Checkout has moved into the chatbot. American Express’s new safety net for agent mistakes comes with one small instruction: try the merchant’s returns desk first.
American Express promises to protect you from eligible mistakes your AI agent makes. Step one of that protection: start a return with the merchant, “where possible.” The industry first, it turns out, comes second. [C7]
Our reading: the merchant’s returns desk is the first stop for the robot’s error. Who finally eats the cost? Nobody has published that, and we are not going to make it up.
Checkouts an agent completes on its own, in Accertify’s data. They had to reach for basis points.
Through Sep 2026 · Accertify [C10]AI-referred retail traffic, year over year. Chatbots are very good at pointing.
Jul 2026 · Adobe via DC360 [C32]Conversion rate of AI-referred visits relative to non-AI visits. The pointing works.
Jul 2026 · Adobe via DC360 [C32]The two Adobe figures cover United States retail sites. A referral is a chatbot saying “try this shop,” not a chatbot buying anything.
“Agentic commerce” covers two very different things: an AI suggesting a product, and an AI actually paying for it. This brief is about the second, the one that can go wrong on your card. Accertify’s consortium puts autonomous checkout well below 1% of transactions. That is a window into its members’ traffic, not a census of the planet. [C10]
United States retail · year-over-year traffic growth (%) · discrete Adobe snapshots
Read as snapshots, not a trend line. Adobe’s windows are holidays, single days, months and quarters, and some overlap. These bars show reported growth in visits, not a continuous monthly trend and not the number of purchases an agent made.
Sources: Adobe 2024 windows [C26] · Jul 2025 [C28] · holiday 2025 [C29] · Q1 / Mar 2026 [C30]; via Digital Commerce 360 May [C31] and Jul 2026 [C32]. Data through Jul 2026.
More visits from a chatbot tell us nothing about how many purchases the chatbot completed. Pointing is not paying.
In an earlier sample, Accertify counted 7,741 agentic transactions out of 3.878 billion: roughly two per million. You will struggle to find them in the chart below; so did we. Its later autonomous-checkout account uses a different definition, so no, you cannot divide one by the other and call it a growth rate. [C11]
Accertify consortium · 1 Dec 2025–15 Mar 2026 · share of transactions (%)
0.0002% agentic
7,741 transactions
About two agentic transactions per million in this historical sample. The other 999,998 were not counted as agentic, which is not the same as untouched by AI.
The orange pointer locates a slice too small to see; its line width is not the share. “Remaining” does not mean no AI was involved. Counts and denominator come from supplied verified research; the denominator is rounded.
Source: Accertify, 26 Mar 2026 [C11]. Share calculated as 7,741 ÷ 3,878,000,000. This is not an estimate of total industry volume.
Amex calls its planned backing an industry first. Read the conditions and the order of events is the interesting part: an eligible agent error goes through a merchant return attempt before this extra protection applies. The merchant gets the mistake first; the industry first arrives later. [C7]
The planned protection covers eligible charges made by agents registered with Amex and integrated with its Agentic Commerce Experiences developer kit. The bot your cousin built over a weekend, unregistered, is not on the list.
Before anything else, the shopper has to start a return with the merchant, where possible. Congratulations to the returns desk, the new front line of artificial intelligence.
Claims are subject to eligibility and review. Full terms are “forthcoming.” Ordinary card disputes remain available, as they were before anyone said “agentic.”
Process, not a monetary waterfall. Source: Amex Agent Purchase Protection [C7a–e], as of 16 Oct 2026.
Shopify opens Shop MCP checkout. The agent completes the purchase once the shopper confirms, and the order lands in the merchant’s admin looking like any other order. The robot does not get a special badge. Shopify [C1]
Amex publishes Intent Intelligence. A record of what you authorised the agent to do, for fraud checks, authorisation and disputes. Receipts for your intentions, essentially. American Express [C5]
Accertify sizes the gap. Its new account separates the big pile of “AI shopping activity” from autonomous checkout, which it measures in basis points. Accertify [C10]
Latest developments in the supplied evidence, within the 18 July–16 October 2026 research window.
On 13 October, checkout moved into the agent and your intent moved into the payment record. Very modern. Neither, on its own, moves the cost of an agent’s mistake anywhere. Shopify’s agent orders still arrive as ordinary orders. Amex’s planned protection asks for a merchant return attempt first, where possible. So when the robot buys the wrong thing, the returns process hears about it first. That is not the same as the merchant losing the money: nobody has published who reimburses whom afterwards. The good news is that the volume is tiny. The bad news is that tiny is exactly when the conditions get written, and nobody reads conditions until they matter.
Adobe’s AI-referred visits went from converting worse than everyone else to converting better. By July 2026, their conversion rate was reported as 60% higher than non-AI visits. Lovely news for the channel. It says precisely nothing about what happens when an agent gets the order wrong. [C32]
United States retail · conversion rate relative to non-AI visits (%)
Zero means equal conversion rates. +60% means 1.6 times the non-AI rate, not 60 percentage points more conversion. These are reported snapshots, not proof that AI caused the improvement or completed the purchase.
Sources: Adobe Mar 2025 / Mar 2026 [C30] · Jul 2025 [C28]; via Digital Commerce 360 May 2026 [C31] · Jul 2026 [C32]. Five observations; no values interpolated.
Accertify’s measured autonomous-checkout share makes it hard to argue for a big burden today. These data do not show a merchant-loss crisis: basis points of anything are not a crisis yet. A consortium is not the whole market, either. [C10]
Starting a return tells you nothing about how it ends or what it costs. Amex’s conditional protection and plain old card disputes may still do the heavy lifting. Nobody has published an agent-error loss rate or a reimbursement split, and we looked. [C7]
Editorial checkpoints, not scheduled company announcements.
Recheck whether Amex has published final coverage conditions: who gets reimbursed, what the return attempt has to involve, and which costs are covered. “Forthcoming” is not a term.
Look for holiday disclosures of agent-completed transactions with consistent definitions. What would actually change the answer: a measured return or chargeback rate, next to realised losses.
31 scoped claims: 27 verified, 1 contested, 3 unverified. The 27 include an explicitly labelled interpretation and a derived calculation. The contested and unverified claims are excluded from the story and reel. “Verified” follows the supplied research protocol; it does not mean independently audited company data.
| Claim IDs | Evidence and publication decision |
|---|---|
| C1, C5, C7, C10 | Primary company material. Current October developments and return-first language rechecked. Protection is a future commitment, not universal live coverage. |
| C11, D1 | Historical counts retained from supplied verified research. Reopened source text states approximately one in 500,000; the exact counts were not freshly recovered. Derived share: 0.0001996%. |
| T1–T8 | Eight Adobe-origin growth snapshots with unequal, overlapping periods. Six directly primary-sourced; two via trade reports. Not a comparable monthly series. |
| V1–V5 | Five relative conversion gaps. Last two via secondary reports relaying Adobe. Observational, not causal. |
| I1 | Editorial interpretation: merchant return handling precedes conditional protection. Final loss allocation is not established. |
| G1–G4 | Excluded: seller-of-record assertion, reported absence of disputes, an unverified rule-change date, and unresolved cumulative Adobe figures. |
The supplied research contains no eight comparable primary observations for those series. Token counts and transaction shares have different denominators. General assistant users are not shopping users. Accertify’s March and October definitions differ. The first exhibit therefore presents eight labelled referral snapshots instead of inventing a continuous trend. No monetary liability split or agent-error return rate was found.
The photograph provides material context, not evidence of an agent purchase or return. Commons metadata confirmed permitted licences. The original is linked directly because asset-hosting credentials were unavailable. Social image is a 1200 × 630 crop of the first photograph.
