What is Agentic Commerce and Why Should I Care?

Agentic commerce is the use of AI agents to carry out shopping tasks on behalf of a person or business, from finding and comparing products to preparing or completing a purchase within defined permissions. The agent works toward a buyer’s goal, uses available information and tools, and may take several steps to reach a result.

You should care because an AI assistant may help a buyer decide what to buy before that buyer reaches your website. For retailers and brands, that makes product accuracy, availability, delivery promises and buying controls part of the customer experience wherever the conversation starts.

This guide explains the business implications and a sensible starting point. For help assessing your own catalog and systems, see Sitation’s Agentic Commerce expertise.

Jump to: How it works · What makes it different · Platform comparison · Business implications · Product data · Readiness plan · Risks · FAQ

How agentic commerce works

Imagine a facilities manager asking: “Find replacement filters for these air purifiers, in the right size, from an approved supplier, delivered by Friday. Show me the total before ordering.” This is an illustrative scenario, not a Sitation customer result.

A useful agent would identify the equipment models, check compatibility and pack quantities, compare available offers, and return a proposed order with its evidence. If compatibility is unknown, it should ask for clarification or stop. If the buyer approves, an authorized connection could submit the order.

1. Understand the request. Capture the intended use, exact product constraints, budget, destination and deadline. Ask about missing essentials.
2. Find and compare. Consult available product pages, feeds or connected catalogs. Match the specific variant and explain why each candidate fits.
3. Check the offer. Revalidate price, stock, delivery, seller and return terms using the systems responsible for those facts.
4. Obtain authority and act. Respect approval rules. Depending on the integration, hand off to checkout, prepare a cart or submit an authorized order.
5. Confirm the outcome. Return a reliable order confirmation or a clear failure, with a route to support. Do not equate an attempted action with a completed purchase.

The important distinction is the degree of delegation. Assisted discovery, cart preparation and authorized purchasing require different controls. Your first project can improve discovery without enabling autonomous buying.

How is this different from automation, chatbots and recommendations?

Rules-based automation

Rules-based automation runs a predefined action when a condition is met, such as a scheduled reorder. An agent can choose intermediate steps to meet a goal. Deterministic rules still belong around approvals and orders.

A conventional chatbot

A conventional chatbot answers a question or follows a conversation flow. An agent may call tools and take actions. A chat interface alone says little about its authority.

A recommendation engine

A recommendation engine ranks or suggests products using defined signals. An agent may combine recommendations with questions, offer checks and an authorized next step.

AI content generation

AI content generation drafts descriptions, extracts attributes or summarizes information. This prepares information for commerce. It becomes part of an agentic workflow when connected to a goal, tools and controlled actions.

These capabilities can work together. A well-designed shopping agent can use an existing recommendation service, a catalog feed and a conventional checkout. It does not need to replace every component of an eCommerce stack.

What commerce and payment protocols change

Payments: proving what the buyer authorized

Google introduced Agent Payments Protocol, or AP2 ↗ (opens in a new tab), on September 16, 2025. It describes signed mandates that record a buyer’s intent and approved cart, including scenarios where permission is granted ahead of time. The point is evidence of authorization: what the buyer allowed, under which constraints, and what was purchased.

AP2 is a protocol initiative, not proof that a particular store supports delegated purchases. Payment providers, merchants and agent builders still have to implement the controls and agree how exceptions are handled.

What the protocol names mean for a business decision

• ACP: Agentic Commerce Protocol. An open commerce protocol ↗ (opens in a new tab) developed by OpenAI and Stripe. Evaluate the integration your chosen channel actually supports.
• UCP: Universal Commerce Protocol. A common model for commerce capabilities across platforms. Choose the capabilities your business needs; a catalog connection and an order-writing connection have different risk profiles.
• AP2: Agent Payments Protocol. A framework for verifiable payment authority. It addresses permission and evidence, not whether a product is correctly described.

Do not begin by picking an acronym. Begin with the buyer journey, the channel you want to support, the data it needs and the actions you will allow. Standards help systems communicate; they do not fix unreliable product information.

Why retailers, brands and distributors should care

Retailers: the shopping journey may begin elsewhere

An assistant can become the place where a shopper defines a need and narrows the options. Your store must still supply a credible offer and fulfill it. Review how loyalty, promotions, delivery promises, substitutions and returns survive the handoff. Decide who handles a customer who says, “The agent told me this would fit.”

The opportunity is to reach buyers with a more specific need. The risk is losing context between discovery and checkout, or accepting an order based on stale information. Evaluate margin, completed orders and service cost alongside traffic.

Brands: differentiation needs supporting facts

A phrase such as “premium performance” does not answer whether a product fits a particular machine or use case. Store the dimensions, materials, compatibility, care instructions and substantiated claims that explain the difference. Give retail partners consistent product facts, with the right images and variant identifiers.

Useful prose still matters. Explain who the product is for, when it is unsuitable and what evidence supports a claim. Avoid manufacturing claims or keyword-heavy descriptions in an attempt to influence an assistant.

Distributors and manufacturers: fit and commercial terms matter

B2B decisions can depend on replacement-part relationships, unit conversions, minimum quantities, account-specific prices and approved supplier lists. A low price on the wrong unit of measure is not a useful recommendation. Start with tightly defined categories and preserve the buyer’s procurement approvals.

What agentic commerce needs from PIM, MDM and product content

PIM manages enriched product information. MDM establishes governed identities and relationships across systems. DAM manages approved assets. An agent needs reliable access to their outputs, alongside current commercial information from commerce, inventory and order systems. The PIM should not be assumed to own every live price or stock value.

• Identity and variants. Stable product and variant IDs, legitimate GTINs where assigned, manufacturer part numbers and clear parent-child relationships prevent the wrong item from being matched.
• Comparable attributes. Explicit units, dimensions, materials, compatibility and pack quantities make comparisons meaningful. Define “unknown” separately from “not applicable.”
• Evidence and content. Approved images, instructions, certifications and source-backed claims help explain fit. Keep the source, approval and review date with sensitive attributes.
• Current offers. Price, currency, availability, shipping eligibility and return terms need an owner and a refresh policy. Validate material changes again before purchase.
• Consistent delivery. Product pages, structured markup, retailer feeds and APIs should express the same approved facts. Monitor rejected records and stale exports.

For example, OpenAI’s product-feed specification ↗ (opens in a new tab) distinguishes an item, its variant group and a seller’s offer. It requires stable item identifiers and explains that search eligibility does not guarantee display. This is why identity and publishing rules deserve as much attention as descriptions.

Google’s merchant-listing documentation ↗ (opens in a new tab) explains how product structured data can communicate offers and product details to Search. Treat markup as one publishing route. It does not grant an agent authority to order, and it does not replace an accurate underlying catalog.

AI can help extract attributes and draft explanations from approved sources. Review the result at a level appropriate to the consequence of an error. See our guide to AI for product content, enrichment and human review for that workflow.

How the major platforms compare

Reviewed October 2026. These are documented capabilities and rollout statements from the linked primary sources, not a hands-on certification of every integration. Availability is specific to the merchant, buyer, market and checkout route.

Platform	Documented experience	Availability and limits
ChatGPT / OpenAI	Product discovery with merchant checkout OpenAI’s March 24, 2026 update describes visual discovery and comparisons. It shifts emphasis from the initial Instant Checkout toward merchants’ own checkout experiences; Walmart has a merchant-specific experience inside ChatGPT.	Checkout differs by merchant and integration. Do not assume a product shown in ChatGPT can be purchased there, or that an agent has permission to spend. OpenAI product discovery ↗ (opens in a new tab)
Google	Live integrations, with eligibility and rollout limits UCP supports connections between merchants and Google shopping experiences. Google’s integration documentation offers native checkout and an optional embedded route for selected, approved merchants.	For Shopify’s documented Google route, direct checkout is rolling out to eligible US stores and US buyers; a valid Merchant Center feed is required. This is not a guarantee that every Google shopping feature is available to every user. Google merchant integration ↗ (opens in a new tab) · Shopify’s Google eligibility ↗ (opens in a new tab)
Microsoft Copilot	Checkout available, subject to onboarding Microsoft documents Copilot Checkout on Copilot.com and the mobile app. The merchant remains the seller; product discovery uses web information and Merchant Center feeds.	Current eligibility is English-language merchants selling to US buyers in USD. A Microsoft Merchant Center account and feed are required; payment-provider onboarding applies. UCP-rich feed features are marked as a selected-customer pilot. Microsoft current requirements ↗ (opens in a new tab)
Amazon	Alexa for Shopping: US availability Amazon’s May 13, 2026 announcement combines Rufus and Alexa+ as Alexa for Shopping, with comparisons, price alerts and scheduled actions. Eligible Buy for Me purchases can extend to other retailers.	The announcement covers signed-in US customers on Amazon’s app, website and Echo Show, with rollout that week. Scheduled cart additions can require review; this is not blanket permission for autonomous purchases. Amazon’s Alexa for Shopping announcement ↗ (opens in a new tab)
Meta	Merchant route documented; surface availability varies Shopify documents discovery and direct checkout on Meta surfaces, including Muse when available. This is a merchant distribution route, not evidence that every Meta assistant supports the same shopping workflow.	Shopify lists direct checkout for customers in the US, Canada and Mexico, with eligible catalog products and checkout enabled. Surface availability and product restrictions still apply. Shopify’s Meta integration and limits ↗ (opens in a new tab)

Shopify’s agentic storefronts ↗ (opens in a new tab) distribute eligible merchant catalogs across several of these channels. Shopify is an enabling commerce layer, not another interchangeable shopping assistant. Its documentation distinguishes ChatGPT discovery and store handoff from direct checkout on supported Google, Microsoft and Meta surfaces.

The practical choice is which buyer journey you can support accurately. Verify eligibility, feed requirements, payment onboarding, returns and attribution for that route before investing in an integration.

Logos identify the platforms discussed; they do not imply endorsement of this guide. All trademarks belong to their respective owners.

A practical readiness plan

Use a small pilot to answer a business question before committing to a broad rollout. The following sequence is a proposed working approach, not a promise about how quickly every business can launch.

1. Choose a buyer problem and an owner. Select one category and one journey, such as finding a compatible replacement part. Agree whether success means a correct shortlist, a completed cart or a verified order.
2. Build a test set from real questions. Include ordinary requests, ambiguous requests, unavailable items, incompatible products and requests outside policy. Record the expected answer and the evidence needed to support it.
3. Audit the record and the offer. Trace each answer back to its source. Fix missing attributes, unit inconsistencies, duplicate variants, stale feeds and unclear responsibilities.
4. Choose one supported channel. Verify its current documentation and eligibility. Start with discovery or cart preparation. Have the commerce owner and security team review any proposed order-writing permission.
5. Test the full handoff. Check what happens when price changes, stock disappears, an approval is declined or a request is retried. A safe refusal is better than an unsupported substitution.
6. Review outcomes and decide whether to expand. Assign owners to failures, repeat the same test set after corrections, and expand only when accuracy and business results justify it.

Measure accuracy and commercial value separately

Keep a repeatable evaluation log with the prompt, date, channel, product, expected facts and observed result. Record your catalog changes so a before-and-after comparison has context. Assistant answers vary; a single prompt run is not a reliable ranking measurement.

• Data quality: required-attribute completeness, identifier errors, rejected feed records and freshness against your agreed targets.
• Answer quality: correct product and variant, supported claims, valid availability and appropriate handling of unknowns.
• Transaction quality: approved actions, completed orders, duplicate attempts, cancellations, returns and support exceptions.
• Business value: attributable visits and orders where observable, margin after fees and fulfillment, and qualified inquiries for higher-consideration sales.

Do not treat bot fetches as shoppers or product mentions as revenue. Where attribution is missing, label it. Compare the pilot with a relevant baseline and avoid assuming every observed improvement was caused by the agentic change.

Risks and limits to plan for

• Incorrect answers. An agent may infer compatibility or repeat an unsupported claim. Require source-backed facts and explicit uncertainty; send high-consequence decisions to a person.
• Stale prices and inventory. A valid discovery result can become an invalid order. Recheck commercial facts at the transaction boundary.
• Overbroad authority. Separate permission to read, build carts, purchase and cancel. Use spending limits, approved suppliers and clear consent records.
• Malicious instructions in retrieved content. Treat product pages and supplier documents as information, never as instructions that can override the buyer’s intent or security policy.
• Duplicate or partial actions. A retry must not create a second order. Keep action logs, reliable confirmation and an exception process.
• Privacy and customer trust. Share only the information needed for the task. Make the seller, costs, returns and approval step clear.
• Platform dependency. Support, eligibility, commercial terms and checkout paths can change. Maintain a usable website and supported alternatives.

An agent can make it easier to act on good information, but it can also repeat a bad record across more interactions. Give product-data ownership, operational monitoring and customer support a place in the project from the beginning.

Turn the first test into a useful decision

Ask your commerce and product-data leads to show one buyer request moving from an approved product record to a verified offer. Where does the evidence break? Who owns the correction? What would the agent be allowed to do next? Bring those questions to the team responsible for the channel.

If the use case is still unclear, Sitation’s AI digital transformation services cover workshops, pilots and implementation. If you already need agents working with governed PIM or MDM information, Managed Agents explains the ongoing build-and-operate model. The Agentic Commerce expertise page connects these capabilities to commerce readiness.

For a closer look at one ecosystem option, see our ReFiBuy partner page. Evaluate any partner against your catalog, channels and test cases before choosing a platform.

Frequently asked questions