Amex Promised to Cover Agent Errors. It Is the Only Network That Could.
The first liability commitment in agentic commerce came from the only network that could make one alone.
The first liability commitment in agentic commerce came from a closed loop. On the open networks, the same sentence needs thousands of signatures.
On 14 April, American Express announced something that nobody else in this industry has announced in two years of trying: a commitment to absorb losses from AI agent error.
Alongside its Agentic Commerce Experiences developer kit, Amex said it will protect eligible Card Members from charges related to AI agent error, where the Card Member authorised the agent and the agent sent Amex the customer's authenticated purchase intent. The company calls it an industry first, and on that narrow claim I think it is right.
I have spent a year writing that the networks were shipping credentials, tokens, intent data and registries while refusing to say who pays when an agent gets it wrong. Somebody finally said it. That deserves to be acknowledged before anything else.
Then it deserves to be read carefully, because the sentence is doing less than the headline suggests, and the reason it came from Amex rather than from anyone else is the most interesting thing about it.
What was actually promised
The developer kit is five services. Agent Registration, which verifies agents so that only trusted ones transact on the Amex network. Account Enablement, which lets Card Members register cards for agentic transactions. Intent Intelligence, which captures purchase intent to support authentication, authorisation and disputes. Payment Credentials, which lets verified agents pay with tokenised credentials. And Cart Context, which shares cart details before or after a transaction to help validation, authorisation and dispute investigation.
Two of those five — Agent Registration and Cart Context — are marked as still in development.
Now the protection. Read it with the conditions attached, which is how it was written.
It is stated in the future tense. It applies to purchases made by registered agents, under a registration service that does not exist yet. It requires the agent to send authenticated purchase intent. It covers charges related to AI agent error, a term the release does not define. It covers eligible customers, a term the release does not define either. And it says terms and conditions will apply, without publishing them. No dollar limits, no claims procedure, no launch date.
I want to be fair about what that means. This is how a large institution announces a direction of travel before it has operationalised it, and announcing early is not dishonest — it signals to agent builders which network to build against, which is the point of a developer kit. Luke Gebb, who runs global innovation at Amex, framed it as making agentic commerce scale with transparency, accountability and confidence. That is a reasonable thing to be doing in April 2026.
But it should not be reported as protection that exists, and in several places it has been. What exists today is a commitment to build protection, contingent on a registration service still being built, covering a failure mode nobody has defined.
They named the right category
Here is where I want to give Amex genuine credit, because they aimed at the correct target.
An agent purchase can go wrong in three distinct ways, and I have written about this distinction before.
The agent was never authorised — somebody compromised the credential. That is fraud, the existing rules handle it, and nothing new is needed.
The agent was authorised and bought the wrong thing. The cardholder said buy a gift under two hundred dollars; the agent bought a gift under two hundred dollars that was wrong. The instruction was followed and the outcome is unacceptable.
The agent was authorised, bought exactly the right thing, and the human changed their mind. That is ordinary first-party misuse with a new excuse attached.
The second one is the gap. It has no reason code, no owner and no rule. And it is precisely what Amex has promised to cover.
The difficulty is that the boundary between the second and the third cannot currently be drawn.
Think about what adjudicating that claim requires. The cardholder says the agent got it wrong. The agent's log says it operated within the mandate. There is no neutral test that distinguishes the machine made a bad choice from I regret the choice the machine made on my behalf, because both produce identical evidence: an authorised agent, a valid mandate, a completed purchase and an unhappy customer.
Amex has committed to absorbing a category whose edge nobody can yet locate. That is a real commercial risk, and I suspect it is exactly why the terms are unpublished. Writing the terms means drawing that boundary, and drawing it is the hardest piece of work in agentic payments.
Why it had to be Amex
Now the part I think is the real finding, and I have not seen it written anywhere.
American Express could make this commitment unilaterally because American Express is, for most of its business, all of the parties at once. It is the network. It is the issuer. It is, substantially, the acquirer. When Amex decides that a loss will sit with Amex, there is nobody else in the room whose money is being allocated.
Visa and Mastercard cannot do that, and it is not because they are slower or less willing.
On an open network, a liability commitment is not a product decision. It is a reallocation of losses among thousands of independent institutions — issuers in a hundred countries, acquirers with their own economics, merchants with contracts — none of which the network employs, all of which get a vote, and many of which will object precisely because it costs them money. A sentence Amex can publish in a press release has to go through consultation, mandate drafting, an effective date and a compliance programme on the other networks.
So the first liability commitment in agentic commerce came from the only network structurally capable of making one alone.
That reframes a complaint I have been making all year, and I want to correct myself in public. I have been reading the open networks' silence on liability as reluctance. Some of it probably is. But a meaningful part of it is architecture: a four-party scheme cannot assign a loss by announcement, and arguably should not be able to, because the parties who would bear it are entitled to a say.
That is a better explanation than the one I was offering, and it changes what I think happens next.
What this does to everyone else
If Amex markets protection from agent error and the open networks cannot match it quickly, two things follow and both land on issuers.
The first is competitive. A card that covers agent mistakes is a straightforward proposition to sell against one that does not, and agentic volume will concentrate where the cardholder feels safest. That is Amex's commercial logic and it is sound.
The second is the one that worries me. Customer expectation moves faster than rules. Once any major brand has said agent errors are covered, cardholders on every other network will arrive at their bank expecting the same thing. The agent made a mistake, my card should cover it, I read that this is how it works now.
And an issuer in Panamá on an open network, asked that question across a counter in 2027, has no authority to grant it. Not reluctance — no authority. The rule that would permit it does not exist, and the issuer cannot write it alone any more than the network can.
So the honest forecast is an awkward middle period: one network offering a protection, several others unable to, and customers on all of them assuming the protection is a feature of cards rather than of a particular brand.
That resolves well if Amex's move creates enough competitive pressure to force the open networks into a rules process. It would be a genuinely good outcome, and it is the most plausible mechanism I have seen for this gap finally closing. It resolves badly if coverage stays a marketing claim on one network and an unmet expectation everywhere else, with issuers absorbing the goodwill cost of saying no.
From a Panamá desk
The direct exposure here is small and the indirect exposure is not.
Amex's footprint in most of this region is thinner than Visa's or Mastercard's, and in several markets it operates through licensed local issuers rather than directly. So the number of cardholders in Panamá who will actually hold a card with agent protection is modest.
That is the wrong thing to measure.
What crosses borders is not the product. It is the expectation. Our cardholders read the same coverage, watch the same channels and are sold the same ideas about what a card should do. A protection announced in New York becomes, within a year, something a customer here believes their card ought to include — and the first time they find out it does not is in a dispute they have already lost.
There is also a registration question this region should be asking now, while it is still cheap to ask.
Amex's Agent Registration service decides which agents may transact. Mastercard described a comparable registry last October. These registries are becoming the gate an agent must pass to touch the payment system at all, and every one of them is being designed in the United States, for agents built there.
So: who registers an agent built in Bogotá, for Colombian merchants, operating in Spanish, serving customers who shop regionally? Nobody has said, and I suspect nobody has been asked, because the people designing these registries have no particular reason to think of it.
That is not a complaint about intent. Registries get built for the users in the room. But if regional agents cannot pass registration, agentic commerce here will consist entirely of foreign agents buying from foreign merchants on our cards — which is, I note, the shape of every other wave of payments technology that has reached us.
The window to raise it is while the services are still in development, and the release says two of the five are.
My judgment
This is the most important announcement in agentic payments since the networks opened their rails a year ago, and it matters for a reason almost nobody has stated.
Not because the protection is operative — it is not. Not because the terms are good — they are unpublished. It matters because somebody with authority finally wrote a sentence in which a named party absorbs the loss, and that sentence now exists for everyone else to be measured against.
That is how liability standards actually form. Not through a committee reaching consensus from nothing, but through one participant with the structural freedom to move first, and competitive pressure doing the rest.
The structural freedom is the part I had wrong for a year. I read the open networks' silence as unwillingness. A good part of it is that a four-party scheme genuinely cannot do by announcement what a closed loop can do by decision. The silence was never going to break from the biggest networks. It was always going to break here.
What I will watch now is whether Visa or Mastercard opens a consultation on agent-error liability within twelve months. If they do, this announcement was the trigger, and the gap closes on a rules timetable. If they do not, we are heading into a period where the protection a cardholder believes they have depends entirely on which logo is on the card — and the explaining falls to us.
Three questions for the people who own this
1. When a cardholder tells you the agent made a mistake, what in your evidence distinguishes that from buyer's remorse? If the honest answer is nothing, then any protection you offer is uncapped and any protection you refuse is arbitrary.
2. If a competitor in your market advertises agent-error protection next year, what does your institution say to the customer who asks for it? That answer should be written now, by someone senior, rather than improvised at a counter.
3. Agent registries are being designed this year, and two of Amex's five services are still in development. Has anyone from your institution, your association or your regulator asked how an agent built in this region would register? If not, who would you expect to ask?
Reply to this email with a line or two. The sharpest answers will shape the next issue.
Sources
American Express, 14 April 2026 (Business Wire release) — the Agentic Commerce Experiences (ACE) Developer Kit and its five services: Agent Registration, Account Enablement, Intent Intelligence, Payment Credentials and Cart Context, with Agent Registration and Cart Context noted as still in development; and Amex Agent Purchase Protection, described as an industry first, under which, in the future, Amex will protect eligible customers from charges related to AI agent error where a Card Member authorised an AI agent to make a purchase and that agent sent American Express the customer's authenticated purchase intent, with terms and conditions to apply. Quotes from Luke Gebb, Executive Vice President and Head of Global Innovation. Link
The release carries no dateline in its body; the 14 April date comes from the Business Wire identifier and from Amex's own newsroom listing. The release does not define AI agent error or eligible customers, does not publish the terms and conditions, and gives no dollar limits, claims procedure or launch date. Those are absences I have noted rather than inferences I have drawn.
The argument that Amex could make this commitment because it is a closed loop, and that open four-party networks structurally cannot, is mine. Amex makes no such claim, and neither Visa nor Mastercard has said anything that confirms or contradicts it.