Riskified launches AI agent identity checks for merchants

Riskified has launched Agent Identity Risk Intelligence for merchants dealing with AI personal assistants in online shopping and customer service. The product is designed to identify the person behind an AI agent and return a real-time risk signal.

The launch comes as consumer AI assistants begin placing orders, filing claims, seeking refunds and handling returns on behalf of shoppers. That creates a new challenge for retailers, which have traditionally relied on device, browser and account signals to detect fraud.

The new offering extends Riskified's existing identity risk tools to interactions initiated by personal AI agents such as Meta's Muse, Instinct and similar services. It is designed to recognise when an order or service request comes from an AI assistant, link that request to the underlying consumer identity and assess whether the activity should be trusted.

That reflects a shift in how merchants may need to view automated activity online. In the past, signals such as a single-use card, relay email address or lack of device history could indicate suspicious behaviour. With AI agents acting for legitimate customers, those same signals may no longer be enough to separate fraud from authorised delegation.

Riskified's UK research points to early consumer interest in using such tools for post-purchase disputes. It said 37.9% of British shoppers would use an AI agent to draft a refund or credit complaint, while 23.4% would use one to initiate a bank chargeback if a return proved difficult.

New risk model

For merchants, the issue is moving beyond whether an interaction is automated to whether the automation has been authorised by a legitimate and trustworthy person. Riskified argued that newer standards for agent verification may confirm that an assistant has permission to act, but do not in themselves address whether the underlying account holder is genuine or whether the request is part of a fraud scheme.

That distinction matters because an authorised agent could still be linked to account takeover, fake accounts or organised refund abuse. The system is intended to sit alongside emerging agent protocols by focusing on the identity of the person who granted the authority.

"Agent protocols solve a real problem: merchants need a secure way to let a customer's assistant act for them. What they don't solve is whether that customer should be trusted. A fraud ring can authorize an agent as easily as a loyal shopper can," said Assaf Feldman, Chief Strategy Officer, Technology and Co-Founder at Riskified.

"Riskified sits on top of those protocols and answers the harder question, using what our identity engine already knows about the person behind the agent, so merchants can welcome a good customer's assistant, decline a bad actor's, and notice the moment a real customer's agent starts doing something that the customer never would," Feldman said.

Checkout to claims

The product covers both checkout and post-purchase interactions. When standard device and browser telemetry is missing, as can happen when an AI assistant acts for a shopper, the system relies more heavily on identity records, network history and the behaviour of the agent itself.

For refund and return interactions, it can engage directly with the assistant to confirm details or apply additional checks. The approach is intended to distinguish between a legitimate assistant acting for a real customer and a scripted process used by fraudsters.

The first deployment will focus on merchants using Shopify, where orders placed through Meta's Muse already arrive marked as agent-originated, according to Riskified. Coverage will expand to other commerce platforms as standards for identifying agents become more widely adopted.

Alongside the core launch, Riskified will support policy controls by agent type in its Decision Studio product. That would allow merchants, for example, to approve lower-risk agent orders tied to verified identities while sending higher-risk refund claims for review.

The same risk category will also feed into the company's Zendesk integration, giving both human support staff and AI service agents visibility when a refund or return request comes through a consumer AI assistant.

Fraud concerns

The broader concern for retailers is that tools designed to help consumers manage shopping and after-sales service could also make organised abuse easier to scale. Riskified pointed to refund fraud groups that already use scripted claims involving missing items or empty boxes against large retailers.

It argued that persistent AI agents could increase the volume and frequency of such activity by allowing bad actors to automate repetitive tasks across orders and support channels. In that environment, merchants may need fraud systems that assess both the source of the automated request and the identity behind it.

Reporting on agent-originated traffic will be added to Riskified's Control Centre, including approval rates, fraud rates and claim abuse rates, with breakdowns by assistant where that information is available. The company said the data would help merchants compare the performance of agent-driven activity with the rest of their business.

Agent Identity Risk Intelligence will be available through Riskified's existing APIs and its AI Agent Approve MCP server on AWS Marketplace.