Riskified agent identity scoring: who is behind the bot refund?

An AI shopping agent with a valid token tells you the agent may act for someone. It does not tell you if that someone is a good customer or a refund-fraud ring. On October 8, Riskified launched Agent Identity Risk Intelligence to answer the second question. It finds the AI personal assistant behind an order or a customer service request, connects it to a real consumer identity, and returns a risk signal in real time. Shopify merchants get it first.

What actually happened

Per Riskified's release on Business Wire:

• Named assistants: Meta's Muse, Instinct, and dots in ChatGPT.
• It works with agent protocols, not against them. Riskified supports the Personal Agent Protocol from Sierra and Meta. Its framing: the protocol proves the agent is authorized, and Riskified checks whether the person who authorized it can be trusted.
• Signals when telemetry is missing. Agent orders often have no normal device or browser data. Riskified uses identity history, network data on orders, claims and chargebacks, and the agent's behavior. For refunds, it can ask the assistant to confirm details or add a step-up check.
• Policy and reporting. Decision Studio gets rules by agent type, for example auto-approve low-risk orders from verified identities and send high-risk agent refund claims to review. Control Center shows what share of your orders and claims come through AI assistants.
• Zendesk. Human agents and Zendesk's AI agents see when a refund or return request comes from a consumer AI assistant, with the requester's risk.
• Availability. First for Shopify merchants, where Muse orders already arrive tagged as agent-originated. Limited enterprise beta in December 2026. General availability expected in 2027. It is also offered through Riskified's APIs and an MCP server on AWS Marketplace.

Co-founder Assaf Feldman: "A fraud ring can authorize an agent as easily as a loyal shopper can." The release gives no performance numbers, so treat effectiveness as unproven.

Why it matters for your business

Refund abuse is the easy target. Riskified cites the Artemis Refund Group, eleven people sentenced in April 2026 in U.S. federal court for refund fraud against online retailers. That was a scheme run by people. An agent can write a hundred polite, consistent refund claims and never get tired.

You do not need an enterprise beta to start:

• Tag agent orders now. If Shopify marks an order as agent-originated, keep that tag in your OMS and reports. You cannot set a policy for traffic you cannot see.
• Split refund rules by channel. An agent-filed refund on a high-value order should go to a person, not to auto-approve.
• Watch the ratio. Track refund and chargeback rates on agent orders separately from normal checkout. If the agent channel trends worse, you will know before Q4 ends.

Sources: Riskified via Business Wire (Yahoo Finance), Riskified investor relations.