An agent buys an annual data subscription for €480. The purchase stays below its €500 limit, uses the approved supplier and produces a successful payment receipt. A week later, operations finds that the subscription excludes the export feature the team needed. The quotation the agent captured earlier included it. The final checkout did not. The purchase is fictional: we invented it to give a payment team an investigation to design.
Someone now has to reconstruct the offer, the instruction, the authorization and the completed purchase, and say which conclusions the records support. This post reviews recent work on agent payment evidence, walks through a worked dispute packet with eight records and six variations, and publishes the packet as a reusable template. It is part of Research notes on agent execution, a research series published in November 2026; the recovery and compensation note covers the other half of resolving an agent’s consequences.
Start with the disputed assertion
For our subscription, the disputed assertion is precise: “The purchased service included export access when the agent authorized payment.” The amount and supplier name cannot answer it. The investigator needs the applicable terms, their version and evidence that connects those terms to the checkout the agent paid.
The worked packet keeps six business records apart: the instruction, the captured quotation, the material sent to the model, the final checkout, the authorization and execution request, and the completion receipt. Two more rows cover a withheld item and a claimed commitment time. Each record has an actor, a time, its content and the question it leaves open. All facts in it are invented, and it contains no executable payment credentials.
From the packet, a reviewer can identify the discrepancy: the quotation includes export access and the final checkout lists a lower tier without it. Why the terms differ is a separate question. The supplier may have changed its offer, the agent may have used an expired quotation, or an intermediary may have removed a field. Each explanation needs further evidence, and an explanation written after the event becomes one more claim to investigate.
The worked packet, record by record
Case SYN-PAY-001 is synthetic. Operations required export access in a one-year subscription from a fictional supplier, Example Data, with a €500 ceiling. The investigator is a fictional internal operations reviewer authorized to inspect service terms and transaction records. Personal contact details are outside that reviewer’s scope.
- Supported within the example: the €480 amount sits under the €500 ceiling, the supplier matches, and Q2 excludes the export feature that P1 requires.
- Contradicted: a statement that the purchased service met the instruction.
- Unresolved: why Q1 and Q2 differ, whether the model received exactly P3, and whether P8’s claimed time holds.
| Record | Actor and time (UTC) | Content | What it supports; what stays open |
|---|---|---|---|
| P1 instruction | Operations manager, 09:00 | Annual subscription, approved supplier, export required, €500 ceiling | States the requirement in this case; no authentic identity evidence supplied |
| P2 quotation Q1 | Captured by agent, 09:02 | €480; export included; expires 09:05 | Shows the captured text; supplier origin and capture accuracy unverified |
| P3 model input M1 | Agent-side adapter, 09:03 | Selected Q1 text, including export and expiry | Shows what the example lists; completeness and provider receipt unverified |
| P4 final checkout Q2 | Fictional merchant, 09:06 | €480; annual Basic tier; export excluded | Supports a comparison with P1 and P2; no merchant signature supplied |
| P5 authorization and request | Agent, 09:07 | €480 to Example Data; references Q2 | Amount and supplier match the budget; the feature requirement conflicts with Q2 |
| P6 completion receipt | Fictional processor, 09:08 | Q2 paid; €480 | Records fictional completion; no settlement or network confirmation |
| P7 contact note | Operations manager, 09:00 | Item listed; personal contact detail withheld under the scope rule | Existence visible; contents unavailable to this reviewer |
| P8 commitment observation | Fictional holder or witness, 09:06:30 | Claimed commitment before P5 | Chronology illustrative only; verification status not_checked |
A stale quotation and a stale signed checkout are different objects
Protocol context matters to how the case is read. The Agent Payments Protocol (AP2) specification at v0.2.0, dated 28 April 2026, defines linked Checkout and Payment Mandates and associated receipts. It leaves outside its scope how the shopping agent establishes the user’s task. Its security guidance addresses a closed Checkout Mandate used with a different checkout session: “Merchant MUST verify that checkout_hash matches the hash of the latest checkout_jwt.” We cite both files at commit b4587ac, chosen to match the version the study below analyses; we do not assert it is the latest release. AP2 specification, security and privacy considerations.
Our scenario involves a quotation the agent captured three minutes before checkout. That raises a question about the context the agent used. It says nothing about whether an AP2 verifier would accept a stale signed checkout, and we ran no AP2 code or samples. AP2 checkout matching would require actual signed objects and verified roles, which the packet does not contain.
Experimental work studies the earlier stages. Beyond the Mandate, an August 2026 preprint by Avital Aviv and colleagues, analyses the security of AP2 v0.2 using a constructed testbed, including attacks on the context that exists before authorization. Its scanner combines static checks, comparisons across roles and adversarial probes. The authors limit how far the results transfer to future implementations and do not measure downstream effects on payment rails. We did not run the scanner, and the companion artifact linked from the paper could not be retrieved through our tools. Paper, version 1.
Keep the points where meaning can change
The packet keeps the captured quotation (P2) and the content actually supplied to the model (P3) as separate records. A summarising step could keep the price and drop the feature restriction. That possibility is part of our investigation design; we have not demonstrated it against a payment implementation. The data provenance guide describes how to record each read an agent makes and link it to the decision.
The packet also keeps the exact checkout approved for execution. Matching the amount to the budget answers one narrow question. Comparing the purchased service against the stated requirement answers another. If the requirement exists only in a chat message the investigator cannot access, the packet should record that gap.
Each suspected change should name an actor and a boundary. “The agent got confused” is too broad to guide a repair. “The checkout omitted a required feature, and the approval record contains no feature-level comparison” gives an investigator something specific to examine.
What a decision commitment can and cannot show
Agent Decision Commitments for Payments (ADC), by Fernando Henrique D. Milan and dated 15 October 2026, proposes committing to an agent’s decision record and releasing relevant evidence during a dispute. It is an individual Internet-Draft, revision 00, with intended status Experimental. An individual draft has no standing as an IETF standard. Its implementation status section reports a prototype holder implementation that is not publicly available, and states that no interoperability testing has taken place. Draft, revision 00, Datatracker record.
The draft is explicit about its limits. It says ADC establishes the integrity and timing of evidence, and lists what that leaves open: whether recorded inputs were authentic or accurate, and whether the context receipt is complete. It produces evidence and does not allocate liability among principals, holders, merchants, issuers, acquirers, networks or model providers.
- Integrity: a released record matches what was committed, once someone runs that check and records the result.
- Timing: the record existed by a stated time, to the extent the timing evidence is checked.
- Left open: the truth of the inputs, the completeness of the record and who bears liability.
Give missing and restricted evidence its own status
P8 in our packet is an illustrative commitment time. We assigned it no ADC attestation level, and we do not claim the rows satisfy the draft’s required fields, canonicalization, disclosure rules or verification steps. Its status is not_checked: no signature, salt, anchor or provider check took place. A real implementation would need to keep the material required to check any claimed timing relationship, plus the result of that check.
Restricted information needs the same care. P7 is a support note containing a personal contact detail. The investigator receives the service requirement, and the contact detail stays withheld with a recorded reason. The packet shows that the item exists without claiming anything about its contents.
A reviewer should be able to say “the amount matches”, “the terms differ”, “this item was withheld” and “this timing claim was not checked” as four separate findings. One green status covering all four would make the packet harder to use. The six variations below are conditional conclusions from invented records. No software produced them, and they report no protocol vulnerability.
| Case | Change | Supported observation | Still unavailable |
|---|---|---|---|
| Changed amount | P6 says €520 | P6 conflicts with the €480 request and the €500 ceiling | Which actor caused it; whether a real rail would accept it |
| Expired offer | Baseline: Q1 expires before Q2 | The captured offer was time-limited and the final terms differ | Supplier misconduct, or automatic invalidity of the payment |
| Missing input | P3 removed | The packet cannot show the listed model input | That the model never received Q1 |
| Delayed timing evidence | P8 moves to 09:09 | The illustrative observation follows authorization | Any independently established earlier commitment |
| Misleading source | Q1 promises a feature the supplier never offered | The packet keeps a false statement, by scenario definition | The truth or authorship of a matching record |
| Restricted detail | P7 stays withheld | The reviewer sees an unavailable item and its reason | Its contents or relevance without further authorized review |
Record risk scores and intent as evidence
Kayode L. Ogunsusi’s paper on the Autonomous Commerce Agent Guard System (ACAGS) appeared on arXiv on 20 October 2026; its journal issue is dated August 2026, so the upload is older work. It evaluates fraud screening and delegated-agent controls on synthetic transaction data with simulated agent characteristics. Its reported performance belongs to that experimental setting, and production latency, scale and live payment integration are outside its evaluation. We have not reproduced it. Paper, version 1.
In a dispute packet, a risk score becomes one more record: the model version, its inputs, the decision threshold and the action it informed. A favourable score leaves the service terms still to be inspected. This is our design recommendation; we have no measurement of its effect on disputes.
Payment standards bodies are also working on shared intent. EMVCo’s draft framework, announced on 1 September 2026, introduces Intent Services for managing authorized intent across a transaction’s lifecycle. Its Q3 newsletter of 9 October 2026 still describes the framework as a draft. We read the announcement and newsletter, did not review the technical framework, and make no compatibility claim. EMVCo announcement, Q3 2026 newsletter.
Test the packet with a reviewer who did not build it
The next exercise is small. Give the packet to an authorized reviewer who did not build it. Ask them to reconstruct the event from the rows alone, list conclusions as supported, contradicted or unresolved, and request the smallest additional evidence set. Record their role, time spent, the items they requested and any disagreements. The reviewer independence note covers why the reviewer’s access to evidence matters.
We have not run that exercise, and we report no comprehension or success rate. The packet is a manual analytical example. It establishes no real payment, protocol conformance, cryptographic verification, settlement, legal acceptance or liability determination, and it makes no claim about the outcome of any chargeback or dispute process.
Frequently Asked Questions
What evidence should you keep for an AI agent payment dispute?
Keep the instruction with its requirements, the quotation the agent captured, the content actually sent to the model, the exact checkout approved, the authorization and execution request, and the completion receipt as separate records. Give each an actor, a time and the question it leaves open. List withheld items with a reason, and mark any timing or signature claim with whether it was checked.
Does AP2 stop an agent paying against a stale quote?
AP2 v0.2 security guidance requires the merchant to verify that the checkout hash in the mandate matches the hash of the latest checkout object. That check concerns the signed checkout. A quotation the agent captured earlier is a different object, and AP2 leaves how the shopping agent establishes the user’s task outside its scope. We ran no AP2 code.
Is Agent Decision Commitments for Payments an IETF standard?
No. Revision 00, dated 15 October 2026, is an individual Internet-Draft with intended status Experimental. Its author reports a prototype that is not publicly available and no interoperability testing. The draft states that it covers the integrity and timing of evidence, and leaves input accuracy, completeness and liability unresolved.
Can an evidence packet decide a chargeback?
This post makes no claim about chargeback or dispute outcomes. Those depend on the parties and rules of the payment scheme involved, which we did not review. A packet helps an investigator separate supported, contradicted and unresolved assertions and request the missing evidence.
Key Takeaways
A useful dispute packet ties every conclusion to an identifiable record and keeps every remaining uncertainty visible. A digest, signature or timestamp needs a statement of what it supports and whether anyone checked it. Start with the dispute-evidence packet, and read the treasury payment release guide for the controls before a payment leaves.
Sources were checked on 26 October 2026 at the versions linked above. The preprints and the draft have not been independently reproduced. The worked packet is invented and was prepared by document inspection only; it contains no real payment, protocol test or KLA result.
