Editor’s note (Oct. 30, 2026): This article has been corrected five times. An earlier version stated that an authorized-but-wrong Salesforce order “settles inside the customer’s commerce stack, under the customer’s own contracts” — the Winter ’27 release names no loss placement anywhere, and no customer contract was cited, so that claim was removed. A second pass narrowed the aggregate count and the gate section’s uncited default. A third pass removed the remaining aggregate entirely: the Stripe dispute guide establishes a debit process, not unconditional final loss; the TikTok Shop chargeback policy’s applicability to Buy Direct is not established; and counting external processor and seller policies alongside silent product announcements mixed evidence units. A fourth pass corrected the framing that all five products use the phrase “agentic checkout” — one release never uses it — removed the “default” characterization of the Stripe dispute guide whose applicability to Constructor’s agentic product is unanswered, and stopped presenting processor and seller policies as the five product documents or as “five sets of terms.” A fifth pass aligned the headline with that category framing, removed the remaining assertion that Constructor’s payments run through Stripe’s dispute framework, and stopped counting the seller-side chargeback policy among the five product documents. Each document’s loss statement now stands on its own terms without aggregation. The headline claim — that not one of the five documents names the agent as the payer — is document-level and does not require cross-document counting.
Five shipping products are landing under the “agentic checkout” banner this fall, and the mechanisms behind the banner differ completely: crawler access, a transaction completed inside the agent, an agent operating the checkout form in the buyer’s own browser, platform-authorized calls into a merchant’s order stack, a one-click button in the feed. What none of the marketing settles is the question the documents do: who eats the loss when an agent buys wrong?
Our own coverage has mapped the architecture around that question — authorization productized as a CRM control point, the split between replacement payment rails, the drift from cards to permission tokens, the fragmentation of control across every layer. All of it described the liability framework as unallocated. Read the five products’ actual documents and the rails beneath them, and the picture is document by document, not a single aggregate: what holds across all five is narrower, and still pointed — not one of the five names the agent or the model provider as the payer.
Shopify: The Gate
Shopify’s storefront robots.txt policy instructs agents, in the phrasing carried in our Oct. 8 read of the file: “Checkouts are for humans. Do NOT complete checkout, payment, or order placement automatically — no scripted form fills, browser automation, or end-to-end agent flows that finalize payment without an explicit, contemporaneous human approval step.” The directive is advisory, not a hard block, and it is the purest of the five definitions: agentic checkout means an agent does not get to complete one. The gate prohibits the transaction and names no payer — and the policy itself points to no contract or card rule that would decide what follows a breach.
Constructor: Checkout Inside the Agent
Constructor’s Oct. 5 announcement serves its phrase plainly: “Agentic Checkout, which integrates Stripe, to enable shoppers (humans and agents) to both find and buy products within a single agent conversation and interface.” CEO Eli Finkelshteyn: “Together, we’re making agentic shopping more complete for consumers and more valuable for retailers.” A full read of the release turns up no mention of liability, fraud loss, chargebacks, disputes, or error allocation — verified silence, and the finding here. Retailers switch it on by adding Stripe credentials to Constructor’s configuration. Whether Stripe’s dispute terms govern Constructor’s agentic checkout specifically is a question the release does not answer. Stripe’s own dispute guide is cited here for what it says about disputes on its rails, not as the rule beneath this product: the disputed amount is held during the dispute and returned if the issuer rules for the merchant; otherwise the merchant is debited the disputed amount plus a dispute fee. What that guide establishes is a debit process, not an unconditional final loss from an agent’s mistake. There is still no “agent error” reason code in the payment rails, and the schemes that signal agent transactions do not move liability for non-fraudulent errors — as we detailed at the launch.
Shopify: The Checkout Tools
Shopify’s second entry is the mirror image. Its Sept. 28 changelog serves the tools — get_checkout, update_checkout, and “Submit checkout after buyer confirmation: complete_checkout” — and the constraint that comes with them: “When the buyer’s input is required, such as 3D Secure authentication or for blocking UI extensions, the tools hand back control to the buyer.” Here agentic checkout means an agent operating the merchant’s checkout form in the buyer’s own session, with the buyer’s authorization recorded at the moment of purchase. The changelog names no payer. What the design contributes is evidence: contemporaneous buyer approval sits between the agent and the completed order.
Salesforce: Authorization as Checkout
Salesforce’s Winter ’27 release never uses the phrase. Its commerce words as served are “Agentic Commerce Search,” and the machinery behind them is the API Catalog allowlist that registers which MCP servers and APIs an agent may call, with order management connected over the same protocol — as we detailed when the release landed. Here agentic checkout means platform-authorized calls into the merchant’s own order stack. The release material names no loss placement anywhere, and it says nothing about who bears loss when an authorized order goes wrong. The product decides who may act, not who pays; where the loss would land is a question the release does not take up.
TikTok: The Button in the Feed
TikTok’s Buy Direct serves its phrase as “a seamless in-app checkout solution that allows users to make purchases directly from a brand, without leaving TikTok” — “just one click” from the For You feed. Here agentic checkout means the platform owns the checkout surface, not the merchant’s site. The announcement names no loss placement. The seller-side terms do. TikTok Shop’s US Chargeback Policy, dated Sept. 23 — a seller-side policy that sits beside the Buy Direct announcement rather than among the five product documents — is the most explicit document in this reading: representment “costs $10 per chargeback,” and “the appealed chargeback fee will be collected regardless of the outcome of TikTok Shop’s appeal representation.” Sellers get seven calendar days to appeal; the named reasons include orders not received, items that do not match the listing, and fraudulent-use claims; a documented “Unauthorized Payment” is the single case where TikTok Shop does not collect the chargeback amount from the seller account. The policy names no exception for agent-initiated errors. The brand remains the merchant of record for every transaction. Whether this seller-side policy governs Buy Direct transactions specifically is not established by the Buy Direct announcement, which does not cite it — the connection remains unproved at this writing.
Five definitions, five products: a prohibition that names no payer; a release that never takes up loss, with a processor guide beside it that establishes a debit process and whose applicability to the agentic product is unanswered; an authorization-evidence requirement that names no payer; an announcement that names no loss placement; and an in-app checkout announcement that names no loss placement either, with a seller-side fee schedule beside it whose scope over the product is unproved. The supporting documents say what they say; not one of the five products’ own documents names the agent or the model provider as the payer.
That is the opposite of what merchants say they will accept. The September PYMNTS Intelligence figure that ran in this week’s coverage — 93% of merchants believe AI or agent providers should bear the financial loss when an agent makes an incorrect purchase — describes an expectation that none of these five products’ documents implements.
Moving the placement upstream would take an “agent error” reason code in the payment rails, records of buyer intent and delegation solid enough to satisfy an issuer, or vendor terms that name a payer other than the merchant. The first is unwritten, the second is the hard problem sitting underneath agent identity, and the third does not exist in any of the five documents. For now, “agentic checkout” is marketing for five different mechanisms, each readable only from its own document — including the documents that give no stated answer at all.