Most writing about agentic AI in finance stays at the level of demos. I wanted one document where a large financial company says, in writing, who pays when an AI agent gets a purchase wrong.
American Express published one on 14 April 2026. The release introduced a developer kit for agent payments. It also announced Amex Agent Purchase Protection, described as "an industry-first commitment to extend its backing to Card Member purchases made by registered AI agents across its network."
I read the core sentence slowly. "In the future, if a Card Member authorizes an AI agent to make a purchase and that agent sends American Express the customer's authenticated purchase intent , American Express will protect eligible customers from charges related to AI agent error."
Look at the first three words. The protection is a promise about the future. You cannot claim on it today.
The kit is called Agentic Commerce Experiences, or ACE. The release lists five services: Agent Registration, Account Enablement, Intent Intelligence, Payment Credentials and Cart Context.
Two of them carry an asterisk. The footnote reads "Services in development". It sits next to Agent Registration and Cart Context.
Amex's own page for the kit is more exact. "Account Enablement, Intent Intelligence, and Payment Credentials specifications are available as of 4/14/2026. Agent Registration and Cart Context specifications are still under development."
This detail matters more than it looks. The protection depends on a registered agent, and registration was the piece not yet finished on launch day.
I should be fair to Amex here. Shipping three of five specifications and saying so on the page is more candid than most launches I read.
I spent most of my time on the terms. They sit at the bottom of the agentic commerce page, under a footnote numbered 2.
They apply to U.S. credit cards from two Amex entities. The agent "must be registered with American Express and integrated with the Amex Agentic Commerce Experiences developer kit." American Express "must receive the Card Member-authenticated purchase intent from the agent."
Then comes the definition of a covered mistake. "Eligible agent errors include transactions where the purchase deviates from the Card Member-authenticated purchase intent."
Two conditions narrow it further. Protection applies "only after the Card Member has initiated a return (where possible) with the merchant". Claims may also fail "where purchase intent is subjective or non-verifiable (e.g. 'best' or 'really nice')".
That example made me stop reading. Then I read it again. It is very precise drafting.
An agent that buys the wrong size of a named product is covered. An agent asked for the best hotel is not. Nobody can check what best meant.
So how often do agents get things wrong? I wanted a measured answer, not a vendor claim.
The closest I found is the Finance Agent Benchmark, posted to arXiv on 20 May 2025 by Antoine Bigeard, Langston Nashold, Rayan Krishnan and Shirley Wu of Vals AI and Stanford. It has 537 expert-written questions, sorted by a task list agreed with people from banks, hedge funds and private equity firms, and each model gets tools including Google Search and EDGAR.
The result is blunt. The abstract says "even the best-performing model (OpenAI o3) achieved only 46.8% accuracy at an average cost of $3.79 per query."
The paper also notes that "no model surpassed 50% accuracy". The authors conclude that this "underscores the need for further advancements before reliable deployment in high-stakes finance settings."
The task list reads like a junior banker's week. One hard question in the financial modeling group, which holds 47 of the 537 questions, asks how much M&A firepower Amazon had at the end of fiscal 2024, counting cash, short-term investments and borrowing up to twice GAAP EBITDA.
I went looking for the hardest category in that table. This one stood out to me. My guess is that most analysts would argue over the borrowing assumption before they argued over the arithmetic.
The same paper has a second number that pulls the other way. The best model "averages just 3.1 minutes per task and costs $3.78, compared to human experts requiring 16.8 minutes and costing $25.66".
The cost gap is roughly sevenfold. The time gap is roughly fivefold. The authors add that the human figure leaves out time spent creating and reviewing questions.
Put both numbers side by side and the economics of agentic AI in finance become clear. Agents are cheap and fast. They are also wrong about half the time on analyst work.
Somebody has to write a liability rule for that. Amex wrote one for purchases and calls it industry-first, and I have not found an earlier one from a card network.
Here is the fact that cuts against the excitement. On 25 June 2025 Gartner predicted that "over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls."
It also named a habit it calls agent washing. This is the rebranding of assistants, RPA tools and chatbots as agents. "Gartner estimates only about 130 of the thousands of agentic AI vendors are real."
The poll behind the release is small but telling. Of 3,412 webinar attendees asked in January 2025, 19% said their organisation had made significant investments in agentic AI, 42% conservative investments, 8% none, and 31% were waiting or unsure.
Anushree Verma, a senior director analyst at Gartner, put it this way. "Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied."
I had assumed the Amex release meant agent payments were already flowing at scale. Read next to Gartner, it looks different. A network is laying down rules early, before most of the agents it will register even exist.
A short aside, because the Amex earnings release says something the product release does not.
On 23 April 2026 Amex reported first-quarter net income of $3.0bn and EPS of $4.28. Stephen Squeri, chairman and CEO, listed the ACE kit and "industry-first Agent Purchase Protection" among the quarter's achievements.
He also said Amex had "decided to increase our investments in marketing and technology". The cautionary note adds that technology spending "will depend in part on the company's success in advancing its agentic commerce initiatives".
So agent work now sits inside the guidance language, next to the 9 to 10 percent revenue growth target. I would not read more into it.
Back to the protection itself.
Amex describes its network as "operating as issuer, network, and acquirer". It sees the card member, the merchant and the transaction in one system.
I think this is why Amex could make the promise first. A four-party network would find it harder. When a dispute arrives, Amex can see the authenticated intent and the merchant record, and the cart details too once Cart Context is finished.
The release says Amex "believes Merchants will benefit from lower disputes and chargebacks". I can see the logic. I have not seen a single number on agent disputes yet, from Amex or anyone else.
I don't know how many registered agents exist today. The page says a full list is "coming soon". Nothing I opened gives a count.
What I can describe is the shape of the rule. Cover the verifiable mistake, exclude the subjective request, and make the customer try a return first.
It is a sensible design for agents that score 46.8% on expert tasks. I keep thinking about the gap it leaves. It still bothers me that what I suspect will be the most common request to an agent, find the best one, is the request the protection leaves out.