merchant, card issuer or fraud system sees a pattern. The agent gets to be a slightly clumsy personal assistant, and for the consumer that is enough.
A thousand orders for one platform
Now flip it. One customer places a thousand orders a day through you: a gifting platform fulfilling gifts the recipient never clicks through for, a bank rewards or loyalty program redeeming points into real merchandise, a procurement tool buying on a company card with nobody at a browser, a vertical AI app in fashion, beauty, home or travel that wants checkout inside its own experience without becoming a retailer. None of them retry failed orders by hand. They have a contract with you, and their end users have a contract with them.
Five things change.
Failure becomes a rate. The rate you tolerated on the consumer side becomes impossible. The 60% that felt fine for one shopper is 400 failed orders a day for a platform. Even 99% is not good enough: that is 10 failed orders a day, 300 a month, each one a support ticket, a refund or a complaint that reaches your customer's CEO.
Volume creates patterns, and patterns get noticed. The thousand orders are spread across hundreds of merchants, but they all come from the same infrastructure: the same IP ranges, the same device fingerprints, headless sessions with no human behavior signals, and a small set of corporate cards doing hundreds of transactions a day. Any one merchant may only see 20 of those orders, and 20 identical-looking sessions is enough for its bot mitigation layer to flag them. The card issuer sees all thousand. What one shopper could do invisibly, a platform has to do openly, with agent identity, consented credentials, a payment reputation, and merchant relationships.
The world under you keeps moving. Merchants redeploy their storefronts, change a checkout step, add a CAPTCHA, rotate a payment provider. Over a protocol, a version changes or a feed goes stale. On a storefront, a DOM change at 2 a.m. breaks a flow that worked at midnight. With one order a day you find out next week. With a thousand, you find out in 20 minutes, from your customer.
Every order needs a paper trail. Which SKU and variant was requested, which was bought, at what final price including tax and shipping, with which merchant order ID, charged to which card, with what confirmation. When a retry fires, was the first attempt actually placed? Duplicate orders at scale are worse than failed ones. Idempotency is not a nice-to-have; it is the difference between a refund queue and a lawsuit.
The money gets complicated. A thousand orders a day across hundreds of merchants means reconciliation: payment captured versus order confirmed versus shipment, partial cancellations, returns, affiliate commission tracking, chargebacks. One shopper never reconciles anything. A platform reconciles everything, daily.
None of this is unique to agentic commerce. It is the same gap that separates a script that posts to an API from a payments processor. What is different is that the industry is currently grading agents on one person's experience at a time and assuming the thousand-order version is a matter of turning up the dial.
The two motions side by side
Two businesses, not one
The personal agents will get very good at one order each for a thousand people. Meta, OpenAI and Instinct, valued at $10B, are competing on it, and the human in the loop forgives a lot while they learn. That motion is a consumer product problem, and the winners will be whoever provides the most convenience and removes the most friction.
A thousand orders for one platform is a different business with different economics. The customer is a company, the volume is concentrated, the failures are a rate, and nobody downstream is going to fix anything by hand. It is won by whoever runs the most reliable execution, payment and fraud-cooperation stack, and it is sold on a success rate, not a demo.
CartAI was built for the second motion. If you run a platform that needs to place orders at volume without a human babysitting each one, get in touch.
