What Is Agentic Commerce? How AI Agents Change Online Buying
your project?
Key takeaways
- What it is: agentic commerce is shopping in which an AI agent searches, compares and buys on a person’s behalf. McKinsey describes it as a model where “AI agents shop, negotiate and transact on behalf of humans.”
- Why it matters now: shoppers already arrive from AI assistants. Adobe measured generative AI referrals to US retail sites growing 693% year over year in November and December 2025.
- How it works: open protocols such as UCP and ACP let an agent read a catalog, build a cart and start a checkout. Other protocols handle consent and payment trust.
- What stays with you: the store remains the merchant of record and keeps customer data, fulfillment and support.
- Where it stands: adoption is early. In March 2026 OpenAI moved ChatGPT shopping toward discovery, and checkout inside AI surfaces is still taking shape.
For twenty years, online shopping has meant a person typing a query, opening tabs and comparing prices. Agentic commerce changes who does that work. The shopper states a goal, and software carries it out.
What agentic commerce is
A traditional shopping assistant answers a question. An agent acts on one. Asked for a waterproof jacket in medium for a trip next month, an agent can search several stores, compare prices and delivery dates, add the best match to a cart and, with the shopper’s authorization, complete the purchase.
Three ideas define it:
- Delegation. The shopper hands over a task, not a single query.
- Action. The agent does things in the world: it calls store systems, builds carts and starts payments.
- Authorization. The shopper sets limits and approves the spending.
Example: “Find a waterproof jacket in medium and order it with my saved address.” The agent searches catalogs, picks a variant that is in stock, builds a cart and starts checkout with the store that has it.
Three stages of agent involvement
Agents do not jump straight to buying. Most stores will meet them in three stages, and each stage asks more from the store.
- Recommend. The agent suggests products and the shopper buys on the store’s site. Product data quality decides whether you appear.
- Assemble. The agent builds a cart, possibly across several items. Live stock and consistent prices matter here.
- Buy. The agent completes checkout with the shopper’s authorization. This needs a callable checkout and a compatible payment setup.
Most activity today sits at the first stage. The second and third are where protocols are doing the most work.
What an agent does in a purchase
- Understands the request and turns it into search criteria.
- Finds products by reading feeds, catalogs and product pages.
- Compares options on price, delivery, returns and reviews.
- Builds a cart with the chosen items.
- Pays with a scoped credential that the shopper has authorized.
- Tracks the order and reports status back to the shopper.
Every step reads data that a store controls. A missing return policy, a stale stock status or a checkout that needs a browser session can stop the chain.
The protocol layer
Agents cannot work with each store through custom code. Open protocols give them one shared set of rules. Different protocols answer different questions.
- UCP (Universal Commerce Protocol) was announced by Google on 11 January 2026 and co-developed with Shopify. It covers catalog, cart, checkout, identity linking and orders. Our guide explains
- ACP (Agentic Commerce Protocol) was built by OpenAI and Stripe and announced on 29 September 2025. It covers a product feed, checkout endpoints and delegated payment.
- AP2 (Agent Payments Protocol) is Google’s protocol for proving authorization. It uses signed mandates, which are verifiable credentials that record what the shopper allowed. It integrates with UCP.
- Card network programs. Visa and Mastercard run their own programs, Visa Trusted Agent Protocol and Mastercard Agent Pay, to help merchants and issuers trust agent-initiated payments.
Under both UCP and ACP, the merchant stays the merchant of record. We compare the two in ACP vs UCP.
Agentic commerce in numbers
Forecasts are large, and early traffic data points the same way. Both come with caveats.
- McKinsey (October 2025) estimates that agentic commerce could orchestrate up to $1 trillion of US retail revenue by 2030, and $3 trillion to $5 trillion globally.
- Adobe Analytics found that generative AI referrals to US retail sites grew 693% year over year in November and December 2025. AI-referred visitors converted 31% better than other visitors, and they were 33% less likely to leave immediately. Adobe’s sample covers more than 1 trillion visits to US retail sites.
Forecasts are estimates. Referral growth from a small base is a different thing from completed agent purchases, so read the Adobe figures as a signal about discovery.
What changed in 2026
Early plans assumed that shoppers would buy inside the chat. On 24 March 2026, OpenAI shifted ChatGPT shopping toward product discovery and deprioritized Instant Checkout, saying the first version lacked the flexibility it wanted. Merchants keep their own checkout, and retailers integrated ACP mainly for product discovery.
Forrester reported that Walmart saw conversion about three times lower for products sold inside the chatbot than for products that sent shoppers to its site. The reading for merchants is practical: today agents mostly influence which store the shopper visits, and checkout inside AI surfaces is still maturing.
Google’s UCP, meanwhile, adds capabilities such as cart, catalog and identity linking for AI Mode and the Gemini app.
What it means for an online store
- Your catalog becomes a storefront. Titles, identifiers, prices, availability and images are what an agent compares. Weak data loses the match before a shopper sees you.
- Policies need to be data. Returns and shipping terms should exist in fields an agent can read, not only on a page.
- Checkout needs an API. An agent cannot click through a visual flow. Every step needs a callable endpoint.
- Payments need new credentials. Your payment provider must accept the scoped tokens agents use.
- Measurement changes. Add a source field so reports show which orders and visits came from which agent.
- Control remains with you. You decide which capabilities to support, and you remain the merchant of record.
Positive effect: shoppers who arrive through AI assistants engage and convert well in Adobe’s data. Risk: a store with stale data or a rigid checkout can be skipped, and the cost of that is invisible in standard analytics.
Where to start
- Audit your product data. Check required fields, identifiers, prices and availability for your top products.
- Check freshness. Make sure feed and site agree on price and stock.
- Review checkout and payments. Ask whether your platform and payment provider support agent flows.
- Pick the surfaces that matter. Decide whether ChatGPT, Google surfaces or both are your priority.
If you want to know where your store stands, our agentic commerce readiness audit checks it in one week for a fixed price.
FAQ
What is agentic commerce in simple terms?
It is online shopping done by an AI agent on your behalf. You describe what you want, and the agent finds, compares and buys it within limits you approve.
Is it the same as an AI shopping assistant?
Not quite. An assistant suggests products and answers questions. An agent also takes actions, such as building a cart and starting checkout, with your authorization.
Will shoppers still visit my website?
Yes. Today many AI-assisted journeys still end on the store’s own site, and merchants keep control of checkout. Agent-completed purchases are growing as protocols mature.
Who is responsible if something goes wrong with an agent order?
Under UCP and ACP the merchant remains the merchant of record, so order acceptance, fulfillment and support stay with you. Payment tokens carry limits set by the shopper’s authorization.
Conclusions
Do I need to choose a protocol now?
Not necessarily. Shopify and the large payment providers already support several protocols. Start with the shared groundwork: clean product data, live stock and a callable checkout.