Agentic Commerce
Agentic commerce is buying and selling in which AI agents search, compare, and pay for a buyer within set limits, using protocols like ACP, UCP, and AP2.
Agentic commerce is buying and selling in which AI agents carry out commercial tasks on behalf of a person or business: finding products, comparing offers, building a cart, and paying within limits the buyer sets. A shopper states a goal, such as "a carry-on suitcase under $200 that arrives by Friday," and the agent works through the searches, comparisons, and checkout steps a person would otherwise click through.
What separates this from earlier e-commerce automation is planning. A recommendation system ranks products a customer might like, while a commerce agent reads specifications, checks stock and delivery dates, asks a clarifying question when constraints conflict, and then acts. A large language model (LLM) typically interprets the request and decides each next step through function calling, and merchant and payment systems stay authoritative on prices, inventory, and whether a transaction goes through.
From Request to Checkout#
An agentic purchase moves through discovery, decision, authorization, and fulfillment. For discovery, the agent queries catalogs through merchant APIs and product feeds. Structured data such as schema.org Product markup gives it prices, availability, dimensions, and return terms in machine-readable form, which is why merchants preparing for agent traffic invest in complete, accurate product data. Where no such interface exists, computer-use agents can shop through the same web pages a person sees, more slowly and with more room for error.
In the decision step, the agent filters candidates against the buyer's constraints, totals the cost including shipping and tax, and assembles a cart through the merchant's checkout interface. Authorization is where agentic commerce differs most from a person at a keyboard. In protocol-based checkouts the merchant never receives the card number: the agent passes it a delegated payment token scoped to that merchant, a maximum amount, and an expiry, and the buyer either approves the final cart or pre-authorizes purchases within a budget. Card networks issue agent-specific tokens for the same purpose through Visa Intelligent Commerce and Mastercard Agent Pay. Once the order is placed, the merchant handles fulfillment, returns, and customer service as it would for any other sale.
The same flow runs between businesses, where procurement agents reorder supplies against contracts and budgets. In stores, shelf cameras running an Ultralytics YOLO26 object detection model trained on a retailer's own products can flag empty facings as they appear, and a replenishment agent can combine those detections with sales and warehouse data to draft supplier orders for a manager to approve. Retail teams can annotate shelf images and train that detector in Ultralytics Platform, then export it to run on in-store edge AI devices as part of broader retail inventory management.
Protocols for Agents and Merchants#
Agents need a common way to transact with thousands of merchants, and several open specifications now define one:
- Agentic Commerce Protocol: Developed by OpenAI and Stripe and released in September 2025 under the Apache 2.0 license, ACP standardizes how an agent creates and updates a checkout with a merchant and passes a payment credential without exposing card details. The business remains the merchant of record and controls which products agents can sell.
- Universal Commerce Protocol: Announced in January 2026 and co-developed by Google, Shopify, and retailers including Etsy, Target, Walmart, and Wayfair, UCP covers the lifecycle from catalog search and cart building through checkout, account linking, and order tracking, with support for A2A, MCP, and AP2 built in.
- Agent Payments Protocol: Introduced by Google in September 2025 as an open extension for A2A and MCP and since donated to the FIDO Alliance, AP2 handles proof of authorization. Cryptographically signed mandates, one for the checkout and one for the payment, record what the user approved, so merchants and payment networks can verify that an agent stayed within those limits.
These build on general agent plumbing: the Model Context Protocol (MCP) connects an agent to external tools and data, and the Agent2Agent (A2A) protocol lets a shopper's agent talk to a merchant's own agent. The commerce protocols add the carts, checkout states, and payment semantics on top.
Agentic Commerce vs. Related Terms#
- E-commerce personalization: Recommendation engines and semantic search help a person find products faster, and the person still decides and pays. In agentic commerce the agent completes the purchase.
- Conversational commerce: A shopping chatbot answers questions inside one retailer's site or messaging channel. A commerce agent works for the buyer across many merchants.
- Robotic process automation (RPA): Scripted bots repeat fixed purchasing steps on known systems. Commerce agents plan each purchase from a goal and adapt to whatever catalog they find.
Risks and Safeguards#
Giving software the power to spend money raises the cost of ordinary model errors. An agent can misread a compatibility requirement, act on a stale price, hallucinate a product detail, or place a duplicate order when a request is retried. Prompt injection is a specific threat: instructions hidden in a product listing or review can try to steer an agent toward a different seller or purchase.
The most effective AI guardrails for commerce sit outside the model. Payment credentials limited by merchant, amount, and time, spending caps, merchant allowlists, and confirmation steps above set thresholds bound what any single mistake can cost, whatever the agent was told. Idempotency keys, unique IDs attached to each payment request, let the payment processor recognize a retried request and charge only once, and complete transaction logs tie every order to the approval behind it for disputes and refunds.










