Limited time BFCM offer for prevent 2.0
Limited time BFCM offer for prevent 2.0
Limited time BFCM offer for prevent 2.0
Limited time BFCM offer for prevent 2.0
/
Litiges et rétrofacturation
7 juillet 2026
Nov 17, 2026

Responsabilité en matière de rétrofacturation des agents IA : qui paie lorsque le commerce par agent tourne mal ?

Logo circulaire blanc comportant, en son centre, des formes entrelacées, entouré de lignes elliptiques qui se chevauchent et ressemblent à des orbites, ainsi que de losanges bleus dispersés.

rétrofacturation?
Ce n'est plus votre problème.

Recover 4x more chargebacks and prevent up to 90% of incoming ones, powered by AI and a global network of 20,000+ merchants.

Plus de 600 avis
Aucune carte bancaire n'est nécessaire.
Responsabilité en matière de rétrofacturation des agents IA : qui paie lorsque le commerce par agent tourne mal ?

En bref :

  • A disputed agent purchase follows ordinary card-not-present rules, so the merchant pays unless it wins the dispute or a tokenized, authenticated network flow shifts fraud liability to the issuer.
  • Visa's October 2026 Core Rules (4.1.23.10) make the cardholder responsible for a registered agent provider's actions, but set no agent dispute procedure.
  • No rule makes the agent platform liable: the merchant stays merchant of record, and American Express's agent-error protection covers cardholders, with no stated merchant treatment.
  • Agent disputes still count toward Visa VAMP and Mastercard ratios, even when a vendor guarantee pays them back.
  • Tag agent orders, keep the agent identity, token and consent record, and pick evidence by reason code.
Chargement du lecteur AudioNative de synthèse vocale d'Elevenlabs…

AI agent chargeback liability asks who takes the loss when an agent's purchase is disputed: the cardholder, the agent platform, the issuing bank or the merchant.

The card networks' public dispute rules were written for human buyers. They say little about agents. A disputed agent purchase mostly follows the rules for online card sales. The merchant carries the loss unless it wins the dispute or authentication moves fraud liability to the issuer.

Choose an agentic commerce guide

This guide covers who pays and which rules decide it. For what counts as an approved agent purchase, and which dispute type each case fits, see our guide to agentic commerce chargeback liability.

What agentic commerce means for chargebacks

Agentic commerce is when AI agents research products, compare options and buy for a consumer on their own. The consumer sets preferences and gives permission. The agent does the rest, often without asking before each purchase.

Cela revêt une importance particulière pour rétrofacturation , car toutes les règles en vigueur en matière de litiges ont été conçues pour un modèle à deux parties : un acheteur et un vendeur. L'un décide d'acheter. L'autre exécute la commande. En cas de problème, les réseaux de cartes bancaires disposent de procédures claires pour déterminer la responsabilité de chacune de ces deux parties.

Les agents IA bouleversent ce modèle. Désormais, un tiers intervient dans les décisions d'achat, et les mécanismes existants de contestation des paiements n'en tiennent pas compte. L'agent n'est ni l'acheteur, ni le vendeur. Pourtant, c'est lui qui a initié la transaction.

The volume is already visible. Visa said in October 2025 that AI-driven traffic to US retail sites had grown 4,700% in a year (Visa press release). The card networks have since launched agent programs, covered below. For the wider picture, see the impact of agentic commerce on fraud and chargebacks and how AI is changing online payments.

Pour les commerçants, cela représente une menace directe. Vous avez accepté une commande qui semblait légitime, vous l'avez honorée de bonne foi, et vous vous retrouvez désormais confronté à un litige dans lequel le consommateur affirme n'avoir jamais validé l'achat. Les règles censées vous aider à vous défendre n'ont pas été conçues pour ce genre de situation.

Where agentic chargeback liability is unclear

Lorsqu'un agent IA effectue un achat qui fait l'objet d'une contestation, trois parties pourraient être tenues pour responsables.

Le consommateur a délégué son pouvoir à l'agent. Il lui a donné l'autorisation de faire des achats, de fixer un budget ou de définir ses préférences. Mais a-t-il autorisé cet achat en particulier ? La réponse à cette question est rarement claire.

C'est le fournisseur d'IA qui a développé et qui exploite l'agent. C'est lui qui a conçu la logique de prise de décision. Si l'agent interprète mal une préférence ou effectue un achat que le consommateur ne souhaitait pas, c'est le produit du fournisseur qui est à l'origine du problème.

Le commerçant a accepté la commande et l'a honorée. Vous avez traité une transaction qui semblait valide, expédié le produit, et vous êtes désormais confronté à un rejet de débit.

The issuer and acquirer sit behind them. The issuing bank decides whether to credit the cardholder and which dispute type applies. The acquirer passes the cost to the merchant through the processing agreement. Neither takes part in the agent's decision, but both decide how the dispute is routed.

Pour déterminer qui est responsable des achats d'agents d'IA, il faut examiner le rôle de chaque partie, ce qu'elle contrôle et où se situe son exposition au risque.

Open a section to compare the details.

Fête
Rôle dans la transaction
Ce qu'ils contrôlent
Financial Exposure Today
Consommateur
Rôle dans la transaction
Pouvoirs délégués à l'agent
Ce qu'ils contrôlent
Preferences, budget, scope of the mandate
Financial Exposure Today
Can dispute the charge; card law caps liability for unauthorized use, but a delegated purchase may not count as unauthorized
Agent platform
Rôle dans la transaction
Built and runs the agent
Ce qu'ils contrôlent
How the agent interprets instructions and records purchase intent
Financial Exposure Today
No chargeback liability under card network rules; exposure comes from its contracts with consumers and partners
Commerçant
Rôle dans la transaction
Commande acceptée et traitée
Ce qu'ils contrôlent
Product, delivery, policies and the evidence it keeps
Financial Exposure Today
Loses the sale, pays dispute fees and counts the dispute toward monitoring ratios, unless it wins representment or the issuer carries fraud liability
Émetteur
Rôle dans la transaction
Holds the cardholder's account
Ce qu'ils contrôlent
Whether to credit the cardholder and which dispute category applies
Financial Exposure Today
Funds the credit first and carries fraud losses on authenticated transactions where the liability shift applies
Acquirer or PSP
Rôle dans la transaction
Processes the merchant's card payments
Ce qu'ils contrôlent
Merchant onboarding, reserves and monitoring-program enforcement
Financial Exposure Today
Passes dispute costs to the merchant and carries the residual loss if the merchant cannot pay

No current framework clearly splits the blame among these parties. No rule says the AI provider is liable when its agent made the purchase. No regulation says a consumer who gave broad permission cannot dispute single purchases.

The result is a liability gap. And right now, merchants sit in the most exposed position. When a chargeback lands, you lose the revenue and pay the chargeback fees and the cost of fighting the dispute. That holds whether the AI agent or the consumer caused the problem.

Who Is the Merchant of Record in an Agent Purchase?

The merchant that sells the goods and takes the payment stays the merchant of record. The agent platform does not take that role. OpenAI's production guidance for its Agentic Commerce Protocol says the seller is the merchant of record. OpenAI and payment service providers are not. The merchant manages refunds and chargebacks.

The Universal Commerce Protocol says the same. Businesses stay the merchant of record and own the customer relationship. So the protocols behind agent checkout also confirm where disputes land: on the seller.

Pourquoi les règles traditionnelles en matière de rétrofacturation ne fonctionnent plus

Chargeback rules were built on two ideas: authorization and evidence. Both break down in AI-agent orders.

L'autorisation devient ambiguë

In a normal sale, consent is simple. The cardholder entered their payment details and clicked "buy." That action is the authorization, and it ties directly to a specific purchase.

With AI agents, authorization gets layered. The consumer authorized the agent to act on their behalf. But that broad permission does not cover every purchase the agent makes. Say a consumer tells an agent "buy me groceries this week." The agent orders a premium item the consumer did not expect. Was that purchase authorized?

Current card network rules have no clean answer. The authorization happened at the agent level, not the transaction level. When the consumer disputes the charge, the merchant must prove consent for a purchase the consumer may not have seen until it arrived.

Consumer law draws the same line. Under Regulation Z, unauthorized use of a credit card means use by someone with no actual, implied or apparent authority. Regulation E also leaves out transfers by a person the consumer gave access to the account.

A consumer who chose to give an agent buying power has arguably given it that authority. So a dispute over what the agent bought may not count as "unauthorized." It may be handled as a billing error or merchant dispute instead.

Regulators have not applied these definitions to AI agents, so treat this as an open question rather than settled law.

Les preuves traditionnelles disparaissent

Merchants fight chargebacks with evidence: IP addresses, device data, browsing history and session data. These show a real person used your site and chose to buy.

AI agents erase all of that. An agent often does not browse your site the way a human does. When it does, its device and network data belong to the agent platform, not the cardholder.

It leaves no session history that shows the shopper comparing and choosing. The behavior signals merchants use to win disputes are missing or misleading on agent orders.

Here is what changes when you move from traditional disputes to agentic commerce chargebacks:

Type de preuveTra­di­tional Trans­ac­tionAI-Agent Trans­ac­tion
Empreinte numérique de l'appareilAccessible depuis le navigateur ou l'application du clientOften absent, or reflects the agent platform instead of the cardholder
adresse IPEn fonction de la localisation du clientConnecté au serveur de la plateforme d'IA
Historique de navigation et de sessionPrésente les étapes de recherche de produits, de comparaison et de finalisation de la commandeLittle or no browsing behavior to capture
Activité du panierLe client a ajouté des articles, a vérifié son panier et a confirmé sa commandeAgent assembled the order pro­gram­mat­i­cally
Signaux d'authentification3DS, correspondance du CVV, vérification AVS3DS often cannot run without a cardholder device; agent programs substitute signed agent identity and tokenized credentials
Éléments attestant de l'intention d'achatLes habitudes de navigation et le temps passé sur la page témoignent d'un choix délibéréNo human decision-making signals; a stored mandate or intent record is the substitute

Cela vous met dans une situation délicate. Vous avez besoin de preuves pour contester un rejet de débit. Les meilleures preuves proviennent du comportement humain. Or, cette transaction ne comportait aucun élément de comportement humain à analyser.

Which Rule Decides Who Pays Today

Before assigning the loss

Three checks before deciding who pays

What is the cardholder claiming?

Separate a claim of unauthorized payment from a complaint about the item, price, or delivery. The actual claim determines the dispute evidence.

How was the payment authenticated?

Check the sale's sign-in and token records, and whether the fraud liability rules were met. An agent label or signature alone does not shift liability.

Which terms cover this transaction?

Confirm the network program, the payment rail and your acquirer or platform contract. Agent-error cover and consumer disputes can be treated differently from fraud.

Use the transaction paths below to assess the case. Current rules for authenticated payments may assign fraud risk. That does not settle every complaint about an agent's choice.

A May 2026 legal analysis by FBT Gibbons says Visa's rules now cover agent sales directly. They need identity checks under Visa Intelligent Commerce and a provisioned token. Mastercard's public Transaction Processing Rules say nothing on agents. It handles them through programs such as Agent Pay and agentic tokens.

The analysis also notes that 3-D Secure struggles with agent traffic. Agents lack the device data and biometrics it expects.

Open a section to compare the details.

Transaction Path
Likely Dispute Category
Who Bears the Loss by Default
What Shifts It
Agent purchase through a network agent program (verified agent, tokenized credential)
Likely Dispute Category
Fraud, if the cardholder claims no authorization (Visa 10.4, Mastercard 4837)
Who Bears the Loss by Default
The issuer typically carries fraud risk on authenticated, tokenized transactions
What Shifts It
Keep the signed agent identity and token data as evidence
Agent purchase with card details entered at checkout, no network program
Likely Dispute Category
The same fraud codes, or consumer dispute codes if the cardholder accepts the agent acted
Who Bears the Loss by Default
The merchant, as with any card-not-present sale
What Shifts It
A completed 3D Secure authentication moves fraud liability to the issuer, but agent flows often cannot complete it
Cardholder accepts the agent acted but disputes the outcome (wrong item, quantity or price)
Likely Dispute Category
Consumer dispute (Visa 13.3 Not as Described, Mastercard 4853 Cardholder Dispute)
Who Bears the Loss by Default
The merchant, unless it shows the order matched the mandate and the listing
What Shifts It
Order confirmation, mandate record, delivery proof and return policy
Purchase covered by a card brand's own agent protection (American Express)
Likely Dispute Category
Cardholder claim handled under the brand's program terms
Who Bears the Loss by Default
Not published in detail; the announcement covers cardholders and is silent on merchants
What Shifts It
Check your acquiring agreement and the card brand's merchant terms
Payment on a rail without chargebacks (stablecoin settlement)
Likely Dispute Category
None by default
Who Bears the Loss by Default
Whoever holds the funds; no reversal or Regulation E protection by default
What Shifts It
Contract terms with the wallet or protocol; see cryptocurrency chargebacks

Reason codes describe the cardholder's claim, not the agent's role. The same agent purchase can arrive as a fraud code or a consumer dispute code, based on what the cardholder says. Review the Visa chargeback reason codes before you decide which evidence to build.

In a July 2026 analysis, Worldpay's head of agentic commerce put the open question plainly. When agents sign in through network programs, liability follows current token rules. The hard cases are everything short of outright fraud, such as an agent that misread an order.

For the general rules on who carries fraud losses, see who bears the loss in payment fraud.

What Visa's October 2026 Rules Say About Agent Purchases

Visa's rules now make the cardholder responsible for what a registered agent provider does. The Visa Core Rules and Visa Product and Service Rules (24 October 2026 edition) set this out in section 4.1.23, Agentic Platform Requirements. Rule 4.1.23.10 says the cardholder answers for any action an Agentic Payment Provider takes, as if the cardholder made the payment.

The section sets duties for agent providers and the wallets that enable them. It does not set a dispute procedure. A disputed agent purchase still runs through the normal Visa dispute conditions. Visa does not say how an issuer should weigh the rule in a given case.

Visa ruleWhat it requiresWhat to keep in your dispute file
4.1.23.3 Pre-transactionCardholder consent and identity checks before an agent stores a card or acts on instructions; instructions carry an expiry dateThe consent and expiry record the agent passed with the order
4.1.23.4 ProcessingThe agent buys only within criteria the cardholder set, passes cardholder details to the merchant and uses the provisioned tokenThe token data and a comparison of the order with the stated criteria
4.1.23.8 Post-transactionAn order confirmation available for at least 120 days from the processing dateYour own confirmation, kept for at least the same window
4.1.23.9 Repeated transactionsCardholder consent to repeat orders, which run until changed, cancelled or expiredThe consent date and expiry for each repeat agent order
4.1.23.10 Cardholder responsibilityThe cardholder is responsible for any actions of the Agentic Payment ProviderProof the order came from a registered agent provider, to answer an "I did not authorize this" claim

Visa can also disqualify an agent provider from its program (4.1.23.11). Earlier editions numbered this section 4.1.24.

Comment les agents IA font naître de nouveaux risques de rétrofacturation pour les commerçants

Les agents d'IA ne se contentent pas de bouleverser les processus actuels de contestation de paiement. Ils créent des scénarios de litige entièrement nouveaux auxquels les commerçants doivent se préparer.

Achats non autorisés liés à l'IA

Un agent IA suit les habitudes d'achat d'un consommateur et passe une commande en fonction de ses préférences passées. Le consommateur reçoit un produit qu'il n'a pas explicitement demandé et conteste le prélèvement. Du point de vue du consommateur, il n'a jamais approuvé cet achat en particulier. De votre point de vue, l'agent a passé une commande valide en utilisant des informations de paiement valides.

This is the most straightforward AI-agent dispute scenario. Whether the claim holds depends on what the mandate covered. If the order falls inside it, the dispute is closer to buyer's remorse than fraud. If it falls outside, the consumer did not approve it.

Without a stored mandate record, you cannot show which case you are in. Expectation gaps cause the same disputes when an assistant's product summary does not match the listing. See AI shopping chargebacks for how that changes dispute volume.

La fraude « amicale » devient plus difficile à combattre

La fraude amicale constitue déjà l'un des principaux défis auxquels sont confrontés les commerçants en matière de rétrofacturation. Les agents basés sur l'IA ne font qu'aggraver la situation.

Il peut arriver qu'un consommateur ne reconnaisse véritablement pas un achat effectué par son mandataire. Il consulte son relevé de carte bancaire, constate un prélèvement dont il ne se souvient pas avoir donné l'autorisation, et dépose une réclamation. Il ne s'agit pas là d'une fraude malveillante. C'est un malentendu, qui, du point de vue du commerçant, ressemble en tout point à une « fraude amicale ».

The difference is that normal friendly fraud at least leaves behavior you can use in a dispute. You can show the buyer browsed your site, added items to a cart and checked out. With agent purchases, that trail does not exist. That makes these disputes harder to fight either way.

Les programmes de surveillance des réseaux de cartes de paiement menacés

Every merchant lives with chargeback ratio thresholds set by the card networks. Visa's VAMP program adds TC40 fraud reports and TC15 disputes, then divides by settled card-not-present Visa sales.

In AP, Canada, the EU and the US, the standard merchant trigger is a 1.5% ratio (from April 1, 2026). The merchant also needs at least 1,500 fraud and dispute records a month.

Apply exclusions and country and regional rules. Mastercard's ECM program uses 100–299 current-month chargebacks and a 1.5%–2.99% ratio to previous-month captured payments. Fees and next steps depend on the program and your acquirer's notice. A ratio alone does not decide penalties.

Thresholds differ by region. Costs are heading up too. Mastercard forecasts that global chargeback costs will rise from $36.9 billion in 2026 to $46.1 billion in 2029 (Mastercard and Datos Insights, June 2026).

Check the Visa VAMP fact sheet and your acquirer's notice. For how processors apply this, and for Mastercard's own reporting periods, see Stripe's monitoring-program guidance.

AI-agent disputes add to your chargeback count, whether or not they are fair. If agent orders become a real share of your sales, those disputes can push you toward the limits fast. And the monitoring programs treat a dispute caused by an AI agent the same as any other.

How Much Can Agent Orders Raise Your Chargeback Ratio?

Hypothetical order-count arithmetic

See how the agent segment changes the blended rate

Blended chargeback ratio = agent share × agent dispute rate + remaining share × baseline rate. These invented scenarios use a 0.60% non-agent baseline; they are not a forecast.

No agent orders

100% × 0.60%

0.60%

0.90 percentage points below a 1.5% comparison line.

5% agent share · 3% agent dispute rate

5% × 3% + 95% × 0.60%

0.72%

0.78 percentage points below a 1.5% comparison line.

10% agent share · 3% agent dispute rate

10% × 3% + 90% × 0.60%

0.84%

0.66 percentage points below a 1.5% comparison line.

10% agent share · 6% agent dispute rate

10% × 6% + 90% × 0.60%

1.14%

0.36 percentage points below a 1.5% comparison line.

20% agent share · 6% agent dispute rate

20% × 6% + 80% × 0.60%

1.68%

0.18 percentage points above a 1.5% comparison line.

This is disputes divided by orders, with made-up inputs. A 1.5% line alone does not show VAMP or ECM status. VAMP adds in-scope fraud reports and disputes, then applies exclusions, count rules and regional rules. Mastercard uses its own timing and counts. Track the agent segment, and work out each official program on its own.

The practical rule: tag agent orders with the identifier you log and track their dispute rate as a separate segment. You then see the problem before the blended ratio moves.

Les protocoles du secteur que les commerçants doivent connaître

The payments industry is building ID and consent layers for agents. None of the protocols below decides who pays a chargeback among merchants, issuers and agent platforms. American Express's program covers cardholders only. Status as of November 2026:

Visa's Trusted Agent Protocol (TAP) launched on October 14, 2025 with Cloudflare. It uses signed requests, so merchants can verify trusted agents and tell them apart from bad bots. It carries the agent's intent, data that helps you recognize the buyer, and payment details.

Visa Intelligent Commerce Connect was announced on April 8, 2026 and is in pilot. It supports TAP, the Machine Payments Protocol, the Agentic Commerce Protocol and the Universal Commerce Protocol.

That signed record is the kind of evidence that can support a dispute response, although Visa's announcements do not describe dispute treatment.

Mastercard's Agent Pay issues agentic tokens. These are dynamic card credentials tied to a registered agent. It also records what the buyer meant to buy before the sale. Its Agent Pay Acceptance Framework lets merchants spot registered agents through Web Bot Auth checks at the CDN layer, with no new code.

Google's Agent Payments Protocol (AP2) was announced on September 16, 2025 and is now developed with the FIDO Alliance. A checkout mandate captures purchase details and user consent. A payment mandate approves a specific amount.

The protocol calls the result a non-repudiable audit trail. That is the record you want when a cardholder says they never approved a purchase. The Universal Commerce Protocol, built by Google, Shopify and others, covers every step from search to checkout. It uses AP2 for payments.

OpenAI's Agentic Commerce Protocol (ACP), built with Stripe, sets how agents pass payment details and order data to merchants through a Shared Payment Token. OpenAI dropped its own Instant Checkout in March 2026 and moved to merchant-owned checkout. ACP still powers merchant setups.

ChatGPT Wallet launched on November 13, 2026. OpenAI built it with American Express, Mastercard and Visa. It passes tokenized card credentials, so the agent and the merchant never see the real card number. Users approve each purchase with a biometric check such as Face ID. See ChatGPT Wallet chargebacks for the merchant side.

OpenAI calls the start limited: only US ChatGPT Pro subscribers aged 18+ with eligible US-issued cards.

OpenAI's ChatGPT Wallet terms say OpenAI is not the merchant. It may share Wallet records, but it does not decide or fund refunds or chargebacks unless the law requires it. The buyer's approval covers only the merchant, items, amount and delivery details shown before checkout. If a key detail changes, Wallet asks again.

Keep each order exactly as the buyer approved it. That match is your best proof the buyer agreed.

American Express announced Agent Purchase Protection on April 14, 2026. It protects eligible cardholders from charges caused by a registered AI agent's mistake. The cardholder must have approved the agent. The agent must have sent signed purchase intent. Terms apply.

The announcement does not say how merchants are treated. American Express runs a closed-loop network, so it can set cardholder protection without waiting for thousands of banks to agree. That is harder under the four-party Visa and Mastercard model.

EMVCo published a draft EMV Agentic Payments framework on September 1, 2026, with public feedback open through September 30, 2026. It adds Intent Services, a shared layer that stores what the consumer approved. It does not address disputes or liability.

Voici un aperçu comparatif de ces protocoles :

Pro­to­colOwnerEvidence You Can Keep / Dispute Treatment Published
Trusted Agent Pro­to­col (TAP)VisaEvidence You Can Keep: Signed agent identity, consumer recognition data and purchase intent
Dispute Treatment Published: None in the announcement
Rémunération des agentsMastercardEvidence You Can Keep: Agentic token tied to a registered agent, plus verifiable purchase intent
Dispute Treatment Published: None on Mastercard's Agent Pay pages
Agent Payments Pro­to­col (AP2)Google, with the FIDO AllianceEvidence You Can Keep: Checkout and payment mandates with a cryptographic audit trail
Dispute Treatment Published: None; the audit trail is meant to support dispute resolution
Universal Commerce Pro­to­col (UCP)Google, Shopify and other companiesEvidence You Can Keep: Checkout data built on AP2 mandates
Dispute Treatment Published: States that the business stays merchant of record
Agentic Commerce Pro­to­col (ACP)OpenAI and StripeEvidence You Can Keep: Shared Payment Token and order data
Dispute Treatment Published: Production guidance says the merchant manages refunds and chargebacks
ChatGPT WalletOpenAI, with American Express, Mastercard and VisaEvidence You Can Keep: Network token in place of the real card number; the buyer approves in ChatGPT with a biometric check
Dispute Treatment Published: None for merchants. OpenAI's terms say it does not decide or fund chargebacks; users keep their card's protections for eligible disputes
Protection des achats auprès d'un agentAmerican ExpressEvidence You Can Keep: Authenticated purchase intent from a registered agent
Dispute Treatment Published: Covers eligible cardholders; merchant treatment not stated
EMV Agentic Payments framework (draft)EMVCoEvidence You Can Keep: Registered, consumer-authorized intent through Intent Services
Dispute Treatment Published: None; the draft does not address disputes or liability

Treat these as an evidence source first. Until the networks publish dispute rules for agent purchases, a signed agent record strengthens your response. Only the tokenized, authenticated flows above move fraud liability toward the issuer.

Does AEP Change Who Pays for an Agent Purchase?

No. AEP records what happened. It does not decide who pays. A valid record can show what the buyer approved and what the agent did. Your provider still applies the rules and contracts for that claim.

Chargeflow joined the A-Comm Evidence Protocol consortium as a founding member. As checked on November 17, 2026, v1.0.3-rc.2 is still a draft. A valid chain is no promise of a win.

Check who can get each record and how it reaches your provider before the deadline. Use the agentic commerce evidence playbook to test that process. Collecting proof and deciding who bears a loss are separate tasks.

Règlement de 2026 sur le commerce agentique : état des lieux

As of November 2026, no government has passed a law that says who is liable when an AI agent buys on its own. The EU is the first to propose rules written for AI shopping agents.

The European Commission proposed the Digital Fairness Act on November 11, 2026. Article 26 covers AI agents that sign contracts for consumers. At first use, the consumer picks whether to confirm every contract, only those that meet set limits such as a maximum price, or none.

If the agent's provider is paid for proposing a deal or has a conflict of interest with the merchant, it must disclose it and get the consumer's confirmation. Fines could reach 6% of worldwide annual turnover.

It is still a proposal, not law, and the European Parliament and the Council must agree a final text. It does not set chargeback liability, but the consent settings it requires could become useful dispute evidence.

The EU AI Act, the broadest AI law on the books, does not cover agent purchases or payment liability. It focuses on AI risk levels and transparency. The EU's AI Omnibus, formally adopted in mid-2026, moved the high-risk deadlines. Stand-alone systems now face December 2, 2027, and systems built into regulated products face August 2, 2028.

PSD3 and the new Payment Services Regulation reached political agreement in November 2025. Neither applies yet. Legal summaries of the agreed texts, such as this April 2026 review, focus on bank-transfer scams and outsourcing duties. Merchants should not expect them to settle agent card disputes.

In the United States, Regulation Z and Regulation E, described above, are still the baseline. No federal rule covers agent purchases. Worldpay notes that Regulation E treats consent as yes or no. It has no answer for an agent that bought something pricier than the consumer meant.

The bottom line: the payments industry is moving faster than regulators. Card network protocols and platform contracts will likely set the rules before laws catch up. Merchants can't wait for regulation to protect them.

How Should Merchants Update Dispute Processes for Agent-Led Transactions?

Treat agent orders as their own segment from checkout to dispute. Tag them, keep the consent record with the order and pick evidence by the reason code the cardholder uses. The rest of your chargeback process can stay the same.

ScèneWhat to changeWhat you keep
Order intakeFlag every agent order with the agent ID, protocol and token typeAn agent flag on each order record
Before shippingScreen agent orders like any card-not-present order and hold orders outside the mandateThe screening decision and the reason
Customer noticeSend an order confirmation with a cancel link and use a clear billing descriptorThe confirmation time and any cancel request
Pre-dispute alertRefund alerts on orders outside the mandate; review alerts on orders inside itThe refund or fight decision
Dispute responseFraud codes: agent identity, token and consent. Consumer codes: mandate vs order, listing and deliveryA response built for the claim type
ReportingTrack the agent segment's dispute rate and win rate apart from the restA monthly agent-segment ratio

Win rates on this segment show whether your evidence works. If agent disputes keep losing on the same reason code, fix the record you keep for that code first.

Comment les commerçants peuvent-ils se protéger dès maintenant ?

Vous n’avez pas besoin d’attendre que les réglementations ou les protocoles du secteur soient finalisés. Vous pouvez dès aujourd’hui prendre des mesures concrètes pour réduire votre exposition à l’ rétrofacturation e des agents IA.

Élaborer une stratégie en matière de preuves pour les transactions menées par des agents

The evidence gap is your biggest weak spot. Start closing it now.

Evidence readiness

Keep the authority chain alongside the order

Agent identifiers

Capture and store the identity of any AI agent that initiates a transaction on your platform. Log the agent type, version, and the platform it operates on.

Mandate and authorization records

Record the authorization chain, meaning what permissions the consumer granted to the agent and when. These logs become your primary evidence in disputes.

Timestamps and interaction records

Document every touchpoint between the agent and your checkout process. The more granular your records, the stronger your dispute defense.

Agreed platform exports

Set up agreements with AI platforms to get signals from the agent's decision process. This data replaces the browsing evidence you lose on agent orders.

Link each record to the order. A signed agent identity shows the source; the mandate and order comparison show the scope of authority. Retain permitted evidence without storing secret API keys or raw payment credentials.

Build this evidence infrastructure before disputes arrive. Retrofitting your systems after chargebacks start rolling in costs more time and money. The chargeback evidence source map shows where each piece of proof lives across your payment stack.

Visa's Compelling Evidence 3.0 needs matching data such as IP address, device and account across past undisputed orders. Agent traffic often comes from the agent platform's servers. So check that your past agent orders give you matching data before you rely on this route.

Chargeflow Automation collects 1,000+ data points per dispute to build the evidence package. Agent IDs and mandate records sit next to the usual order, delivery and customer history in one response.

PRÉVENTION rétrofacturation Avant qu'ils ne se produisent

Preventing agentic commerce fraud starts with stopping disputes before they become chargebacks. This matters most for AI-agent orders, where your evidence is weaker from the start.

Chargeback alert programs such as Verifi and Ethoca tell you when a cardholder disputes a charge. You can then fix the issue, often with a refund, before it becomes a formal chargeback on your record.

Post-purchase fraud detection adds another layer on top of standard ecommerce fraud prevention. These tools check order patterns after the purchase and flag risky orders.

That includes orders from a hijacked or rogue agent the consumer never approved. These often start with account takeover. For the steps to verify real agent orders, see preventing fraud and chargebacks in agentic commerce.

Chargeflow Alerts works through Verifi RDR, Ethoca and the Chargeflow Network. It refunds each alert for you and claims up to 90% deflection. Chargeflow Prevent scans each order after checkout and before shipping. It scores device, IP, email and payment behavior against a network of 20,000+ merchants, so you can stop risky orders before they ship.

What SLAs and Chargeback Guarantees Should You Require From a Fraud Vendor for AI Agent Orders?

Ask whether the guarantee names agent orders, which dispute types it pays on and what data you get back. Chargeback guarantees often cover only fraud chargebacks on approved orders. Many agent disputes come in as consumer disputes instead.

Contract termWhat to askWhy it matters for agent orders
Order scopeAre agent and tokenized agent orders covered, or excluded as a new channel?An exclusion leaves your newest risk uncovered
Dispute typesFraud codes only, or also "not received" and "not as described"?Agent disputes often use consumer codes
Decision speedHow fast does the vendor approve or decline an order?Agent checkouts expect a fast answer
Data accessCan you export the decision, score and agent identifiers for rep­re­sent­ment?You need them to fight what the guarantee skips
Ratio supportDoes the vendor report your agent-segment dispute rate?Reimbursed chargebacks still count toward VAMP and Mastercard ratios
Claim termsWhat is the claim deadline, and who answers the chargeback?A missed claim window can void coverage

A guarantee pays back money. It does not take the dispute off your monitoring count, so pair it with alerts and a response process.

Automatisez votre réponse aux litiges

Manual chargeback management cannot keep up with AI-agent dispute volume. Every dispute needs evidence, formatting and a response before a tight deadline. Miss a deadline, and you lose by default.

Automated chargeback platforms handle the whole dispute process without waiting on staff. They collect evidence, build the response, submit it on time and learn from each result. For merchants facing a new kind of dispute on top of their usual volume, this is no longer optional.

Chargeflow Automation handles chargebacks end to end, from evidence to submission. It carries a 4X ROI guarantee and a 25% fee on recovered chargebacks only. A dispute on an agent order is still a card dispute with the same reason codes and deadlines. What changes is the evidence you attach.

Can an AI Agent Handle Your Chargebacks End to End?

Yes, for the repeatable work. Agentic AI for card dispute handling can match each dispute to its order, pull evidence, write the response and submit it before the deadline. Keep people on policy: which disputes to accept, refund limits and unusual cases.

Check three things before you trust a dispute agent. It should pull data from every processor you use. It should map evidence to each reason code. And it should show you what it submitted, so you can audit the result.

Définir des lignes directrices claires pour les clients

La prévention passe d'abord par la communication. Veillez à ce que vos clients sachent quand un agent IA lance une transaction en leur nom.

Reduce surprise purchases

Give the customer a clear view of the agent’s order

Immediate confirmation

Send a clear confirmation as soon as an agent completes a purchase. Include what was purchased, the total cost, and a simple cancellation option.

Recognizable billing descriptor

Use billing descriptors that clearly identify your business. Consumers who do not recognize a charge on their statement are far more likely to dispute it.

Review before fulfillment

Give customers time to review and cancel agent purchases before you ship. This reduces disputes driven by surprise or confusion.

Agent activity summary

If your platform supports AI agents, provide customers with regular summaries of what their agents purchased. Transparency reduces the "I did not approve this" disputes that drive chargeback volume.

A notice or an unused cancel option does not prove consent by itself. Keep it with the mandate, the order details and any cancel history.

Chargeflow covers all three stages. Alerts stop disputes before they post, and Prevent screens orders before shipping. Automation builds and submits the response when a chargeback still lands.

Démarrer gratuitement

FAQ

Qui est responsable lorsqu'un agent IA effectue un achat non autorisé ?

No card rule assigns liability to the agent platform. An agent purchase is handled like a card-not-present sale. If the cardholder claims fraud and the sale was authenticated through a network agent program, fraud liability usually sits with the issuer. Otherwise the merchant bears the loss unless it wins the dispute.

Visa's October 2026 rules make the cardholder responsible for what a registered agent provider does. They add no new dispute process. Consumer rules such as Regulation Z do not cover agents yet.

Les commerçants peuvent-ils contester les « rétrofacturation s » issues de transactions effectuées par un agent IA ?

Yes. You respond with the same process and deadlines as any chargeback. Build the case from agent identifiers, the mandate record and order confirmations, since device and browsing data may be missing. See how to fight a chargeback for the response workflow.

Quel est l'impact du commerce par agent sur les taux de rejet de paiement ?

AI agents can raise dispute volume through unauthorized purchases, duplicate orders and customer confusion. Use each network's own counts, exclusions and time periods, as shown above. Keep this rough math apart from the official math. A won dispute or an internal alarm does not make you safe with your processor.

How Much Can Agent Orders Raise My Chargeback Ratio?

Blended ratio equals agent share times agent dispute rate, plus the remaining share times your baseline. A merchant at a 0.60% baseline where 10% of orders come from agents that dispute at 3% lands at 0.84%. No public source reports a dispute rate for agent orders, so track your own agent segment.

Qu'est-ce que le protocole « Trusted Agent » de Visa ?

TAP is Visa's framework for verifying AI agents acting on behalf of consumers. Visa launched it in October 2025 with Cloudflare. It uses cryptographic signatures to help merchants tell trusted agents from malicious bots, and it carries agent intent and consumer data. Visa has not published how disputes are treated.

Who Is the Merchant of Record When an AI Agent Completes a Purchase?

The merchant that sells the goods and takes the payment. OpenAI's production guidance and the Universal Commerce Protocol both say the agent platform and payment providers are not the merchant of record. The merchant handles refunds and chargebacks.

Does American Express Agent Purchase Protection Cover Merchants?

The announcement covers eligible cardholders who approved a registered agent that sent signed purchase intent. Terms apply. It does not say how merchants are treated, so confirm the dispute process with your acquirer.

Can an AI Agent Be Held Legally Responsible for a Purchase?

An AI agent is software, not a legal person, so no card rule makes the agent itself liable. Responsibility falls on the parties around it: the consumer who gave it authority, the company that runs it and the merchant that accepted the order. Who pays a chargeback is decided by card network rules and contracts.

How Can Merchants Prevent Chargebacks From AI Agents?

Merchants should use chargeback management best practices. Catch disputes early and screen agent orders after purchase. Automate evidence, and tell customers clearly about AI-made purchases. Before you give a support agent refund or cancel rights, compare which AI agents manage post-purchase actions.

What Is the Biggest Risk of Accepting AI Agent Payments for Merchants?

The biggest risk is losing disputes you cannot prove. Device and browsing data often point to the agent platform, not the cardholder. Unless you keep the agent identity, token and consent record, an "I did not approve this" claim is hard to beat.

What Chargeback Rate Increase Should Merchants Tolerate From Agent-Led Transactions?

No network gives agent orders extra room. Your limit is the gap between your current ratio and each program's trigger, such as Visa's 1.5% VAMP ratio. Set an internal alarm well below that gap and track the agent segment on its own.

Who Pays When an AI Agent Overspends or Buys the Wrong Item?

The merchant pays by default if the cardholder files a consumer dispute and the merchant can't show the order matched the listing and the mandate. Visa's rules make the cardholder answer for a registered agent provider's actions. So keep the consent record and the order confirmation. The cardholder's claim against the agent provider is separate from your dispute.

How Should Merchants Allocate Liability Between PSPs, AI Agents and Customers?

Card rules do not move chargeback costs to the agent platform, so use contracts. Your acquiring or PSP agreement sets what you pay per dispute. A direct deal with an agent platform can add data sharing and an indemnity for agent errors. Your customer terms should cover agent confirmations, cancel windows and refunds.

PARTAGER CET ARTICLE
Logo circulaire blanc comportant, en son centre, des formes entrelacées, entouré de lignes elliptiques qui se chevauchent et ressemblent à des orbites, ainsi que de losanges bleus dispersés.

rétrofacturation?
Ce n'est plus votre problème.

Récupérez 4 fois plus d'rétrofacturation s et PRÉVENTION jusqu'à 90 % des messages entrants, grâce à l'IA et à un réseau mondial de 20 000 commerçants.

Plus de 600 avis
Aucune carte bancaire n'est nécessaire.
s'abonner

Les dernières actualités sur l'rétrofacturation, la fraude et le commerce électronique, directement dans votre boîte mail. Chaque semaine.

Inscrivez-vous dès maintenant pour ne rien manquer des dernières tendances !
En indiquant votre adresse e-mail, vous acceptez nos Conditions d'utilisation et notre Politique de confidentialité
Schéma composé de lignes pointillées et courbes formant des arcs segmentés, mis en évidence par trois repères en forme de losange bleu situés à gauche.Motif abstrait en forme de grille circulaire, avec des repères en forme de losanges bleus sur un fond moitié noir, moitié blanc.