Disputes and Chargebacks When an AI Agent Buys: What Evidence a Merchant Will Need

Disputes and Chargebacks When an AI Agent Buys: What Evidence a Merchant Will Need

Picture an agent that buys a pair of trainers for a customer on Tuesday. On Friday the cardholder disputes the charge. The merchant's first question is not who bears the loss. It is simpler and more urgent: what can we put in front of our acquirer to show this purchase was what the cardholder asked for?

Most agentic commerce commentary stops at authorisation and liability. That debate matters, but it leaves merchants and platforms without an operating answer. This guide is the operational follow-on. It sets out the three records a dispute response is likely to lean on, how a dispute may flow, and what to log from the first live transaction. It is not legal advice, and it does not quote scheme dispute rules, because we have not reproduced any here.

The question has moved from "can agents pay" to "how is authority proven"

The industry conversation has shifted. Innovify's October market review, drawing on Finextra, The Fintech Times, PYMNTS, Innovation & Tech Today and Payments Dive, found that who authorises and who is liable for an agent's transaction is now the dominant agentic-commerce theme, for a third consecutive monthly run. The debate has moved from whether agents can pay to who is liable and how authority is proven.

Two other October developments show why proof matters. JPMorgan Chase blocked outside AI agents from direct access to customer accounts and backed open-banking routes instead, as reported by Finextra, citing CNBC, in mid-October 2026. Large institutions are choosing sanctioned, authorised access over agents acting with customer credentials. And consumer surveys reported by Yahoo Finance UK (carrying a Wildfire Systems release) and Forkast News show that willingness to use AI to research a purchase runs ahead of willingness to let it complete one. Those surveys are vendor-commissioned, so treat them as directional rather than precise.

Read together, they describe a trust gap. Customers are cautious about delegating payment, and institutions are cautious about agents acting for customers. A dispute is where that caution becomes a cost: someone has to show that the customer did hand over authority, and that the agent stayed within it.

Who bears the loss is a separate question. It is the subject of our guide on who authorises, and who is liable, when an AI agent pays, and we do not repeat that argument here. This article assumes you are working that through with your acquirer, your agent partners and your advisers. Whatever answer you reach, you will need evidence to act on it.

What we know, and what we do not

It is worth being plain about the limits of the public evidence before going further.

What is reported: three card networks now have agent-intent frameworks. American Express published an "Intent Intelligence" framework linking authenticated purchase intent to agent transactions, according to The Paypers on 14 October 2026. It sits alongside Visa Intelligent Commerce and Mastercard Agent Pay. Mastercard also opened Early Access to Agent Pay for Machines, with buyer-side and seller-side APIs and a sandbox, aimed at agents paying for services (Mastercard, The Block and Decrypt, October 2026).

What is not in our evidence: how any scheme will decide a disputed agent purchase, which dispute categories apply, or what documents a scheme will accept. The American Express framework is vendor-authored and we are describing it only at the level The Paypers reported. Read each scheme's own material, and ask your acquirer for its current dispute rules, before you rely on anything in this article.

So the sections that follow are Innovify's analysis and a practical framework, not reported scheme rules. We have kept them to what is logical from first principles: a dispute asks whether the buyer agreed, and your job is to be able to show the chain of agreement.

The three records: mandate, intent and checkout

A purchase made by an agent has three links in its chain of authority. A dispute response will be stronger if you can show each one, and weaker wherever one is missing.

Record one: the mandate

The mandate is the delegation itself: the moment the cardholder gave an agent authority to act for them. This is the record most merchants will never see unless they plan for it, because it usually sits with the agent provider or a wallet, not with the merchant.

A useful mandate record shows:

• Who delegated. The authenticated customer or account, and how they were authenticated when they granted authority.
• Which agent received it. A stable identifier for the agent and the organisation operating it, not a display name.
• What the authority covered. Spending limits, categories, permitted merchants, number of purchases, and the period it stays valid.
• When and how it was granted. A timestamp and the screen or flow the customer used, including any terms they accepted.
• Whether it was changed or revoked. A dated history, so you can show what was in force on the day of the purchase.

The mandate answers the first question a disputed purchase raises: did the cardholder authorise this agent to buy at all?

Record two: the intent

Intent is narrower than the mandate. It is the specific request that led to this particular purchase. A customer who delegates authority to buy groceries has not necessarily asked for a particular basket on a particular day.

The American Express framework, as The Paypers describes it, links authenticated purchase intent to agent transactions. That framing is useful whatever scheme you work with, because it names the gap: a general mandate proves the agent was allowed to act, but intent shows the agent acted on what the customer wanted.

A useful intent record shows:

• The instruction. What the customer asked for, in their words or selections, captured at the time.
• The constraints. Price ceilings, size, colour, delivery date, and what the agent was told to do if an item was unavailable.
• The agent's interpretation. What the agent decided to buy and why it matched the instruction, including any substitution it made.
• Any confirmation. Whether the customer approved the final choice before purchase, and how.

In our analysis, intent is where most agent disputes will be won or lost. A customer who says "I never asked for that" is making a claim about intent. Without a record of the instruction and the agent's reading of it, a merchant has little to answer with.

Record three: the checkout

The checkout record is the part a merchant controls directly. It shows what was offered and accepted at the moment of payment.

A useful checkout record shows:

• What the agent was shown. Product, price, taxes, delivery charges and the terms on the page or in the interface the agent used.
• How the agent presented itself. The agent identifier, session and any token or credential supplied with the payment.
• What was agreed. The order, the confirmation sent, and timestamps for each step.
• What happened after. Dispatch, delivery confirmation, and any customer contact about the order.

This overlaps with what merchants already keep for ordinary online orders. The new element is the agent's identity and the link back to the mandate and the intent.

How an agent-purchase dispute may flow

The route a dispute takes depends on the scheme, the acquirer and the type of claim. What follows is a generic sequence to plan against, not a description of any scheme's rules.

1. The cardholder raises a claim with their card issuer. The reason they give may be that they did not authorise the purchase, that the agent bought something other than what they asked for, or that the goods did not arrive or were not as described.
2. The claim reaches the merchant through the acquirer. There will be a deadline, and your acquirer will tell you what it is.
3. The merchant assembles a response. This is where the three records matter.
4. The response is assessed and an outcome follows. Who assesses it, and on what basis for an agent-made purchase, is a question for your acquirer and the scheme's published rules.

The point worth taking from this sequence is a structural one. A dispute on an agent purchase may turn on two questions at once: did the customer delegate, and did the agent act within that delegation? The merchant normally holds only the checkout record, while the mandate and the intent sit with other parties. That is a design problem and a contract problem, not only a dispute-handling one.

Who holds which evidence

Map the three records to the parties involved, and you will quickly see where the gaps are.

• The merchant or platform holds the checkout record, and perhaps the order and delivery history.
• The agent provider is the likeliest holder of the intent record: the instruction, the interpretation and any confirmation step.
• The wallet or credential provider may hold the mandate, or the authenticated grant of authority, depending on how the agent is set up.
• The acquirer and the payment service provider hold the transaction data and will define the dispute process you follow.

None of this is settled by a published standard. The practical step is to agree, before launch, what each party will capture, in what format, and how quickly it can be handed over if a dispute arrives. Ask for it in the commercial agreement, not in the middle of a dispute with a deadline running.

What regulators have said so far

On the UK regulatory position, we can offer only an outline, and the primary source should be read directly. In a speech on 19 October 2026, the FCA signalled that it is applying existing frameworks, including Consumer Duty and senior-manager accountability, to AI rather than creating new rules for agents, and set expectations that AI providers add disclaimers to AI financial advice, as reported in the speech itself and by Finextra.

For a merchant, the sensible reading is modest. Regulators are looking at how firms already under their rules use AI, which makes a clear record of what an agent did and why useful in itself. What that means for your own circumstances, including any obligations you or your payment partners carry, depends on your business model and on who in your chain is regulated. This is not legal advice. Take advice from a qualified adviser, and check the FCA's own material.

What to log from the first live transaction

You do not need a perfect system to start. You need a record you will wish you had kept. A working checklist:

1. A stable agent identifier on every transaction, mapped to the organisation operating the agent.
2. A reference to the mandate in force at the time, with its limits and validity period.
3. The customer's instruction, captured as given, with a timestamp.
4. The agent's interpretation and any substitution, in plain language you could show a non-technical reviewer.
5. The price, terms and options the agent was shown at the point of purchase.
6. Any customer confirmation step, including how and when it was given.
7. The payment credential or token used, and how it was issued.
8. Order, dispatch and delivery events, linked to the same transaction identifier.
9. Customer contact about the order, so a later claim can be set against what the customer said at the time.
10. A change history for mandates, limits and agent settings, so you can show what was in force on a given day.

Two cautions. First, logging customer instructions and agent activity involves personal data, so agree what you hold, why, and for how long with your data protection adviser. Second, a log nobody can retrieve under time pressure is not evidence. Test that you can assemble a full record for one transaction within the time your acquirer allows.

Keep your evidence portable while standards multiply

Agent-commerce standards are multiplying without converging. October coverage lists Google's Universal Commerce Protocol, Mastercard's Agent Pay for Machines, American Express's Intent Intelligence, and the x401 and x402 approaches, and our market intelligence found no sign that any of them is displacing the others. Agent identity and authorisation tooling is multiplying too: FF News, Business Wire and PYMNTS reported several new offerings in a single week in October, and no single standard has emerged.

For anyone designing a dispute record, that is a reason for caution. Clients that commit early to one scheme's approach risk rework. The more durable move, and the one we would take on any agentic commerce and payments build, is to design an evidence record that is scheme-neutral at its core: the mandate, the intent and the checkout, each with its own identifier and timestamp, that can be mapped to whichever framework a given scheme or acquirer asks for.

There is also a useful signal in where agent payments are landing first. The Visa, BILL and Billtrust live pilot of agents paying business-to-business invoices through Visa Intelligent Commerce was reported by Finextra in October 2026, and "first" in that story is Visa's own claim. Reporting from Visa, PYMNTS, TechCrunch and Finextra suggests accounts payable and finance operations are the first beachhead, because approvals, supplier relationships and audit trails already exist there. That is our reading as well: delegated payments are easier to authorise, and to evidence, where a paper trail is already part of the process. Open-web consumer checkout lacks that scaffolding, so it has to be built deliberately. The approval logs and audit trails of accounts payable are a good template.

Where Innovify fits

Innovify is a technology partner with deep expertise in banking, financial services, fintech and insurance. We help mid to large businesses build AI-native products, modernise their technology stack and launch fintech products, and we design compliance in from day one. We build for regulated clients; Innovify is not itself FCA-authorised or regulated.

For this topic, that means working with commerce and embedded finance teams on the practical layer this article describes: designing the mandate, intent and checkout records so they are captured at the point of purchase, shaping the evidence trail so it can be retrieved when a dispute arrives, and keeping it portable as schemes and standards change. Our Agentic Commerce & Payments work is where this sits. If you are deciding how your agent purchases should be recorded before they go live, start with one product line and one agent, and work backwards from the dispute response you would want to send.

FAQ

What evidence does a merchant need to defend a dispute on an agent-made purchase?

A merchant needs to show the chain of authority from customer to purchase: the mandate, the intent and the checkout. In practice that means a record of who delegated authority to which agent and within what limits, the customer's instruction and the agent's interpretation of it, and what was offered, accepted and delivered at checkout. Exactly what a scheme or acquirer will accept is set by their own rules, so confirm it with your acquirer.

Is a disputed agent purchase treated differently from an ordinary online card payment?

We cannot say, because our evidence does not include any scheme's published dispute rules for agent purchases. Reported developments show that the card networks are publishing agent-intent frameworks, which suggests that proof of delegated intent will matter. Ask your acquirer how it expects to handle disputes on agent-made purchases, and read each scheme's own material.

What is American Express's Intent Intelligence framework?

It is a framework American Express published that links authenticated purchase intent to agent transactions, as reported by The Paypers on 14 October 2026. It joins Visa Intelligent Commerce and Mastercard Agent Pay as a card-network approach to agent payments. The report is vendor-authored coverage, so read American Express's own documentation for the detail.

Should merchants wait for a single standard before capturing this evidence?

No, but they should avoid tying the record to one scheme. Standards in agent commerce are multiplying and none has emerged as dominant, so a scheme-neutral record of mandate, intent and checkout can be mapped to whichever framework you later need. Waiting means the first disputes arrive with no record to rely on.

Has the FCA set specific rules for AI agent payments?

According to the FCA's speech of 19 October 2026 and Finextra's reporting, the FCA is applying existing frameworks, including Consumer Duty and senior-manager accountability, to AI rather than creating new rules for agents. That is an outline only and not legal advice. Read the FCA's own material and take advice on your circumstances, including who in your payment chain is regulated.

Conclusion

The industry has spent months on who authorises and who is liable. The next test is practical: when a customer disputes an agent-made purchase, can the merchant show what was delegated, what was asked for and what was bought? The public evidence on how disputes will be decided is still thin, which is the reason to act on what is within your control. Capture the mandate, the intent and the checkout. Agree who holds each record. Keep the structure portable. Then read your acquirer's and each scheme's current rules, and take advice for your own circumstances.