The Chargeback Problem Agents Created
| Dispute Category | Friendly Fraud Claim | UCP Cryptographic Defense | Evidentiary Artifact | Card Network Rule (Visa/MC) | Merchant Liability Shift |
|---|---|---|---|---|---|
| Unauthorized Transaction | Claim: ‘My autonomous AI assistant made this purchase without my permission.’ | Hardware-attested WebAuthn biometric keypair signature bound to transaction payload. | FIDO AP2 Signed Attestation Record with hardware enclave certificate. | Visa Compelling Evidence 3.0 & Mastercard Dispute Rules. | 100% liability shift to card issuing bank. |
| Item Not as Described | Claim: ‘The physical product delivered does not match what the agent evaluated.’ | SHA-256 Merkle root hash of product catalog specifications at transaction timestamp. | Signed Catalog State Snapshot timestamped in transaction envelope. | Card Scheme Merchandise Quality & Specification Guidelines. | Conclusive proof of exact merchant specifications agreed to at checkout. |
| Price / Total Dispute | Claim: ‘The total amount billed exceeded the quoted price.’ | RFC 9421 cryptographic signature over immutable checkout quote payload. | Asymmetric Quote Token with rate lock digest and decomposed taxes/shipping. | Truth in Lending Act (TILA) & Electronic Fund Transfer Act (EFTA). | Irrefutable mathematical defense closing the dispute instantly. |
| Cancelled Subscription Mandate | Claim: ‘I cancelled the recurring agent reorder mandate before renewal.’ | Verifiable revocation ledger timestamped in append-only cryptographic audit log. | Signed Revocation Receipt with microsecond timestamp (or lack thereof). | FTC Negative Option Rule & Visa Stored Credential Framework. | Clear proof that authorization mandate was active and unrevoked. |
| Duplicate Processing | Claim: ‘The merchant charged my account twice for the same autonomous order.’ | Unique cryptographic idempotency key enforced across payment processor rails. | Gateway idempotency transaction response log with identical digest. | ISO 8583 duplicate transaction processing rules. | Immediate automated dispute dismissal in merchant favor. |
Agentic commerce introduced a liability gap: when an AI agent completes a transaction autonomously, who bears the chargeback risk if the purchase was unauthorized, fraudulent, or disputed by the cardholder?
With agents making autonomous decisions, merchants lose that human signal—and must align dispute evidence with retry and escalation policies. A cardholder can claim “I never authorized this purchase” even though their agent did.
Visa and Mastercard have not published agent-specific chargeback evidence checklists beyond Visa’s CE 3.0 login-ID clause. The lists below are operational practices, not network rules. General background on consent and signed intent: consent management, verifiable intent.
How Chargebacks Happen in Agentic Systems
Vector 1: Unauthorized Agent Activation
A cardholder grants an agent access to purchase on their behalf. Later, they claim the agent was never authorized or that the authorization was revoked. The agent has no record of revocation and processes a legitimate transaction under the original grant.
Vector 2: Agent Overspending
A cardholder sets a budget cap: “Buy me groceries under $100.” The agent, interpreting “groceries” broadly, purchases $145 in items and crosses the threshold. The cardholder disputes the overage as unauthorized.
Vector 3: Policy Violation Disputes
An agent purchases restricted items (alcohol, age-gated products, hazardous materials) on behalf of a cardholder who then disputes the transaction. The chargeback code is “merchandise not received” or “unauthorized transaction,” not “age restriction violation.”
As an operational practice for restricted categories, a merchant can retain evidence that the cardholder explicitly consented to the agent purchasing age-gated items and that age verification was performed. No network rule in the programme’s sources establishes that standard KYC is insufficient, and no sheet settles a consent-proof mandate for restricted items.
Vector 4: Agent Impersonation
A fraudster compromises an agent system or device and uses it to make purchases on a victim’s account. The victim’s agent logs show legitimate API calls, but the initial compromise isn’t recorded. The chargeback defense collapses because the agent’s authorization chain appears clean.
Chargeback Prevention Architecture
Layer 1: Real-Time Agent Authorization Logging
Every agent action must be logged with:
- Timestamp (millisecond precision)
- Cardholder identifier (hashed for privacy)
- Agent identifier and version
- Explicit instruction captured (text or semantic representation)
- Merchant confirmation of instruction receipt
- Agent decision rationale (why this product, why this price)
- Device fingerprint and geolocation at time of request
Layer 2: Agent Revocation Protocol
Cardholder revokes agent access. This must trigger:
- Immediate notification to the agent system (API call, not email)
- Cryptographic signature from the cardholder’s identity provider
- Timestamp recorded in both merchant and payment processor logs
- Blocking of any pending transactions initiated before revocation
- Cardholder receives confirmation (SMS, push notification, email) within 30 seconds
Layer 3: Behavioral Anomaly Detection
Machine learning models flag suspicious agent behavior in real-time:
- Purchase patterns diverging from historical agent behavior
- Transactions in unusual geographies for that cardholder
- Price thresholds exceeded relative to cardholder’s typical purchases
- Rapid-fire transactions (bot behavior)
- Purchases of restricted categories not previously bought
When anomalies are detected, the system can:
- Pause the transaction and request human cardholder confirmation
- Log the anomaly in the authorization record (chargeback defense material)
- Adjust the agent’s authority for future transactions
Layer 4: Explicit Consent Capture
Before an agent makes a high-value or restricted purchase, capture explicit cardholder consent:
- Show the item, price, and merchant to the cardholder
- Require a signed approval (digital signature, biometric, or cryptographic token)
- Store the consent proof linked to the transaction
- For restricted items, verify cardholder eligibility (age, location)
Dispute Recovery Strategy
Pre-chargeback preparation (timing varies by network/process)
When a cardholder disputes a transaction, use this time to:
- Retrieve all agent authorization logs
- Verify the agent’s authorization grant is still active
- Check for any revocation requests
- Extract the cardholder’s explicit instructions
- Gather behavioral data (is this consistent with past behavior?)
Chargeback Evidence Package
Submit to the payment processor:
- Agent authorization grant (date, time, cardholder confirmation)
- Complete transaction log from agent to merchant
- Merchant confirmation of transaction completion
- Cardholder profile (purchase history, typical spend, geolocation)
- Agent decision rationale (why this item was selected)
- Proof of revocation receipt (if applicable)
- Behavioral anomaly analysis (or absence of anomalies)
- For restricted items: age/eligibility verification
Post-Chargeback Phase (if merchant loses)
If the chargeback succeeds, use the evidence to adjust agent policy:
- Lower the agent’s spending cap for that cardholder
- Require additional consent steps for similar purchases
- Implement stricter behavioral detection
- Flag the cardholder’s account for future disputes (not as punishment, but as risk indicator)
Compliance and Standards
November 2026 update: Per Sheet 5, agent-transaction disputes still resolve by challenge plus existing chargeback processes. The only settled agent-specific network term is Visa CE 3.0 Dispute Condition 10.4 (effective for transactions on/after 24 Oct 2026). Mastercard’s Verifiable Intent (SD-JWT + Key Binding, co-developed with Google, donated to FIDO Alliance) provides signed intent evidence but does not itself set chargeback outcomes. AP2 mandates (Sheet 2) are an authorization/evidence layer, not a settlement rail, and do not determine liability.
Visa Compelling Evidence 3.0 — Dispute Condition 10.4 (agent clause)
For transactions processed on or after 24 October 2026, an agent provider’s login ID can serve as one of the matching data elements tying an earlier undisputed transaction to a disputed one. [Sheet 5 item 12] Liability outcomes beyond that clause text are not settled — disputes resolve through challenge plus existing chargeback processes.
Frequently Asked Questions
How does FIDO AP2 biometric attestation stop ‘unauthorized purchase’ chargebacks?
Because the user’s private key resides inside their device’s hardware Secure Enclave and requires biometric authorization (TouchID/FaceID) to sign the transaction token, card issuers accept this as conclusive proof of user authorization.
What evidence must a merchant provide to win an AI agent merchandise dispute?
The merchant provides the UCP Signed Catalog State Snapshot, which proves the exact product description, dimensions, return policy, and pricing active at the millisecond the agent confirmed the order.
How do cryptographic idempotency keys prevent duplicate billing disputes?
Idempotency keys ensure that even if an agent retries an order due to network latency, the payment processor recognizes the unique transaction fingerprint and returns the original receipt without creating a second charge.

Leave a Reply