Agentic checkout

Agentic checkout is a purchase workflow in which an authorized AI agent prepares or completes an order for a buyer within defined limits, such as budget, merchant, or timing. The agent may assemble a cart and hand the buyer to the merchant’s checkout, or complete the purchase inside an AI interface. 

Either way, the merchant’s systems confirm the product, price, availability, and final total before accepting and fulfilling the order. Buyers can approve each purchase at checkout or set rules in advance for certain orders. Agentic checkout is the transaction stage of agentic commerce, not the entire journey of finding, comparing, and buying products.

What agentic checkout is and what it is not

Agentic checkout begins where recommendation ends. Much of conversational AI in e-commerce helps buyers find, compare, and understand products, then leaves the purchase to them. Agentic checkout asks for more. The agent acts on the order itself through connected commerce systems by building the cart, confirming the selected variant, and moving the cart into checkout for the buyer. That work depends on live product, pricing, inventory, promotion, shipping, and policy data from the merchant’s systems.

Setting it beside neighboring terms shows exactly where that difference sits.

Embedded checkout is the term most often confused with agentic checkout. “Embedded” only tells you where the checkout screen sits, such as a buy button inside a social or messaging app. “Agentic” tells you who does the work. Say a buyer asks an assistant for running shoes in size 9, under $120, delivered by Friday. If the agent finds the pair, checks the size and price, and builds the cart, that’s an agentic checkout, whether the buyer pays in the chat or on the store’s site. If the buyer does all of that themselves and simply pays inside an app, it’s an embedded checkout. Some vendors call the agent-led version agentic AI checkout, which means the same thing.

In e-commerce today, both versions are already in use. In Google’s AI Mode and Gemini, eligible US shoppers can pay without leaving the AI interface through UCP, with Google Pay and Google Wallet providing the payment and shipping details. 

In ChatGPT, the agent helps choose, and the buyer finishes on the merchant’s own checkout, a change OpenAI made in March 2026 when it stepped back from in-chat Instant Checkout. 

For merchants, this choice matters. When checkout happens on a third-party platform, merchants may have less direct control over the customer interface, behavioral data, and post-purchase relationship, a central part of the disruption agentic commerce brings to retail.

How agentic checkout works

Agentic checkout starts with a buyer instruction. The instructions may be specific, such as “reorder the same coffee pods from my usual store,” or broader, such as “find a carry-on suitcase under $250 that arrives before Friday.” The agent’s first job is to turn that natural language request into something it can act on.

From there, the agent works through the merchant’s live commerce services. A product description helps with discovery, but it can’t confirm that an item is still in stock or still at the displayed price. The merchant’s catalog, pricing, inventory, and order systems remain the source of truth throughout six stages:

1. Interpret the request. The agent identifies the product attributes, budget, merchant preferences, delivery deadline, payment method, and any standing approval rules.
2. Check the product and delivery options. It confirms that the SKU and variant are in stock, using live data from the merchant’s inventory management system. It then checks price, tax, shipping, delivery date, and return terms.
3. Build the cart and checkout session. The agent adds the confirmed items and applies only the offers that the merchant’s pricing and promotion rules allow. It then creates or updates a checkout session with the merchant.
4. Present the final order. The buyer sees exactly what will be bought, how it will arrive, and the total before submitting anything. In a conversational shopping flow connected to the storefront, this review happens inside the chat.
5. Record authority and pay. The buyer approves the order, or the agent acts under a preapproved rule limited by merchant, product category, amount, timing, and payment method. Payment travels as a token, so the agent never sees the buyer’s card details.
6. Accept and confirm. The merchant accepts or declines the order, processes payment, reserves inventory, and sends confirmation. From there, its order management system handles fulfillment, tracking, and any later changes.

When a step fails

A reliable flow doesn’t assume every action will succeed. Inventory can change after a product is selected, a promotion may expire, shipping may miss the delivery date, or payment can be declined. A network timeout can even leave the buyer unsure whether the order was placed.

Each action needs a clear recovery path. The agent requests fresh merchant data before retrying a time-sensitive step. Idempotency keys, a unique ID for each order attempt, stop a retry from creating a duplicate order. When the price, availability, or delivery promise changes, the buyer is shown the change. If the order can’t proceed within the buyer’s limits, the agent pauses and asks for a decision. It never substitutes a product or charges a different amount without approval.

This stage is also where human escalation matters. The buyer should always be able to review the cart, cancel the process, or move to a support channel when an exception can’t be resolved safely.

Protocols and architecture behind agentic checkout

Every stage of the checkout flow involves the agent talking to a merchant system: checking stock, creating a cart, passing on the buyer’s approval, and requesting payment. If each AI platform phrased those requests differently, merchants would need a separate integration for each one. Open protocols lower the integration burden by giving agents and merchants a common set of rules for each exchange. Three are shaping agentic checkout today, and each handles a different part of the transaction.

Agentic Commerce Protocol (ACP)

ACP lets merchants share products and offers with ChatGPT so they can appear in shopping recommendations. Since March 2026, OpenAI has emphasized product discovery, with ChatGPT generally directing shoppers to merchant checkout. The ACP specification still includes checkout capabilities, but those should not be confused with ChatGPT’s default shopping experience. In-chat checkout depends on the supported merchant integration and eligibility.

Universal Commerce Protocol (UCP)

UCP covers the whole purchase, from finding a product to tracking the order after it ships. It’s what lets shoppers buy without leaving Google’s AI Mode or the Gemini app. Merchants can either let Google display the checkout screen or keep their own checkout inside the experience.

Agent Payments Protocol (AP2)

AP2 answers one question: did the buyer really approve this payment? It records the buyer’s instructions and payment conditions through digitally signed mandates that merchants and payment providers can verify. AP2 can also plug into the UCP checkout. 

Two more protocols often appear alongside these:

• MCP lets an agent connect to tools and business systems, such as product search, inventory, or order lookup. It is an integration layer, not a payment or commerce protocol.
• A2A lets one agent pass work to another, for example, a shopping agent handing a payment task to a separate payment agent. 

Neither handles the purchase itself. They carry requests to the systems that do. ACP and AP2 also take different approaches to approving payments, which is worth comparing before choosing which agentic payment protocols to support.

The systems behind the transaction

Protocols carry the messages, but the work happens in the merchant’s own systems. A production setup usually includes:

• An agent interface that understands the request and coordinates each step
• Identity and policy controls that confirm whom the agent represents and what it’s allowed to do
• Catalog and search services that hold product details and variants
• Pricing, inventory, and promotion services that return what’s available and at what price
• Cart and checkout services that calculate the final total and track the purchase
• A payment provider that protects card details and processes approved payments
• Order management and post-purchase services that handle tracking, cancellations, refunds, and returns
• Monitoring and audit services that record every request, approval, response, and failure

The agent reaches these services through controlled APIs and events. An agentic commerce platform built on a composable stack can then support several protocols, including both owned and third-party agent experiences, without rebuilding catalog, cart, payment, and order services for each new channel.

Benefits and high-value use cases

Agentic checkout pays off when a buyer knows what they want, and only the transactional steps remain. Re-entering addresses, picking shipping methods, recalculating totals, and typing payment details all take time. When an assistant handles those steps directly, journeys that would otherwise stall can become completed orders. Merchants can also capture sales that start on external AI platforms, provided their products stay visible to AI agents comparing options.

These gains should be tested, not assumed. Faster ordering helps only if the agent selects the right variant, and larger baskets add little value if return rates rise.

Exact-fit purchases where wrong parts mean returns

In categories like automotive parts, near matches lead straight to returns. For checkout to work, the agent must carry three details into the cart:

• The customer’s vehicle specification
• The confirmed part number
• Proof that the part fits that specific vehicle

This step relies on accurate fitment data tied to live inventory. One automotive retailer introduced a conversational ordering agent on WhatsApp. Customers identify parts through text, photos, or voice notes, confirm compatibility through a short questionnaire, and complete the purchase without leaving the app. Order placement now takes 3 to 5 seconds, 95% faster than handling it manually.

Routine replenishment on standing rules

Household goods, pet supplies, and personal care items follow regular purchase cycles. A buyer sets simple rules once: preferred items, store choices, and a spending limit. When supplies run low, the agent prepares the cart and asks for confirmation in a single prompt.

Saved sizes, formulas, and delivery addresses carry straight into checkout. At the payment step, the agent can also show the total with and without eligible loyalty rewards, so the buyer decides whether to apply points before completing the order.

Price-triggered buying within preset limits

Some buyers wait for sales on expensive items. With price-triggered checkout, the shopper sets a target price and a delivery deadline. The agent checks live pricing and submits the order only when the merchant’s total meets both. Because payment can happen while the buyer is offline, the transaction runs on a payment limit authorized in advance.

B2B procurement on contract terms

Business buyers work within rules that consumer retail doesn’t have:

• Approved vendor lists and negotiated prices
• Pack sizes and minimum order quantities
• Delivery schedules
• Department budgets and approval limits

Finding the right products is the first hurdle. In commercial kitchens, staff might ask for “rags” while the catalog lists “microfiber cloths.” Search and catalog enrichment map everyday terms to the correct items.

Payment works differently too. Many B2B purchases run on credit terms, with a purchase order now and an invoice later. The agent checks that the order stays within the account’s credit balance and spending cap, then submits the purchase order to the merchant’s billing system. Anything over the budget limit goes to a manager for approval, a practical use of agentic automation in routine purchasing.

Assembling one basket across several merchants

A request like “everything needed for a weekend camping trip” often draws on several stores. An agent can assemble carts across retailers, with each merchant confirming stock, calculating tax, and processing its own payment. Separate delivery fees and return policies can quickly outweigh the convenience, so the agent shows itemized shipping costs and terms before the buyer approves.

Trust, security, and governance

Agentic checkout depends on every party being able to verify who authorized a transaction. Giving software access to payment credentials introduces real risks, including duplicate orders, wrong variants, and unauthorized charges. Managing those risks takes controls in four areas: authorization, authenticity, payment data, and accountability.

Buyer authorization and limits

An agent should never act on open-ended assumptions. Every checkout needs explicit consent or a rule the buyer set in advance. Those rules go further than naming merchants and categories. They also set:

• spending caps per order and per billing cycle
• an expiry date for standing permissions
• a threshold that triggers a new prompt when the total exceeds the estimate
• a clear way for the buyer to revoke access at any time

If an order falls outside these limits, the agent pauses and asks for live confirmation. Designing for clear consent and user verification keeps convenience from overriding the buyer’s control.

Agent authenticity

Merchants need to tell legitimate shopping agents apart from scrapers and malicious bots. An emerging practice, often called Know Your Agent (KYA), gives agents verifiable digital identities. Using signed requests and public-key cryptography, the merchant’s gateway can confirm three things before accepting a cart:

• The agent is authentic
• It is acting within its permissions
• It represents a verified customer account

Card networks are building their own agent recognition schemes, such as Visa’s Trusted Agent Protocol and Mastercard Agent Pay.

Scoped payment credentials and data minimization

Sharing raw card numbers with an AI agent creates avoidable fraud and PCI compliance exposure. Secure setups use network tokenization instead. The payment provider issues a restricted token that is valid only for one merchant, one checkout session, and a set amount. The agent coordinates checkout without ever seeing or storing card details.

The same principle applies to other data. Addresses and order history go only to the merchant fulfilling the order. Conversation logs follow the same privacy rules as any other customer record.

Accountability and dispute management

When an item arrives damaged, late, or in the wrong size, responsibility should be clear:

• The merchant typically remains the merchant of record and handles catalog accuracy, fulfillment, cancellations, and returns.
• The agent platform is responsible for passing accurate session data, quantities, and buyer confirmations.
• Audit logs should capture each buyer instruction, agent action, merchant response, and mandate.

If a buyer disputes a charge, those records help establish proof of authorization with banks and card networks. Exact cost allocation depends on contracts, network rules, and local law. For routine exceptions, post-purchase support workflows can process returns and refunds through chat or messaging under the store’s standard policies.

How merchants can prepare and implement

Businesses implement agentic checkout by exposing the commerce logic they already run through secure, stable interfaces, then proving it on one journey before expanding. Seven steps take a merchant from data to launch.

1. Pick a narrow, measurable starting flow. Choose a journey with high intent and clear rules, such as routine replenishment or reordering from order history, and set a baseline before building. Leave cross-merchant and multi-category baskets for later.
2. Make catalog and policy details machine-readable. Wrong-variant orders usually trace back to vague product data. Sizes, materials, pack counts, compatibility tags, and return terms need to sit in structured fields an agent can compare. Filling those gaps by hand is slow. G-Star used AI catalog enrichment to extract attributes from images and existing content, cutting product onboarding from weeks to days.
3. Expose live commerce capabilities through stable APIs. Owned and external agents need direct, consistent access to inventory, promotions, shipping quotes, and checkout sessions. Merchants running an aging monolith can get there without pausing trade. A British luxury jeweler moved from a decade-old SAP Hybris platform to a composable commerce architecture with 15 external integrations, and online sales continued throughout.
4. Connect tokenized payments and consent controls. Confirm that your payment provider supports network tokens and agent-initiated transactions. Then enforce consent rules and spending limits in your own checkout, independent of the agent platform.
5. Set up runtime governance and monitoring. Log every agent action, approval, and merchant response, and track each session end to end. Continuous AI agent evaluation helps catch drift and failed calls early. It also gives teams a kill switch or human handoff before shoppers feel the problem.
6. Test failure and recovery on a real SKU. Run live transactions from instruction through delivery, cancellation, and refund. Then deliberately break them: stock that runs out at the moment of payment, an expired promotion, a timeout during card capture, and a declined payment.
7. Expand across channels and protocols. Once the first flow holds, add support for standards such as UCP or ACP to reach shoppers on external AI surfaces. Measure each new channel against the original baseline.

A practical first move is an agentic commerce readiness assessment across your catalog, checkout APIs, payments, and order systems. It surfaces integration gaps before any agent code is written.