The shopper approved a maximum of $120, including delivery. The item cost $112. Shipping pushed the total to $126. The payment went through.

The card approval tells the merchant the charge was accepted. The customer's instruction tells a different story. A reviewer needs the price at the moment of purchase, what the shopper actually approved and whether the agent asked again when the total changed.

This distinction matters even in flows that authenticate the buyer. A useful agent prepares the purchase and brings the customer back in when a material term changes. A system that cannot reconstruct that decision will have trouble explaining the charge when the receipt lands.

How often did shopping agents fail, buy and repeat a success?Visa tested 22 models across 15 challenges, with three attempts per model and scenario. Each bar has its own denominator.
  • 42.8%Failed the shopping taskShare of all attempted runs.
  • 54%Completed a purchase despite failingShare of failed runs only.
  • 34.2%Passed on all three attemptsShare of model–scenario combinations.
Visa benchmark announcement · 5 Nov 2026

Two tests, two meanings of success

Visa benchmark, November 2026. Visa tested 22 models from 12 labs across 15 shopping challenges, repeating each model–scenario combination three times. 42.8% of runs failed. 54% of failed runs still ended in a purchase. All three attempts passed in 34.2% of model–scenario combinations. Multiplying the first two reported, rounded shares gives about 23% of all benchmark runs that both failed and purchased. This is a derived figure for the benchmark. It is neither a fraud rate nor a production payment-loss estimate. The test measures shopping-task success, which can fail even when a valid payment is processed.

Worldline, ING and Mastercard, June 2026. The partners reported a live end-to-end purchase for a Dutch merchant and ING cardholder. The agent proposed concert tickets within budget and the customer explicitly approved the transaction. This demonstrates feasibility for that partner flow. A single reported event does not establish repeatability, open participation, failure rates or results for unattended purchases.

Visa Intelligent Commerce. Visa documents tokenized credentials and customer-approved purchase instructions, alongside production onboarding and eligibility requirements. The merchant should verify what its network, issuer, acquirer and payment provider support in the intended country and flow. An authenticated agent can still select the wrong item, and a valid payment token can still be used for an order outside the shopper's instruction unless the system checks the order independently.

What the dispute reviewer will ask for

A folded receipt, two event tickets and a phone on a desk
The order, the approval and the receipt tell different parts of a transaction. A reviewer should be able to connect them. Editorial illustration

In the $126 example, the checkout should ask for fresh approval before it submits the charge. If payment times out after submission, the system should check the first attempt's state before a retry. The buyer should receive a receipt linked to the final order and payment state.

QuestionEvidence to retainTest condition
Who delegated what?Customer approval, agent identity, merchant, item scope, cap, currency and expiryRevocation arrives before execution
Was the attempt inside scope?Exact offer and order total checked at executionDelivery or currency changes the total
Did the payment settle once?Order and transaction IDs, status and reconciliation recordFirst request times out after approval
Can a reviewer resolve the dispute?Approval and order evidence with refund or dispute outcomePartial refund for one line on a multi-item order

Restrict the evidence to what reviewers need for their role. Keep payment secrets out of logs. Give the risk owner an explicit stop control and establish retention with the relevant payment and compliance teams.

Where permission is likely to move next

The easiest place to test delegation without a fresh approval is a routine purchase with stable terms: an approved merchant, a known category, a narrow cap and an expiry that the system can enforce. A one-off purchase with a changing total belongs in a flow where the customer can review the final terms.

Watch how often the merchant can explain an outcome from the retained evidence alone. If revocations take effect and every accepted purchase reconciles, a narrow recurring mandate becomes easier to defend. If the reviewer needs to reconstruct consent from a long chat, keep the buyer in the confirmation loop.

Earn each step of delegation

  1. Sandbox. Create a scenario set for authorization, cancellation, expiry, price drift, duplicate submission and partial refunds. Score authorization violations and duplicate settlements separately from purchase completion.
  2. Shadow approval. Run proposed actions through the controls without charging. Require a reviewer to reconstruct each approval and resulting decision from the retained records, without relying on the conversational transcript.
  3. Limited live traffic. Use one merchant, currency and permitted category, with explicit customer confirmation and reconciled daily settlement. Stop on an unapproved charge, duplicate settlement or missing evidence for an actual charge.
  4. Expand only with evidence. If a representative set of 1,000 independent trials has zero failures, the approximate one-sided 95% upper bound is 0.3%, using the rule of three. Correlated or incomplete trials weaken that inference. Size the sample to the loss tolerance and separately stress rare scenarios. Review customer recourse and partner responsibilities before widening authorization.

Evidence limits. The benchmark, documented controls and a production demonstration establish different things. None estimates the loss distribution for a merchant. Local measurement must cover settlement, refunds, disputes and customer harm.

Sources and related research