Shopify let browser agents check out. Who pays when they buy wrong?
Shopify opened checkout to browser AI agents on 28 September. Six banks say nobody has settled who pays when an agent errs. What a retailer carries.
| What changed | on 28 September 2026 Shopify gave browser AI agents checkout tools (get_checkout, update_checkout, complete_checkout) that need no merchant configuration (Shopify developer changelog) |
|---|---|
| The safeguard | Shopify tells agent builders to show the buyer the order and total and get permission before complete_checkout; the tools do not accept new card details (Shopify Checkout WebMCP documentation) |
| The banks' view | on 22 September 2026 ASB, Bank of America, Capital One, Commonwealth Bank of Australia, ING and NatWest said liability is unclear when an AI agent exceeds the authority a customer gave it (Building Trust in Agentic Commerce) |
| Visa's rule | a cardholder is responsible for actions taken by a registered Agentic Payment Provider as if the cardholder initiated the transaction (Visa Core Rules, 18 April 2026, ID 0031176) |
| Evidence rule | Visa's Compelling Evidence 3.0 counts a login ID for an Agentic Payment Provider or the merchant's site as a matching identifier against a fraud dispute (Visa Core Rules, 18 April 2026) |
| Network monitoring | Visa flags a merchant as excessive only at a fraud-and-dispute ratio of 150 basis points or more and at least 1,500 such cases a month in the US, EU, Canada and Asia Pacific (Visa VAMP fact sheet) |
On 28 September 2026 Shopify gave AI agents running in a shopper's browser the tools to fill in and submit checkout on Shopify stores, with no setup by the merchant. Six days earlier, six banks including Bank of America, Capital One, ING and NatWest published a paper saying nobody has yet settled who pays when an agent buys something its owner did not want.
The verdict for a retailer with a few hundred staff: this changes less for you than the headlines suggest. An order placed by a browser agent reaches you as an ordinary online order, and the dispute rules you already live with decide who carries it. The work this week is to make sure your evidence and your product pages would win that dispute.
What actually happened
Shopify's changelog lists the new tools plainly: one reads the checkout, one updates fields such as address and delivery option, one submits the order. Shopify says they need no merchant configuration. Its documentation puts two limits on the agent: the tools do not accept new card details, and before submitting, the agent must show the buyer the order and the total and get permission to place it.
That permission step is an instruction to whoever builds the agent. The merchant cannot see whether it happened, and the documentation does not describe a flag on the order that tells you an agent placed it.
The banks' paper, Building Trust in Agentic Commerce, is voluntary and binds nobody. Its useful part is the admission. It says expectations can differ on who is liable if an agent exceeds its authority, that dispute processes do not involve every party in the chain, and that agent providers should keep auditable records of what the customer instructed. It also notes merchants' worry that chargebacks may rise for reasons outside their control.
Who pays if an AI agent buys the wrong thing?
For now, the same party as before. Visa's rulebook, in force since 18 April 2026, makes the cardholder responsible for what a Visa-registered agent does, as if they had bought it themselves. Any other "my agent got it wrong" claim is an ordinary card-not-present dispute, and the merchant answers it with ordinary evidence.
That Visa rule covers only agents enrolled in its Intelligent Commerce programme and paying with a Visa token. Shopify's documentation does not say whether browser-agent orders arrive marked that way, so plan as if they do not.
The new risk is not a new kind of dispute. It is an old dispute with a buyer who never looked at your page.
What it does not mean
It does not put a mid-sized shop near a card-network penalty. Visa's monitoring programme flags a merchant as excessive only when fraud and disputes reach 150 basis points of transactions and at least 1,500 cases in a month in the US, Europe, Canada and Asia Pacific. A retailer of your size would feel the cost one refund at a time long before that.
It also does not mean you should block agents, buy an agent-fraud product or wait for the networks to finish their rules. The banks themselves say a second paper on implementation is still to come.
What to do this week
Check one disputed order from the last quarter and ask whether you could win it if the customer said software placed it. Visa's Compelling Evidence 3.0 rules let you answer a fraud dispute with earlier undisputed orders from the same device or IP address plus a second match, and the April 2026 rulebook names a login ID on your site, or with a registered agent, as one such match. That only helps if your shop keeps device, IP, login and delivery data for each order.
Then make the product page carry the facts an agent reads: size, material, compatibility and the returns terms, in plain text rather than in images. An agent that buys the wrong size because the size chart was a picture is your return, whoever is liable on paper.
Last, tell your customer-service team that "my AI bought it" is now a reason they will hear. Treat it as a normal return within your policy, log it, and count the cases each month. If that count is still near zero by spring, you have your answer about how much of this is real for you.