

Agentic commerce is a way of carrying out commerce tasks in which an AI agent can select and perform permitted steps on behalf of a buyer or merchant. A buyer-side agent might compare products or prepare a checkout; a merchant-side agent might carry out a configured support or order-related action. What the system can do depends on its tools, permissions, and the participating services.
A recommendation, a prepared checkout, and a confirmed purchase are different outcomes. A useful system should make clear what has happened, what still requires the buyer’s authorization, and which merchant and payment records verify the result.
Who does what in agentic commerce?
The term covers different arrangements, not one universal shopping flow. In a buyer-side scenario, the buyer specifies a task and sets or confirms the limits. An agent interface interprets the request and may retrieve product information or initiate permitted actions. The merchant supplies products, prices, order processing, and fulfillment. A payment provider handles the applicable payment processing and reports its status.
Responsibilities vary by product, merchant, region, and integration. An agent can recommend a filter without creating a cart. It can prepare a checkout without placing an order. An order record and an appropriate payment state are needed before claiming that an order and payment have been confirmed. Even then, confirmation does not mean the item has shipped.
Merchant-side workflows are another category. For example, an agent may help answer an order-status question or prepare a support case using approved data. That is useful commerce automation, but it does not by itself demonstrate delegated buying or payment.
For the broader concepts behind tool use and action limits, see What Is an AI Agent?.
The layers behind an agentic commerce workflow
Different standards describe different parts of an integration. They can coexist, and an implementation does not necessarily use all of them.
|
Protocol |
What it handles |
What the merchant must still verify |
|---|---|---|
|
A standard way for AI applications to connect to external tools, data, and workflows. |
Which tools are exposed, their permissions, the data they return, and whether actions are safe to execute. |
|
|
Documented commerce interactions, including a checkout specification. |
Whether the specific product, merchant, region, and integration support the intended flow, and who handles order and payment status. |
|
|
Commerce capabilities and interactions between consumer surfaces and businesses. |
Supported capabilities, implementation details, merchant policies, and reconciliation with existing systems. |
|
|
Verifiable authorization for supported agent payment flows. |
Whether the specific payment and merchant partners support the flow and whether the authorization is valid for the action. |
The Google UCP explanation describes compatibility with AP2 and several integration approaches. A protocol’s specification does not prove that a particular store has deployed it, that a specific purchase will be supported, or that its configuration is secure.
ACP: announcement and later product focus
OpenAI and Stripe introduced the Agentic Commerce Protocol in September 2025 alongside an initial checkout experience. In a March 2026 product discovery update, OpenAI described an expanded focus on product discovery and options involving merchant checkout. A specification and an earlier announcement should not be read as proof that the same checkout is available to every shopper or merchant today.
An illustrative purchase: from request to verified result
Suppose a buyer asks an agent to find a replacement water filter compatible with a particular appliance. The buyer specifies a maximum total budget and a delivery window. The following example uses a fictional merchant; it is not a tested CogniAgent workflow.
1. Discover a compatible item
The agent reads the buyer’s requirements and, if allowed, checks a product source. It identifies the appliance model and exact filter SKU rather than relying on a product title alone. If the compatibility information is missing or ambiguous, it asks for clarification or stops.
The result at this point is a recommendation. No order or payment has occurred.
2. Review stock and the full price
Before checkout, the system checks current stock, the item price, shipping, applicable taxes, and the delivery estimate through the sources available to the merchant. The buyer’s total budget applies to the resulting total, not just the product price.
It also records whether the buyer permits a substitute and, if so, under what conditions. A similar-looking filter is not automatically an approved replacement.
3. Obtain buyer authorization
The buyer reviews the item, total, delivery details, and any permitted substitution. The implementation may request an explicit approval or return the buyer to the merchant’s checkout. A stored preference does not automatically authorize an unlimited purchase.
If the authorized scope includes a spending limit or expiration time, the system must check those conditions before submitting the action. The exact mechanism depends on the implementation.
4. Attempt the permitted action
Only when the integration and the buyer’s authorization permit it can the merchant and payment systems attempt the order and payment. A checkout being created is a preparation step; it is not proof of a completed payment.
5. Verify the order and payment states
After an attempt, check the merchant’s order record and the payment provider’s result. A payment can be authorized or captured where those statuses apply; they are not interchangeable. The merchant may confirm an order before fulfillment begins. Report only the states that the relevant systems actually establish.
|
State |
What it indicates |
What it does not establish |
|---|---|---|
|
Checkout created |
A checkout or purchase session was prepared. |
That an order or payment exists. |
|
Payment authorized |
The provider reports an authorization, where applicable. |
That funds have been captured or goods shipped. |
|
Payment captured |
The provider reports capture, where applicable. |
That fulfillment is complete. |
|
Order confirmed |
The merchant has confirmed an order in its system. |
That the order has shipped or arrived. |
|
Fulfillment |
The merchant reports the relevant shipping or delivery state. |
A successful return, replacement, or other later event. |
A checkout, a payment state, and a merchant order are separate stages that require separate checks.
Handle exceptions before retrying
- The price increases: recalculate the complete total. If it exceeds the buyer’s authorized scope, ask for fresh approval or stop.
- The item is out of stock: do not substitute unless the buyer’s instructions and the implementation permit the exact substitution. Otherwise, ask or return to discovery.
- Authorization is denied or expired: do not submit the purchase. Request a new authorization if appropriate.
- A request times out: the outcome may be unknown. Check the merchant’s order and payment state before retrying, and use the integration’s duplicate-protection mechanism where available.
These branches are part of the design, not exceptional details to add after launch. A buyer needs a clear status even when the system cannot yet determine the outcome.
A real-world example: Alexa for Shopping
Amazon describes Alexa for Shopping as combining Rufus and Alexa+ on the Amazon Shopping app and website. Its published description includes product comparisons, deal finding, cart building, and routine purchases, with availability dependent on the service and market.
This is a vendor-described example of shopping assistance and selected shopping actions. It does not establish that all merchants, products, or other agent platforms support the same experience. Check the provider’s current availability and the specific action offered before using it as a model for an implementation.
How merchants can start with a controlled workflow
Start with a narrow problem rather than delegating an entire purchase. For example, expose read-only product or order data in a sandbox, then test one permissioned action with explicit approval and a manual fallback.
Before expanding the scope, check:
- Freshness: where stock, price, shipping, and applicable taxes come from and when each was updated.
- Authority: who may approve the action, what it covers, and any spend or time limit.
- Failures: what happens when access is denied, data is missing, the buyer declines, or a request times out.
- Duplicate protection: how the system checks whether a previous attempt already produced an order or payment.
- Reconciliation: which merchant order record and payment status are used as evidence.
- Ownership: who handles order support, cancellations, and exceptions after the interaction.
Use test inputs that exercise these cases before enabling a live action. The AI agent testing checklist provides a wider framework for pre-launch checks.
Where CogniAgent fits: a merchant-side support example
CogniAgent’s documented eCommerce Support Pack is an example of a merchant-side support workflow, not a checkout or payment implementation. Its template page describes a shared workbook with order and policy data, a verification worksheet for configuring checks and permitted disclosures, and a case workflow for issues requiring support. The page states that a failed identity check does not reveal order details, and describes deduplication and support notification as part of that configured template.
This example has not been run for this article. Before adopting it, inspect the template’s requirements, connections, permissions, worksheet settings, and actual case records in your own test environment. A generic connection to a commerce system does not prove that checkout, payments, or operator takeover are available.
An illustrative support flow based on the documented template; it is not a checkout or payment flow.
For the platform’s wider merchant context, see eCommerce Automation. The AI Use Cases in eCommerce article covers broader applications beyond the purchase flow.
An AI agent builder or no-code AI builder platform can help configure parts of AI workflow automation, but the permitted action, its authorization, and the external record remain the relevant boundaries. Start with a process whose result you can verify and whose exceptions have an owner.


