What fraud controls should retailers add for AI shopping agents?
Treat an agent purchase as delegated. Confirm which agent is acting and for which account. Keep purchase limits, payment authentication and velocity checks at the level you apply to people. Record the agent's identity with the order, and watch for automation that impersonates a known agent.
A delegated purchase has three parties: the customer, the agent acting for them and the merchant. Fraud controls built for two parties assume that the entity at the keyboard is the customer. With an agent, that assumption needs to be replaced by two questions. Is this the agent it claims to be, and did the customer authorise it to do this?
Start with the agent's identity. Verify it through the attacker's published network ranges or request signatures, as you would for any declared crawler. An unverified request that claims an agent's name should receive no special treatment, since the name is what an impersonator would borrow to pass through relaxed rules.
Then confirm the link to the customer. The purchase should be tied to a customer account or a payment credential the customer has approved for that use. Where the agent logs in as the customer, treat a new device or location as you would for any sign-in. Where the agent presents a delegated payment credential, check that it is valid for your store and for the amount.
Keep the existing controls in force. Purchase limits apply to the customer behind the agent and are counted across device, card and address as before. Payment authentication rules still apply to the transaction. Velocity checks matter more, because an agent can place orders faster than a person, and a compromised or misused agent could place many.
Record what happened. Store the agent's identity, how it was verified and which account it acted for alongside the order. When a customer later disputes a charge or says they did not ask for the purchase, this record is your evidence of who placed it and under what authority.
Think about returns and disputes in advance. An agent may buy the wrong item or misread a condition. Decide how your returns policy treats agent purchases and say so.
Watch for promotion and stock abuse at scale. An agent that can register, apply a code and check out is also a tool for claiming a welcome offer many times or taking limited stock.
Introduce agent checkout in monitoring first. Observe which agents arrive, what they do and how their orders perform before adjusting any rule for them.
How Kairal helps
Agent Traffic Intelligence, currently in early access, checks agent identity and purpose to support your access policies. A recognised agent is not proof of a customer’s authorisation; keep account, purchase and payment checks on the operations it can request.
Agent Traffic Intelligence