Agentic Commerce Checkout: Who Should Build It

Some of your traffic is already not a person

When a shopping agent lands on a site built for thumbs, it usually gets most of the way to a purchase and then stops. It hits a form, a CAPTCHA, or a session cookie that has to be earned by loading three pages of JavaScript. The transaction dies there. There is no support ticket, and nothing in your logs that says "a customer wanted to buy and we would not let them." Just one more visitor who bounced, indistinguishable from a window shopper.

That's the most expensive kind of lost revenue, because you cannot see it. There's a mirror image of the same problem, and Checkout.com describes it well: when a shopper's assistant reaches the moment of buying, it often kicks the shopper back out into a browser checkout built for humans, which breaks the experience and increases basket abandonment.^[1] Same failure either way. A purchase that was ready to happen dies at a form. The answer isn't a handoff to a human checkout, it's a checkout that both the buyer and the agent can finish.

The receipts

Agents are not a next-year thought experiment. Forter's fraud network, which sees orders across a large set of retailers, counted steady growth in agentic orders through the summer of 2026: up 24% month over month from August to September, and 48% once Instinct and Grok Bot are counted. When Meta's Muse showed up in late September, agentic orders in its first detected week ran 229% above the prior month. By late September, all agentic orders on that network sat at 4.4x their early-August level.^[2] That's one network, so don't read it as the whole market. But these are real orders at real merchants with real dollars attached.

The demand side lines up. NielsenIQ found that 51% of U.S. consumers had used at least one AI tool to help them shop in the past month.^[3] Global Payments found consumers across seven countries now expect agents to make 15% of their purchases within five years, up from 9% a year earlier, and that 45% of Americans have already used an AI shopping agent or would consider it.^[4] Visa Acceptance's 2026 index, built by PYMNTS Intelligence from surveys in the U.S., Brazil, and the UAE, found 30% of consumers now use ChatGPT to research products before buying, against almost nobody two years ago, and 64% expect to use AI shopping agents within two years.^[5]

Where this is happening should worry a business owner more than the totals do. A Salesforce survey of 4,690 consumers found that between August 2025 and May 2026 the rate of shoppers discovering products through brand-owned properties fell 7%, while traditional search fell 15% and the share choosing newer channels (AI assistants, social AI, delivery apps) grew 38%.^[6] Separate behavioral data from a panel of more than 1.5 billion shoppers shows traffic referred from AI chats growing by 150% to 428% year over year in every quarter Salesforce measured, against overall traffic growth in the single to low double digits.^[6]

Discovery is leaving the places you built and arriving from places you don't control. The rails for selling into those places exist and are past the announcement stage: OpenAI's Instant Checkout has been live inside ChatGPT since September 2025, and Google handed AP2 to the FIDO Alliance to run as an open standard rather than a vendor's product.^[7]^[8]

Two standards carry most of the weight. ACP comes from Stripe, OpenAI, and Meta.^[9]^[10] UCP comes from the Google and Shopify side of the industry.^[11]

Three questions that decide whether this is your problem

Bad advice here comes in two flavors. One says every business needs an AI strategy. The other says agentic commerce is hype. Both skip the filter that matters.

Is the buying decision routine, or does it need a person? An agent is good at "reorder the printer toner we bought last month." It is bad at "help me pick a venue that feels right for our wedding." The more spec-driven and repeatable the decision, the better the fit.

Does the order happen again? If your customer buys the same thing on a schedule, the agent has something to automate. Forter's own reading argues against me here: it describes agents as "largely used for more considered purchases," and calls Google's smaller orders, which it says "suggest single-item or repeat purchases," atypical.^[2] So the case for replenishment below is my argument from the shape of the demand, not a finding I can hand you.

Can a machine resolve your catalog without calling you? If an agent cannot determine the exact SKU, the price, and whether it's in stock, nothing else matters. Inventory accuracy blocks more businesses than protocol support does. Among the Singapore organizations Salesforce surveyed, 46% of multichannel ones said inventory was not synchronized in real time, and 36% named inconsistent pricing and promotions as a failure point.^[6]

Three yeses and you're a candidate. If your answer to the first question is "the human relationship is the whole product," stop here. This isn't your problem yet, and probably won't be for a while.

Who should build this first

Replenishment and consumables. Office supplies, janitorial, ink and toner, packaging, lab reagents, coffee. The buyer already wanted to stop thinking about it. A standing order list is the most automatable purchase in commerce, and a fifty-dollar order makes up in frequency what it lacks in size.

Industrial and B2B supply. A business purchase order is a spec list: part number, quantity, sometimes a due date. It isn't a conversation. These buyers don't want a prettier catalog. They want the order to place itself, without replacing anyone's ERP or accounting system.

Parts and supplies for the trades. HVAC, plumbing, electrical, auto repair, dental and veterinary clinics. Same shape as B2B supply: catalogued, consumable, urgent, and ordered by someone who is currently on a ladder or under a truck.

Food service and hospitality supply. Restaurants, food trucks, hotels. Weekly staples, thin margins, and an ordering cadence that is already a written list. UCP's roadmap has food on it, and lodging already has a draft specification.^[11]

Bookable inventory. Travel, dining, tickets, appointments, service windows. Forter found that the agents doing the most scraping work were concentrated in tickets, travel, and digital services, categories where there is no shipping address.^[2] Those also happen to be the categories where the hard part was never payment. It was searching across dates, party sizes, and options, which is work an agent does better than a form does.

Digital goods, APIs, and credits. Buying an API credit or a data license is often a program's job already, triggered by a build pipeline or a scheduler rather than a person clicking a button. The humans here mostly want to set a budget and be left alone.

Long-tail marketplaces. If your catalog runs to 400,000 SKUs and your own search box can't find the right one, an agent with a spec will beat a human with a search bar. This is the case where agents are better than the alternative rather than just cheaper.

Who should wait, and why that's a real answer

Consultative, high-ticket B2B. RFQs, negotiated pricing, custom tolerances. The relationship is the product, and an agent cannot hold it.
Made-to-order and configuration-heavy goods. Some agents configure well, most configure badly, and a wrong spec costs you a rebuild.
Luxury and brand-experience retail. When the experience is the product, an automated purchase is a downgrade. Premium brands have always guarded the packaging as closely as the product, and that instinct holds here.
Regulated, age-gated, or prescription goods. Identity and authority have to stay in the loop, and that's what an agent naturally abstracts away. Payment mandates and verifiable credentials are starting to put some of it back,^[11] but they are early and they don't replace checking a date of birth or a prescription.
Businesses that can't tell an agent from an attacker. This one bites hardest. Fraud systems were built to spot automation, and legitimate agents look like automation.^[12] If your defenses block agent traffic, you don't have an agentic commerce problem, you have a false-decline problem that never appears in a dispute report.^[12]
Undifferentiated resellers competing on nothing but price. When the agent's comparison is price plus spec and you differentiate on price, you've automated your own margin away. Build an operational edge first.

Why it pays

The case is duller than the pitch decks suggest.

The demand is moving whether or not your site can take it. Brand-owned discovery down 7%, traditional search down 15%, newer channels up 38%.^[6] More keyword spend does not reverse that. Being a valid endpoint is cheaper than fighting for attention.

The handoff is where you lose. An agent that found your product is a high-intent buyer. Failing it at checkout costs the whole order, not a fraction of one.

Agent orders are decent orders. On Forter's network only 26% of agentic orders were $50 or less, with the $50 to $200 band as the sweet spot. Agents spend 14% less per order than humans, but only because Google's pilot restricts checkout to a single item. Take that pilot out and agents spend slightly more per order than humans do.^[2] These are considered purchases, not impulse candy.

You keep the customer relationship. ACP is an open standard created by Stripe, OpenAI, and Meta under Apache 2.0, and it leaves businesses as the merchant of record, in control of what can be sold, how it's presented, and how orders are fulfilled.^[9]^[10]

UCP is co-developed by Google, Shopify, Amazon, Walmart, Target, Meta, Microsoft, Stripe, Booking.com, and DoorDash, and it defines a full checkout capability, with operations to create, get, update, complete, and cancel a checkout.^[13] The spec is blunt about where it stops: the checkout "has to be finalized manually by the user through a trusted UI unless the AP2 Mandates extension is supported."^[13] Turn mandates on and the agent can finish without the buyer stepping in, which is the same consent question that comes up in what this costs.

UCP also defines continue_url, a handoff from the platform to the business's own UI so the buyer can finalize, and Shopify's cart flow uses it the same way.^[14]^[13] When Amazon, Walmart, and Shopify land on the same specification, the specification tends to stick.

What it actually costs

The protocol is the easy part. The work is data.

Product data a machine can read. Names, GTINs, dimensions, variants, materials, compatibility. If a person has to read your marketing copy to work out what the thing is, an agent cannot buy it.
Price and stock accuracy in real time. Stale inventory is worse than no inventory, because an agent will cheerfully order what you don't have.
Legible policies. Shipping, returns, warranty, lead time. Agents relay these to buyers, and gaps become abandoned carts and disputes.
A consent trail. As the fraud conversation matures, the question shifts from "does this look human?" to "was this agent authorized, within what limits, and did the purchase match the permission?"^[12] Payment mandates in AP2 and delegated authorization in ACP exist for exactly that.^[15]^[10]
Distinguishable traffic. Most merchants still can't tell an agent from a person: in Visa Acceptance's survey of 1,185 merchants, only 23% could distinguish AI-driven traffic from human traffic.^[5] If agent orders aren't tagged separately in your reporting, you can't see what's working, and your fraud rules will treat your best new channel as abuse.

If you want to know how far customers will actually go, Visa's trust gradient is the useful chart. More than half will let an agent search and compare products on their behalf, but fewer than 40% will let one touch their payment credentials.^[5] Build for the first group today and let the second arrive with the mandates above.

Order of operations

Fix structured product data and inventory accuracy for your top-selling SKUs. It pays off in AI search and human search, and you can start today.
Pick one rail. ACP for the ChatGPT-style channels, a UCP profile if you want the Google and Shopify ecosystem, or both if you have the appetite.^[9]^[11] Don't build one integration per agent. That's the point of a standard.
Instrument agent orders as their own segment so finance and risk can actually see them.
Set spend limits and approval thresholds around what your customers told the researchers: a third want to approve every transaction, and 42% worry the agent will buy the wrong thing.^[4]

What you don't need: a chatbot bolted to your homepage, a replacement for your checkout, or an end to human sales. ACP is additive by design, and its premise is that the seller keeps their existing backend and payment processing.^[9]

If you sell anything online and you're not sure how you'd answer those three questions, that's worth a conversation. We build the product data and the checkout that agents can use. See what we build or get in touch.

Agentic commerce checkout: who should build it and who shouldn't bother