What is agentic commerce? A plain-English guide for 2026

Agentic commerce lets a shopper delegate parts of a purchase to an AI assistant. The assistant might compare products, check availability or arrange checkout. This guide explains how those steps work and what your business needs to support them.

Reviewed October 1, 2026 · 7 min read

Agentic commerce means an AI agent does shopping tasks on someone’s behalf. A customer might ask it to find a physiotherapy appointment this week or coffee under $20 a bag. The agent finds options, checks the details and, where the service and merchant support it, helps complete the purchase. The customer’s instructions and approval still determine what it may do.

The opportunity includes several different behaviors: researching a product, following a recommendation and authorizing a purchase. Keep those separate when reading market forecasts. A visit referred by an assistant is evidence of discovery, not proof that an agent completed checkout.

Below, we walk through a purchase, put the market figures in context and explain the integrations involved. For a closer comparison of the technical standards, see UCP vs ACP vs MCP.

How an agentic purchase actually works#

A purchase can involve the following stages. Some assistants handle only research; others can take additional steps through supported tools. Missing or outdated information can interrupt the process at any point.

1. Intent. The person tells their assistant what they want in natural language, with constraints: budget, timing, brand preferences, size, location.
2. Discovery. The agent looks for candidates through search, catalogs or connected tools. Accessible pages and accurate listings give it useful places to start.
3. Evaluation. It compares price, availability, reviews, delivery and fit with the customer’s requirements. Clearly labeled facts make comparison easier, although agents can still make mistakes.
4. Configuration. It checks the required variant, appointment, delivery window or service scope. Negotiation is possible only when the seller supports it and has defined acceptable terms.
5. Transaction. Depending on the integration, the assistant can prepare a cart, hand the customer to a checkout page or use an authorized purchase or booking tool.
6. After the purchase. Supported tools may let the customer check an order, request help or change a booking through the assistant. These actions need their own permissions and reliable records.

A chatbot on your site usually helps someone who is already visiting. An external assistant may help them choose a business before they reach a website at all. That makes your public product details and listings part of the customer experience, even when you do not see a pageview.

What the market figures tell us#

The figures below measure different things and should not be added together. McKinsey’s projections and Morgan Stanley’s outlook describe possible future markets. Adobe’s retail analysis measures referred visits, while Salesforce’s holiday report includes broader AI influence.

For a merchant, the useful question is which of these behaviors customers already bring to your business. Measure referrals, qualified enquiries, completed orders and the steps where people still need help. Those results are a better guide to your next investment than a market forecast on its own.

How ACP, UCP and MCP fit together#

The Agentic Commerce Protocol (ACP), introduced by Stripe and OpenAI, describes interactions between an agent and a merchant during checkout. A protocol specification and a particular shopping feature are separate things: a change to ChatGPT Instant Checkout does not automatically retire every ACP integration.

Google’s Universal Commerce Protocol (UCP) describes commerce capabilities that platforms and merchants can support together. Model Context Protocol (MCP) is broader: it lets compatible assistants discover and call tools, such as checking a calendar or requesting a quote.

Choose an integration based on the assistant, platform and customer task you need to support. Publishing a feed or running a server does not automatically enroll a business in every shopping service. The ACP guide covers that distinction, and the MCP server guide explains how tool access works.

Who feels it first#

Repeat purchases are a useful place to consider delegation. A customer can give clear instructions about the product, budget and frequency. Even then, the system needs rules for substitutions, price changes and final approval.

Services create a different opportunity. A request such as “find a haircut Thursday after 5pm” depends on location, availability and the type of appointment. Clear service descriptions and a booking integration can reduce the work involved. The service-business booking guide walks through those requirements.

What merchants should do now#

Start with the information customers need, then add supported ways to act on it:

1. Make key facts accessible. Publish prices, hours and policies as readable text, and review crawler rules by purpose. Search access and permission to use content for training are separate decisions.
2. Add accurate structured data. Describe your business, products and supported offers using appropriate schema.org types, keeping the markup consistent with the page.
3. Maintain listings and feeds. Follow the requirements of the shopping services relevant to your business. Submission, eligibility and display are separate steps.
4. Connect useful actions. Where an assistant supports your integration, let it check availability or prepare a booking with clear permissions and customer approval.
5. Measure results. Track identifiable referrals and completed actions, and investigate where customers or agents get stuck.

Helping people find your business in AI answers is related to, but different from, supporting a transaction. The generative engine optimization guide explains how to improve the information available to search and answer systems.

What businesses should protect#

Start with authorization. A request labeled as an AI agent is not evidence that a customer approved a purchase. Verify the caller where possible, check its permissions and require appropriate approval before charging money or making a booking.

Keep the customer informed about who is selling, what is included and where to get help. Clear records and reliable support matter when an assistant handles part of the interaction. Price is also easier to compare, so explain differences in service, delivery and product quality.

You do not need to predict the size of the market to begin. Accurate listings, readable pages and a dependable checkout already help customers. Start there, then add integrations where you can measure a useful result.