%201.svg)
Agentic Commerce Chargebacks: The Evidence Playbook Merchants Need
Chargebacks?
No longer your problem.
Recover 4x more chargebacks and prevent up to 90% of incoming ones, powered by AI and a global network of 20,000 merchants.
600+ reviews.avif)
TL;DR:
- Agentic commerce chargebacks run through existing reason codes, so you win or lose on proving what the customer authorized the agent to do.
- Log the agent’s verified identity, the customer’s mandate, the cart the agent approved and the notification sent for every agent-initiated order.
- Visa’s April 2026 rules added agentic payment requirements, but as far as the public text shows they define no separate agentic dispute condition or liability shift.
- Compelling Evidence 3.0 covers only Visa 10.4 and needs an IP or device match that agent traffic rarely provides.
- Track agent-led orders separately, then calculate Visa VAMP and Mastercard ECM using each program’s event definitions, reporting periods, exclusions and count conditions.
Agentic commerce chargebacks are disputes that occur when an AI shopping agent completes a purchase on a customer’s behalf and the customer later challenges that transaction with their bank or card issuer. They are emerging as a distinct dispute category because the evidence merchants have always relied on to prove a sale, built for human buyers, cannot show that an AI agent acted within the boundaries a customer actually set.
This page is the evidence side: what to log for every agent-initiated order, which records answer each dispute claim, and how to write the response. For who ultimately pays, see AI agent chargeback liability. For what counts as an authorized agent purchase, see agentic commerce chargeback liability.
What Is Agentic Commerce?
Agentic commerce is when AI agents handle the purchase process on behalf of a consumer. The agent browses products, compares options, selects a match, and completes the transaction, sometimes without the customer clicking “buy.” A product recommendation tells you what to buy. An agentic commerce system buys it for you.
Agent-led checkout is live but uneven. Google announced its Universal Commerce Protocol in January 2026 with Shopify, Etsy, Wayfair, Target and Walmart, Mastercard has offered Agent Pay since April 2025, and Visa introduced its Trusted Agent Protocol in October 2025. One major AI assistant provider reportedly pulled back in-chat checkout in March 2026 and now sends shoppers to merchants’ own checkouts, which means agent-assisted orders can reach your normal storefront, where your own logs are the only record you control.
Demand is moving fast. Adobe Analytics reported that traffic from AI sources to US retail sites grew 393% year over year in the first quarter of 2026. Adobe measures AI-referred visits, not purchases completed by agents, and public dispute data for agent-initiated orders is not available yet, so your own order and dispute logs are the earliest signal you will get.
For merchants, the buyer at your checkout may not be human, does not read a product description the way a person does, and may not fully understand your customer’s intent.
How Agentic Commerce Creates New Types Of Chargebacks
These disputes are not traditional chargebacks. The core problem is the gap between “authorized” and “wanted.” A customer who gives an AI agent permission to shop authorizes the agent to act, but not necessarily to buy that product, at that price, from that merchant. Traditional dispute handling assumes a human reviewed the purchase and clicked “buy,” so these patterns do not fit neatly into existing reason codes:
| Dispute Pattern | What Happens | Why It Leads to a Chargeback |
|---|---|---|
| Forgotten Agent Purchases | A customer sets up an AI agent, forgets about it, and discovers unexpected charges on their statement | The purchase was authorized, but the customer does not remember it |
| Misinterpreted Intent | The agent understood “buy affordable running shoes” differently than the customer meant | The product technically matches the criteria, but it is not what the customer wanted |
| Fraud Hiding Inside Automation | Bad actors use AI agents as a layer of separation between themselves and fraudulent purchases | The agent provides cover, making it harder to trace intent back to the actual person |
| Trust-Break Chargebacks | The customer trusted the AI agent to make good decisions, and the agent made a bad one | The customer blames the merchant instead of the platform that built the agent |
If you have dealt with friendly fraud before, this will look familiar: a legitimate customer disputing a legitimate transaction, with an AI intermediary that makes the “legitimate” part harder to prove. Auto-purchases and renewals are covered in AI shopping chargebacks, and the attacker side, where criminals use or impersonate agents, in preventing AI agent fraud.
The Agentic Commerce Liability Gap
When an AI agent buys something your customer disputes, the merchant absorbs the loss by default, because the dispute process was not built to separate a human buyer from an agent. Worldpay’s July 2026 analysis found no formal liability shift for agent purchases and described disputes short of outright fraud, such as a misread instruction, as unsettled between merchant, issuer and agent platform.
Industry opinion is split too. In a Darwinium survey of 500 fraud, risk, and security leaders, 39% pointed to the AI provider as liable when an agent-driven transaction goes wrong, 20% to the customer, 14% to the merchant and 11% to the bank, with a smaller group favoring shared liability. Until the rules settle, you win or lose disputes on the evidence you can produce.
New Protocols For Agent-Led Transactions
Card networks and platform companies are building protocols that create a verifiable chain between the consumer, the AI agent and the merchant. Each one leaves records you can use as evidence.
| Protocol | Owner and Date | What It Does |
|---|---|---|
| Trusted Agent Protocol (TAP) | Visa, October 2025, built with Cloudflare | The agent sends a signed request that merchants verify against a registry of agent keys, with signals for agent intent, consumer recognition and optional payment information |
| Agent Pay | Mastercard, April 2025 | Requires agents to be registered and verified, and issues agentic tokens tied to the consumer, who authenticates on their own device |
| Agent Payments Protocol (AP2) | Open protocol that records the customer’s constraints and approval in signed checkout and payment mandates, building an audit trail for disputes | |
| Universal Commerce Protocol (UCP) | Google, January 2026, with Shopify, Etsy, Wayfair, Target and Walmart | A common language between agents, merchants and payment providers across the shopping journey; works alongside AP2, and retailers keep seller-of-record status |
| Agentic Commerce Protocol (ACP) | OpenAI and Stripe | Open standard for checkout between buyers, agents and businesses, passing payment credentials to agents through a shared token without exposing the underlying card |
| Verifiable Intent | Mastercard with Google, March 2026 | Open-source specification for a tamper-resistant record of the consumer’s identity, their instructions and the agent’s resulting interaction with the merchant; Mastercard said at launch that it would integrate it into Agent Pay |
Visa has also written agentic payments into its rulebook. The Visa Core Rules edition of 18 April 2026 includes section 4.1.24, which sets requirements for agentic payment enablers and providers, including obtaining cardholder instructions before processing. As far as the public text shows, it does not create a separate agentic dispute condition or liability shift, so these protocols give you evidence rather than a new defense by themselves. None of them carries the force of law either, and agentic commerce regulation tracks where lawmakers stand.
Why Chargeback Evidence Standards Are Changing
Traditional chargeback evidence was built for a world where humans buy things, and agentic commerce makes it incomplete:
| Evidence Type | Traditional (Human Buyer) | Agentic Commerce (AI Agent Buyer) |
|---|---|---|
| Identity | AVS matches, IP addresses, device fingerprints | IP address belongs to the agent platform’s server, not the customer |
| Delivery | Tracking numbers, signed receipts | Same, but the customer may not have chosen the shipping address |
| Authorization | Transaction logs, 3D Secure records | Delegation of authority to the agent, not direct purchase confirmation |
| Intent | Customer clicked “buy” after reviewing the product | Customer may never have seen the product page at all |
Traditional evidence proves a human was on the other side of the transaction. Agentic evidence has to prove the agent acted within the boundaries the human set, which takes new records alongside the compelling evidence you already collect: proof of delegated authority, the limits the customer set, scope compliance and notification timestamps.
What To Log For Every Agent-Initiated Order
Capture these records at order time, keyed to the order ID. Pulling them from an agent platform after a dispute is open is slower than reading them from your own system.
| Record | Where It Comes From | What It Proves |
|---|---|---|
| Agent identity and signature check | TAP request signature fields (key ID, timestamp, session ID, domain binding) checked against the registry, or the Agent Pay agentic token | The order came from a registered agent, not an anonymous bot |
| Customer mandate | AP2 checkout and payment mandates, a Verifiable Intent record, or the agent platform’s consent export | What the customer authorized: item or category, price cap, merchant scope and expiry |
| Cart as approved and as placed | Your order system compared with the closed mandate or cart record | The agent stayed inside the customer’s limits on price, quantity and item |
| Customer recognition | Account ID, prior undisputed orders, and any consumer or loyalty identifier the agent passes | An existing relationship with the cardholder, and the inputs Compelling Evidence 3.0 asks for |
| Notification and cancel window | Email, app or SMS logs with timestamps and any cancel-link activity | The customer was told about the order and had a chance to stop it |
| Payment authentication | 3D Secure result, token type, AVS and CVV responses | How the payment was authenticated, and whether fraud liability shifted |
| Fulfillment | Carrier tracking, delivered address compared with the address on file, signed receipt | The goods reached the customer or the address they use |
Keep these records at least as long as the longest dispute window that could apply. Some Visa and Mastercard conditions for non-receipt and interrupted service reach 540 days, and your response deadline runs from the case notice, not the order date, as the guide to chargeback time limits explains. Store identifiers and signed records, not raw card data or API keys. Ask each agent platform you accept orders from, in writing, which mandate and consent records it can export per order and how quickly.
Evidence By Dispute Scenario
Agent orders are card-not-present transactions, and disputes over them arrive under existing reason codes: mainly Visa 10.4 (card-absent fraud), 13.1 (merchandise not received) and 13.3 (not as described), and Mastercard 4837 (no cardholder authorization) and 4853 (cardholder dispute). The claim the cardholder files decides which records matter.
| Claim | Likely Codes | Evidence To Submit | Where It Is Weak |
|---|---|---|---|
| Customer says they never authorized the purchase | Visa 10.4, Mastercard 4837 | Mandate with timestamp, agent signature check, notification log, prior undisputed orders, billing descriptor match | No mandate record: the agent platform holds it, so request it before the deadline |
| Agent bought the wrong thing | Visa 13.3 or 10.4, Mastercard 4853 or 4837 | The customer’s instruction text, the options the agent was shown, the listing as displayed, the cart the agent approved | Ambiguous instructions favor the customer; consider a refund before the dispute escalates |
| Agent exceeded the customer’s limits | Visa 10.4, Mastercard 4837 | Cart compared with the mandate on price, category and merchant scope | If the order broke the limits, the response will not hold; accept it and pursue the agent platform under your contract |
| Item not received | Visa 13.1, Mastercard 4853 | Tracking, delivery to the address on the customer’s profile, proof that address came from the customer and not the agent | An address the agent chose without customer confirmation |
| Order came from a spoofed or hijacked agent | Visa 10.4, Mastercard 4837 | Little: a failed or missing signature check means this is a prevention case, not a response case | You shipped without verifying the agent |
For the code-by-code response rules, see our guides to Visa chargeback reason codes and chargeback reason codes and how to fight them.
Compelling Evidence 3.0 And Agent-Initiated Orders
Compelling Evidence 3.0 applies only to Visa dispute condition 10.4. Adyen’s documentation summarizes the test: at least two earlier transactions on the same payment method that settled 120 to 365 days before the dispute and were never disputed, with at least two data elements (user ID, IP address, delivery address, device ID or fingerprint) matching the disputed order, one of them the IP address or device ID. It applies where 3D Secure liability shift does not, and following it is optional.
Agent orders strain the IP and device match. An agent runs on its platform’s servers, so the IP address you log usually identifies the platform rather than the customer’s device, and a match shows only that both orders came through the same service. Account ID and delivery address can still match. Log any stable consumer identifier the agent passes, such as the optional consumer recognition fields in TAP, and ask your acquirer how it treats those fields in a Compelling Evidence 3.0 submission.
Representment Template For Agent-Initiated Disputes
Lead with the claim, then answer it in the order the issuer reads it: who bought, what the customer authorized, what the agent bought, what the customer was told, and delivery. Fill the brackets from your logs, attach each exhibit, and keep the tone factual. The response deadline comes from your case notice; the general process is in our guide on how to fight a chargeback.
Editable representment template
Build the response around five linked exhibits
Case [case ID], [network] reason code [code], order [order ID], [amount] on [order date].
1Who bought · Exhibit A
The order was placed by [agent platform], a registered agent whose request signature (key ID [ID]) was verified at [time] against [registry], acting for account [customer ID], open since [date] with [number] prior undisputed orders. Exhibit A: signature verification record.
2What the customer authorized · Exhibit B
On [date and time] the customer approved [item or category], up to [price cap], for [merchant scope], valid until [expiry]. Exhibit B: mandate record.
3What the agent bought · Exhibit C
[item, quantity, price], within the approved limits. Exhibit C: cart as approved compared with the order as placed.
4What the customer was told · Exhibit D
Confirmation sent to [channel] at [time], with a cancel option open until [time]. The order was not cancelled. Exhibit D: notification log.
5Delivery · Exhibit E
Shipped via [carrier], tracking [number], delivered [date] to [address], which matches [source of address]. Exhibit E: tracking and delivery proof.
Request: the cardholder authorized this purchase through a delegated agent, and the order stayed inside the scope they set. We ask that the dispute be reversed in the merchant’s favor.
Fill placeholders only from records you can substantiate. Use the authorization request above only when the exhibits support it; adapt the response to the actual claim and case deadline. A signature or successful delivery alone does not establish every fact.
How Agentic Commerce Affects Your Chargeback Ratio
You do not need more actual fraud for agentic commerce to increase your dispute volume. When a person buys something themselves, they remember the decision. When an AI agent handles the whole process, the charge can look like any other unfamiliar transaction, and disputing it is the easy response.
The baseline is already rising. Mastercard’s chargeback cost research (updated June 2026, by Datos Insights) projects about 286 million chargebacks in 2026, growing 37% from 2025 to 359 million a year by 2029, and puts the global chargeback cost at $36.9 billion in 2026, rising to $46.1 billion in 2029.
More disputes can increase your chargeback ratio, and the same chargeback thresholds apply to agent-led orders as to human ones. Visa VAMP uses the combined count of TC40 fraud reports and TC15 disputes divided by the count of settled TC05 card-not-present VisaNet transactions. Its standard Excessive Merchant criteria in AP, Canada, the EU and the US pair at least a 1.5% ratio (from April 1, 2026) with at least 1,500 monthly fraud-and-dispute records, when the acquirer is not Above Standard or Excessive. Apply exclusions and country and regional rules. The Mastercard ECM program uses current-month chargebacks against previous-month captured payments and its own count and ratio criteria. Confirm any identification, fees and remediation from the applicable notice; neither a ratio alone nor this hypothetical internal alarm establishes program status. Merchants on Shopify Payments also face Shopify’s Network Dispute Resolution Program, where disputes settle automatically and no evidence can be submitted.
Check the Visa VAMP fact sheet and your acquirer’s applicable notice. For processor implementation and the separate Mastercard reporting periods, see Stripe’s monitoring-program guidance.
Hypothetical internal alarm
Work backwards from your own dispute-count ceiling
Total monthly orders
10,000
9,000 non-agent orders + 1,000 agent orders.
Internal alarm line
1.0%
10,000 × 1.0% = 100 total disputes at the line.
Disputes on other orders
45
Assumed disputes on the 9,000 non-agent orders.
Agent disputes at the line
100 total disputes − 45 other disputes
55
55 ÷ 1,000 agent orders = 5.5%.
Agent rate at the line = (ceiling × all orders − other-order disputes) ÷ agent orders. At these inputs, 55 agent disputes brings the blend to exactly 1.0%. Staying below the alarm requires fewer than 55 agent disputes.
Invented inputs, aligned reporting periods, one dispute per order, and a chosen internal alarm. This is not an official VAMP or ECM formula or an agent-dispute forecast. Network programs apply their own event populations, dates, exclusions, and count conditions. If other-order disputes already reach the ceiling, there is no remaining allowance in this model.
How To Prepare Your Chargeback Strategy For Agentic Commerce
The protocols and rules are still taking shape, but these steps work with what exists today.
Set Guardrails For Agent-Initiated Purchases
Give your customers clear controls over what AI agents can and cannot do on your store as part of your broader chargeback mitigation plan.
- Spending limits: Let customers set maximum purchase amounts for agent-initiated transactions
- Category restrictions: Allow customers to define which product categories the agent can and cannot buy from
- Approval requirements: Require human confirmation for high-value or unusual orders before fulfillment
- Real-time notifications: Send immediate alerts when an agent-initiated purchase is completed so the customer can review and cancel if needed
Each control produces a record for your evidence log. A recognizable billing descriptor also keeps an agent order from reading as an unknown charge on the statement.
Capture Intent-Based Evidence From Day One
Do not wait for new protocols to become mandatory. Add the fields from the logging table to your order record, agree record exports with each agent platform, and test the process before a real dispute does:
Before the first live case
Rehearse the evidence workflow on a recent order
1Add the order-level fields
Add fields for agent platform, signature check result, mandate reference and notification time to every agent-initiated order
2Assemble and time a full packet
Pick one recent agent order and assemble a full response packet from your logs, timing how long it takes and noting which exhibit was missing
3Report the agent segment
Report agent-initiated orders as their own segment, so disputes on them show up before they move your overall ratio
Agree order-level record exports with each agent platform and identify the owner of every missing exhibit. An operational rehearsal does not predict whether an issuer will accept the evidence.
Prevent Chargebacks Before They Happen
Fraud does not disappear because a transaction runs through an AI agent. A failed or missing signature check is a reason to hold an order, and standard controls such as AVS, CVV and velocity rules still apply; see ecommerce fraud prevention for the full set. Chargeflow Prevent scans orders after checkout and before fulfillment, so you can hold a risky order before it ships. Chargeflow Alerts receives notifications from card-network alert programs such as Verifi RDR and Ethoca before a chargeback is filed, so you can refund first; chargeback alerts explains how that works. For the operations side of legitimate agent traffic, see preventing fraud and chargebacks in agentic commerce.
Automate Your Dispute Response
If agent-initiated orders add disputes, manual evidence gathering is the first thing to break. Chargeflow Automation collects evidence from your connected systems, builds the response and submits it, and you pay 25% only on chargebacks it recovers, backed by a 4X ROI guarantee. Keep your agent records retrievable by order ID so they can go into any response.
Frequently Asked Questions
Is Agentic Commerce Actually Happening Yet, or Is It Still Theoretical?
Agentic commerce is live but early. Mastercard has offered Agent Pay since April 2025, Visa introduced its Trusted Agent Protocol in October 2025, Google announced the Universal Commerce Protocol in January 2026, and the 18 April 2026 edition of the Visa rules added requirements for agentic payment providers. Dispute data for agent orders is not published separately yet, so the evidence gap is a current operations task.
What Are Agentic Commerce Chargebacks?
Agentic commerce chargebacks are disputes that occur when an AI shopping agent makes a purchase on behalf of a consumer and the consumer later challenges the transaction with their bank or card issuer.
Who Is Liable When An AI Agent Makes A Purchase?
No formal liability shift exists for agent purchases, according to Worldpay’s July 2026 analysis, so disputes follow existing chargeback rules and merchants often absorb the loss unless their evidence shows the customer authorized the agent and the agent stayed within scope. Protocols such as Visa TAP and Mastercard Agent Pay build trust infrastructure but do not settle who pays.
Will Agentic Commerce Increase Chargeback Volume?
It can. Even without more actual fraud, customers have less memory of a purchase an agent made for them, which makes an unfamiliar-looking charge easier to dispute. Public dispute data for agent orders is not available yet, so track agent-initiated orders as a separate segment in your own reporting.
How Does Agentic Commerce Affect Conversion Rates and Dispute Rates?
Adobe Analytics found that AI-referred traffic to US retail sites converted 42% better than other channels in March 2026, after converting 38% worse a year earlier, so agent-assisted shopping can lift sales. No network has published a dispute rate for agent-initiated orders, so measure your own: tag agent orders, track their dispute rate against your other orders, and check the result against your ratio ceiling.
What Chargeback Rate Increase Should Merchants Tolerate From Agent-Led Transactions?
Do not treat the internal alarm in this model as an official network threshold. Agent-led orders remain subject to the applicable Visa and Mastercard event definitions, exclusions, reporting periods and count conditions. Calculate each program separately, and confirm the applicable country, regional and acquirer criteria; being below the chosen comparison line does not guarantee safety from processor review.
How Should Merchants Update Dispute And Chargeback Processes For Agent-Led Transactions?
Add agent-specific fields to the order record (agent platform, signature check result, mandate reference, notification time), agree record exports with each agent platform, keep the records for the longest dispute window that can apply, and rehearse a full response packet on a recent agent order. Add agent-specific exhibits to your response template and track agent orders as their own segment.
Can Merchants Win Chargebacks On AI Agent Purchases?
Yes, when the records show the customer authorized the agent and the agent stayed within the limits the customer set. A verified agent signature, a mandate with timestamp, a notification log and proof of delivery answer most unauthorized-purchase claims. If the order broke the customer’s limits or the instructions were ambiguous, the response is unlikely to hold and a refund is usually cheaper.
Does Compelling Evidence 3.0 Work For Agent-Initiated Orders?
Only in part. Visa Compelling Evidence 3.0 applies to dispute condition 10.4 and needs two matching data elements, one of them an IP address or device ID, across prior undisputed transactions. An agent’s IP address usually belongs to its platform rather than the customer, so account ID and delivery address matches carry more weight, and merchants should ask their acquirer how consumer identifiers passed by the agent are treated.
How Can Merchants Prevent Agentic Commerce Chargebacks?
Combine customer-facing guardrails such as spending limits and approval requirements with a post-checkout fraud scan before fulfillment, chargeback alerts that let you refund before a dispute is filed, and automated dispute management to handle any added volume.
What Evidence Do Merchants Need For Agentic Commerce Disputes?
Beyond traditional delivery and identity proof, merchants need the agent’s verified identity, the customer’s mandate showing what they authorized, a comparison of the order with the limits the customer set, notification timestamps showing the customer could cancel, and proof of delivery to the customer’s address.
Chargebacks?
No longer your problem.
Recover 4x more chargebacks and prevent up to 90% of incoming ones, powered by AI and a global network of 20,000 merchants.
.png)
