AI orders on Shopify grew 15x, and the checkout underneath still assumes a human

AI-powered search now sends Shopify fifteen times the orders it sent in January 2025, a figure Shopify president Harley Finkelstein put on the company's February 2026 earnings call and which Greg Jarboe reported in Search Engine Journal on September 4. Almost none of the systems taking those orders were built for a buyer that is software.

Three transaction protocols appeared in roughly a year: Google's Universal Commerce Protocol, OpenAI's Agentic Commerce Protocol, and Salesforce's Agentforce Commerce, which lined up behind UCP rather than launch a rival. Etsy sellers were first to trade inside ChatGPT; Glossier, Spanx and Vuori followed from Shopify's side. OpenAI has since backed away from closing purchases in the chat window, sending them into retailer apps instead. The protocol layer keeps moving. The layer underneath, which takes the money, has not.

An agent does not shop like a person

Konstantin Klyagin, who founded the QA agency QAwerk in 2015, told Jarboe that being surfaced by an AI platform is the simple half of the problem. People shop erratically: they stall, leave a cart, come back that evening. An agent works through the criteria it was handed in a burst of API calls and commits within seconds. Bot detection and rate limiting were built to stop automated traffic, and a sanctioned shopping agent is indistinguishable from the traffic they exist to block.

Klyagin sells the testing he recommends, and he declined to oversell it: he told SEJ he has not seen a verified production incident in which an agent broke a client's checkout. What his team keeps finding is duller — components that disagree about the state of a purchase. On one client funnel, the same plan data was being stored twice by two different components, with the price and attributes drifting slightly apart. A person refreshes the page. An agent that picks a plan from one store and meets a checkout validating against the other stalls in a state it cannot resolve.

Anyone who has shipped software pricing knows that failure. Plan data usually lives in at least three places: the marketing CMS, the billing system, and whatever component gates the product. Keeping them in agreement was cosmetic while the reader was a person who could work out that $49 and $50 meant the same tier. It becomes a revenue problem the moment the reader has nothing to look at and nothing to reason with. Klyagin expects the same shape elsewhere — a catalog feed and a checkout that disagree on stock, a retry after a timeout hitting an endpoint nobody made idempotent, a refund settled on your side that the calling agent never hears about.

The visibility work of 2026 sits entirely upstream of the only question that decides revenue: can the buyer finish?

The buy side already built the paperwork

On the same day, Digiday's Sam Bradley reported the mirror image: agencies putting AI into planning and buying are learning that an unwatched agent saves nobody money. Gartner surveyed 1,300 senior marketers in April and found 56% of companies have no clear policy governing how AI tools get used, marketing leaders least likely to attach financial controls. Gartner also expects three in five AI-using organizations to overspend, for the plain reason that nobody is counting.

So the agencies started instrumenting. Rise, part of the Quad group, has run AI media-buying agents for an unnamed supermarket client since June on top of PubMatic's audit log, which timestamps every change an agent makes and flags one straying past its limits. PubMatic sells that tooling, so read it accordingly. Dept now routes every request through an "AI gateway" that picks the model centrally, after what technology evp Jonathan Whiteside described as people who "burn through 1.5 million tokens in a day". Brainlabs issues tokens in tiers; PMG caps each user daily. Whiteside's rule is the one worth copying: "Every deliverable has an accountable human."

The asymmetry is hard to miss. The side spending money through agents built logs, ceilings and named owners within months. The side taking money from agents spent the year on schema markup and citation counts. Both are half-instrumented, in opposite directions: buyers are metering cost and not correctness, sellers are measuring visibility and not completion.

What to change this quarter

• Find every place a plan name, price or feature limit is stored and prove the copies agree. Three copies of your pricing is three answers to the same question.
• Put the checkout or trial-signup API under a burst of concurrent calls and see whether it holds. Years of human traffic never tested that.
• Compare what your pricing page publishes in machine-readable form against what billing actually charges. A contradiction there is worse than silence.
• Run a refund, a downgrade and a cancellation with nobody watching, and confirm the new state reaches the caller. Nothing else will catch a half-finished one.
• Ask who reads the log when your own marketing agents run, and where the ceiling sits. No cap and no named owner means you are the 56%.

The point is sequencing. Software marketing spent 2026 competing to be found by machines, on the assumption that being found was the hard part. Both sides now suggest the hard part is downstream, in systems that tolerated inconsistency because humans do. Agents do not, and they will not tell you: a stalled agent purchase leaves no abandoned-cart record, no support ticket and no analytics line explaining what happened.