From Budget to Purchase: Mastercard’s Canadian AI Trial

An AI assistant has completed a purchase within a shopper’s stated budget in a transaction announced in Canada. The task went beyond recommending a product.

For an online retailer, the practical question is what happens after a recommendation. Does the customer place the order, or does the AI continue? The latter requires current stock, a final total and a reliable order result, as well as discoverable product information.

Mastercard, Flybits and Rogers Bank announced the transaction on September 21, 2026. This article explains its scope and proposes a preparation exercise for merchants, based on information checked on September 22. It is not a guide to an open merchant enrolment programme.[1]

What did the AI actually do?

According to the announcement, a Flybits-powered assistant followed a Rogers Red credit cardholder’s instructions, selected an eligible product within a specified budget, and completed the purchase using Mastercard Agent Pay. The transaction used validated consumer instructions and predefined spending limits.[1]

Agent Pay is Mastercard’s framework supporting payments when AI acts for a shopper. We have not inspected the purchased product or the actual checkout screen. The announcement does not establish a measured time saving or an increase in merchant revenue.[1]

The potential customer benefit is straightforward: less work between finding something suitable and buying it. Whether a particular store can offer that experience depends on its supported connections and operating conditions.

Three stages to examine in your own store

The following is our preparation exercise. It does not reproduce Mastercard’s enrolment instructions or the transaction interface.

1. Can a suitable product be selected?

A product price alone may not show whether a purchase fits a budget once delivery is included. Different sizes or colours may also have different prices.

Open one product’s public page, the page customers see rather than its administration screen. Check whether the name, price, availability, variants and delivery restrictions are clear.

Showing shipping costs at checkout is not automatically a problem. But an AI purchasing integration needs a defined point at which the relevant system calculates those costs. Do not add an invented flat shipping fee to the description just to make the information look complete.

2. Can the total and terms be checked before purchase?

A suitable product is not necessarily an order that can still be fulfilled. Stock may run out or delivery charges may change with the address.

Add the product to your own basket and inspect the steps before order confirmation. Note where quantity, item price, shipping, tax and delivery expectations become final. Use a test environment for transaction tests; otherwise stop before submitting an actual order.

The design question is when the AI should receive this information. Decide how a connected service should stop a purchase outside the permitted conditions or ask the person to confirm. A budget limit should not become optional simply because software handles the purchase.

3. Can the result be verified afterwards?

Reaching a payment page does not prove that an order exists. The store needs an order number associated with the correct items and quantities.

Inspect the confirmation process, including what happens when payment fails or an item sells out. For a future AI connection, add tests for tracing the same order across systems and preventing duplicate orders.

This inventory is useful before adopting a new integration. It identifies which information already exists and what still needs to be connected.

Does accepting Mastercard mean the store is ready?

The announcement does not establish whether any particular store is eligible. It describes a transaction connecting an AI service, a card issuer and payment infrastructure. It is not evidence that every store accepting Mastercard can immediately receive the same AI-led purchases.[1]

Check official information from the payment provider or commerce service under consideration:

Check	Decision it supports
Eligible countries, merchants and products	Whether the business can participate
Product submission and update method	Whether prices and stock can stay current
Order and payment connection	Whether the existing store can integrate
Consumer permission and spending limits	When to stop or ask for confirmation
Failure, cancellation and return handling	How staff can trace and resolve a problem

Leave unanswered questions marked as unconfirmed. Keep planned features separate from functions available under an actual contract.

Product discovery still comes first

A checkout connection is useful only when an AI can find a suitable product. Conversely, appearing in an AI recommendation does not mean the order process is connected.

First establish which sales channel will receive which product information. Then investigate its ordering requirements. The ACP merchant preparation guide explains this relationship for ChatGPT-related commerce.

For the role of permission records, read AP2 and purchase authorization. Those shared rules and Mastercard Agent Pay should not be treated as interchangeable service names.

A concrete next step is to document one product’s selection information, final checkout information and order result. That makes a future integration discussion specific enough to evaluate.