Agentic AI in retail

Shopping where an AI agent does part of the work for the customer: it searches, compares, builds a cart and, with permission, pays. The agent can be the retailer's own assistant or a third-party tool such as ChatGPT or Gemini. McKinsey found that in an average shopping category, 17% of US AI users had bought directly through an AI tool by mid-2026.

A chatbot answers questions. An agent takes actions in your systems: it changes a delivery address in the OMS, starts a return, drafts a purchase order or completes a checkout, within the permissions and limits you set. That is why agents need integration and guardrails, not just a knowledge base.

Order status and address changes, return requests and triage, product discovery and cart building, merchandising reports, markdown proposals and replenishment drafts. Start with tasks that are frequent, rule-based and easy to measure, and keep a person approving anything that moves money above a set limit.

Through the same APIs your other systems use. The agent gets a defined set of tools, such as read order, update address or create return, each with its own permissions. We build one integration layer so several agents reuse the same connections, logging and access controls.

Make product data complete and machine-readable, keep availability and prices accurate in real time, and expose catalog, cart and checkout through the protocols your platforms support, such as ACP, UCP or AP2. Decide which agents you allow, verify them, for example with Visa's Trusted Agent Protocol, and expect uneven support: some AI platforms have scaled back instant checkout.

Give each agent only the actions it needs, cap spending, refunds and discounts, and require confirmation from the shopper or a colleague above those limits. Treat product content, reviews and messages as untrusted input that may try to redirect the agent, keep personal data out of prompts unless needed, verify external agent traffic and log every action for audit.

Most retailers will need both. McKinsey found that 56% of consumers used an AI chat or search tool in the past three months against 27% for a retailer's assistant, but at checkout retailer assistants were slightly preferred, 15% against 14%. Open your catalog to external agents for discovery, and use your own assistant where live inventory, loyalty and service matter.

We usually start with a 4-week proof of concept on one workflow and your real data, which typically costs $15K. A production agent integrated with your commerce platform, OMS and helpdesk starts from $60K, depending on the number of integrations, data preparation and security requirements.