Shopify Lets Shoppers' Agents Check Out, and Brand Agents Lose the Pre-Purchase Conversation

Two kinds of AI agent now meet at a Shopify store. The brand agent is the one a merchant pays Siena, Gorgias, Intercom Fin or Sierra to run on its own surfaces: website chat, Instagram DMs, email. The shopper's agent works for the buyer. Shopify's Fall '26 developer release (October 28) packages the plumbing that lets a shopper's agent search a store, build a cart and check out, with the shopper approving the purchase.

If a shopper's agent can call "search," "add to cart" and "check out" directly, it has no reason to open a conversation with the brand agent. None of the four vendors documents a public way for a shopper's agent to shop through them.

Two brand-side positions survive. The first is the set of actions only the merchant can perform, mostly after the sale: refunds, exchanges, warranty claims, subscription changes. The second is selling on channels outside agents don't reach yet. Pre-purchase chat on the brand's website is the exposed position, and the exposure runs into the vendors' revenue. All four tie their usage charges to a conversation or a contracted outcome. When the shopper's agent completes a purchase without talking to the brand agent, the vendor collects nothing beyond fixed platform or seat fees.

What's live, and what's still a draft

Three terms first:

• Shop MCP is a Shopify-hosted endpoint any agent can call for catalog, cart and checkout.
• WebMCP is a browser mechanism that lets a web page expose its own tools to an agent running in the shopper's browser, so the agent calls "add to cart" instead of clicking through the page.
• Shop Pay is Shopify's saved-payment wallet, tied to a Shop account that works across every Shopify store.

If you're dating the shift, use September 28. Since then a browser agent has been able to finish a purchase on a Shopify store. October 28 gathered scattered pieces into one developer entry point.

Who approves the purchase

All four vendors stop short of payment approval in their documentation. Gorgias and Fin document recommendations and cart help that end in the store's normal checkout (Gorgias, Fin). Siena's product page shows Add to cart buttons and no payment flow. Sierra's retail material stops at recommendations.

Brand agents owned the step before approval, which is persuasion. On the new path the shopper's agent handles that step, comparing across stores as it goes.

Where approval sits now depends on which row of the table the agent is using. There are three places.

• Browser route (Checkout WebMCP). The shopper confirms on the merchant's own checkout page, in their own tab. Shopify's reference says an agent's technical ability to check out doesn't grant permission, and a changed total means asking again. The merchant keeps the approval screen.
• Direct route (Shop MCP). The agent renders the order review: merchant, items, total, warnings, legal links. The shopper picks a payment method, and Shop Pay checks that payment matches what was confirmed (guide). The merchant's terms reach the shopper through another company's interface. Shopify specifies how agents must display them; nobody has independently tested whether they do.
• Budget permission (described, not dated). Shop's consumer help describes a setting, "Allow agent to pay for me," where the shopper sets a limit and stops approving each purchase. If it ships as written, approval moves off the purchase and onto a settings page. The help page offers weekly, monthly or total budgets. The payment handler spec lists only monthly. Neither gives a launch date.

On the merchant's side, Shopify says the merchant stays merchant of record, so it carries refunds, taxes and chargebacks. It receives the order and, on some Shop Pay paths, an approval_id. That ID tells the merchant an approval happened. It doesn't show what the shopper saw or which terms they accepted, and Shopify documents no merchant-facing record of either.

Picture the dispute: "My agent bought the wrong size." The party liable for the chargeback holds the least evidence about the moment of consent. That reading comes from the docs, and Shopify may pass more than it documents.

Google's AP2 spec, the Agent Payments Protocol, handles this differently. The user signs a standing permission with limits, such as a price cap. The agent later binds that permission to one specific purchase, and the resulting proof is built to travel to whoever has to verify it, merchant included. Sierra's protocol lists verifiable signed approvals as an open item. On Shopify today, approval rests on an instruction to the agent plus a wallet check.

Which customer context persists, and with whom

Brand agents build their advantage from the merchant's records plus the live visit. Siena's shared context pays off when a customer comes back to the brand's site or social channels. Gorgias's recommendations read the current session: pages viewed, cart contents. A purchase through Shop MCP never loads the storefront, so there is no session to read. Neither vendor says this directly; it follows from setting the two sets of docs side by side.

So the first agent-placed order leaves the merchant with its thinnest record. Why the shopper picked this brand over three others stays with the shopper's agent, and that reasoning is what a brand agent would want in every later conversation.

What the merchant can still see

Outcomes, in aggregate. Agent-placed orders appear in the Shopify admin labeled by agent. An Agentic dashboard reports sales, AI-referred sessions, conversion and the queries agents ran. It does not show how the agent decided.

The clearest documented case is Microsoft Copilot's direct checkout, which belongs with the direct route. The purchase completes in Copilot and the storefront never loads in the shopper's browser. Google Analytics and the merchant's tracking scripts don't run; Shopify sends "checkout started" and "checkout completed" from its own servers.

The add-on modules merchants place in checkout, for chat, loyalty, upsells and marketing consent, may not appear. Chat is where a brand agent lives. This is one channel's documented behavior, so don't extend it to the browser route, where the storefront still loads.

Two cautions. The dashboard blends Shop app sales, AI referrals and direct agent checkouts, so it can't tell you what share outside agents placed. And brand-agent vendors sold transcripts as part of the product: what shoppers asked, where they hesitated, which objections kept coming up. That data feeds merchandising and the vendor's own analytics, and it shrinks as agents finish more purchases.

Where each brand-side agent stands

Read the billing column first. In the published pricing, no fee from any of the four attaches to a plain Shopify order. Outside fixed platform or seat fees, every usage charge needs a conversation or a contracted outcome. Agent-to-agent interaction, which this release makes more likely, has no settled price anywhere; Intercom says Agent API pricing is still being finalized.

Sierra is closest to being the endpoint a shopper's agent calls. Its retail scope already covers the merchant-only actions (refunds, warranty claims, subscriptions, account changes), and it is the only one of the four drafting how an outside agent reaches it.

Fin's Agent API docs don't show it reaching the Ecommerce role's catalog, cart or checkout. What other agents can call today is customer service.

Siena is betting on social DMs, where the shopper is still a person. Meta co-authored the protocol with Sierra, so the company that owns Siena's strongest channels is helping define how personal agents reach company agents. What that means for Siena is speculative, but worth tracking.

Gorgias sits where post-purchase work lands no matter who bought: returns, tracking, exchanges. Its Shopping Assistant is the exposed piece.

Where the category is heading, and what to watch

Observable now. Agent purchasing is live on Shopify. Approval rests on agent instructions and wallet checks. Merchants see orders and aggregates. No brand-agent vendor documents catalog, cart or checkout access for a shopper's agent.

Not observable: volume. Adobe's October 29 data shows September AI-referred visits up 71% year over year, converting 90% better and earning 83% more per visit. Those are people clicking through to a website, where brand chat still gets a turn.

Two retailers interviewed by Digital Commerce 360 put AI referrals at about 5% and under 10% of total site traffic. Those are self-reported anecdotes, not a market share. Referred visits and agent-placed orders are different denominators, and only the first gets published.

Projection, moderate confidence. The brand-agent category splits in two:

• Service endpoints that shoppers' agents call for merchant-only actions.
• Selling on human channels outside agents can't reach.

Website pre-purchase chat erodes slowly rather than collapsing, which is what the referral numbers suggest. Meanwhile the brand's pitch to a shopper's agent increasingly arrives as catalog data: price, compare-at price, images and attributes. Maintaining that data becomes a larger part of the selling work.

Watch for:

• A commerce endpoint for shoppers' agents from any of the four, and its price. Fin's Agent API pricing will be the first number.
• The Personal Agent Protocol settling payments and signed approvals, the likeliest source of merchant-side proof.
• Universal Checkout leaving early access, and how it recovers when a multi-merchant checkout partly fails.
• A launch date for budget permission, and which budget periods survive.
• context.intent appearing in the merchant's customer profile. That would be the first piece of the agent's reasoning to reach the brand.
• Shopify reporting agent-placed orders separately from Shop app and referral sales.

The hardest design problems here are approval screens rendered by someone other than the merchant, and brand agents built to be called by another agent rather than to open a conversation. A company in this space that is hiring for either is working on the part of the problem that matters most.