Agentic commerce is buying and selling in which an AI agent acts for the buyer, the seller or both: searching for products, comparing offers, placing orders, arranging payment and handling returns within limits a person has set. The customer states what they want and how far the agent may go; the agent carries out the purchase.
Why does agentic commerce matter?
Much of commerce assumes a person is at the keyboard. Fraud checks look for human patterns. Consumer protection assumes the buyer saw the price and the terms. Disputes turn on what the customer agreed to. Once an agent places the order, each of those assumptions needs a substitute: a record of what the person authorised, and proof the agent stayed inside it.
The seller side changes too. Merchants deploy agents to answer questions, adjust orders, process returns and check sellers or buyers against their own rules. Those agents act on customers' money and data, so the same questions about limits and evidence apply. How regulated businesses decide what agents may do is the subject of AI agents in banking.
How does an agentic purchase work in practice?
- Mandate: the person sets what may be bought, the budget, approved sellers and an expiry.
- Discovery: the agent searches and compares. Product pages and offers are claims from outside, useful for choosing, not for authorising.
- Order: the agent places it within the mandate, or holds it for the person if it falls outside.
- Payment: the payment step follows the payment provider's rules, including customer authentication where required.
- After the sale: tracking, returns and refunds, each with its own limits.
The step where the agent asks a person is the most important design decision. How an agentic operating system handles human approvals covers how that is built so the request is clear and the answer is recorded.
What does that look like in an example?
A small architecture practice lets a purchasing agent reorder office supplies from three approved suppliers, up to 500 a month. The agent notices toner is low and places an order within budget. A week later an email arrives, apparently from one supplier, asking that future payments go to a new bank account. The email is an external claim. It cannot change where money goes. The agent flags the request and holds any payment to that supplier until the office manager confirms the change by phone using the number already on file.
How is agentic commerce different from agentic payments?
Agentic payments are the money movement step: an agent initiating or managing a payment. Agentic commerce is the whole purchase around it: choosing, ordering, paying and dealing with what comes after. A payment can be agentic without any shopping, such as an agent paying approved supplier invoices. A purchase can be agentic while the payment step stays entirely with the person, who confirms it in their banking app.
Which related terms should you know?
- Agentic AI: AI that plans and acts rather than only answering.
- Mandate: the recorded authority a person gives an agent to act for them.
Compliance teams at online sellers face these questions directly; for a market view, see AI agents for compliance teams in e-commerce in Canada.
Questions people ask
Who is liable if a shopping agent buys the wrong thing?
That depends on the contract terms and the consumer law that applies. In practice, the party who can show what the person authorised, and that the agent stayed within it, is in the stronger position, so keep that record.
Do merchants need to change anything?
They need to recognise when an agent is acting, decide what agents may do on their site, and keep evidence of the mandate behind each agent-placed order. Returns and disputes will turn on that evidence.