Multiple companies join forces to insure "AI agent shopping": How can merchants reduce risks sta

As more and more consumers use AI to find products, compare prices, see discounts, and even place orders, e-commerce merchants not only see growth opportunities, but also face new risks head-on. To this end, Riskified, a global e-commerce anti fraud and refund protection company, has announced a partnership with Human Security to launch a unified risk control framework: on the one hand, it identifies and manages the access and interaction of AI shopping agents, and on the other hand, it provides stronger protection against fraud, refunds, and policy abuse, helping merchants to "enjoy the benefits of AI" without being dragged down by "AI abuse".

John Searby, Chief Strategy Officer of Human Security, stated that they provide a "layer of trust and visibility" that allows merchants to identify AI agents and develop "trust/distrust" strategies for them.

Riskified co-founder and CTO Assaf Feldman emphasized that in the era of "AI proxy trading", identity and trust are more complex; Through new proxy tools and rules, try to approve as many real orders as possible, reduce false rejections, and maintain profits.

Why do 'AI agents' make it harder for merchants?
Rule based risk control is easily bypassed. AI agents will learn thresholds (such as access speed, coupon restrictions, IP segments, etc.) and quickly adjust their behavior to avoid static rules.

Highly anthropomorphic. Agents are adept at mimicking human operation paths, rendering traditional "robot recognition" ineffective.

The risk of LLM recommending traffic is relatively high. Early samples of Riskified show that the LLM recommendation traffic risk of a top ticketing merchant is 2.3 times higher than that of Google search; A certain electronic product merchant is about 1.8 times.

Automatic 'dealer arbitrage'. Agents can empty inventory during new listings or promotions, and then resell at high prices through grey channels, disrupting pricing and damaging reputation.

The promotion period is particularly fragile. A more intelligent agent can decode verification codes, read site structures, research promotion mechanisms, and dynamically adjust, disguised as a "normal buyer" to wash away profits during peak hours.

Simply put, AI brings efficiency but also leads to "super efficient abuse". If risk control is not upgraded, businesses may face a triple blow of revenue loss, inventory manipulation, and damaged brand trust.

What problem does the joint framework aim to solve?
The goal of this framework is not to "block AI one size fits all", but to distinguish between "good AI" and "bad AI" - to enable compliant agents to place orders smoothly, and malicious/abnormal agents to be identified and intercepted, thereby achieving:

Less false rejections (protecting genuine customer conversions)

Stronger abuse interception (blocking arbitrage/coupon fraud/subsidy manipulation, etc.)

Clearer visibility (knowing who, where, and what has been done)

More stable profit margin (finding a balance between "speed" and "risk control")

8 'landing actions' that merchants can take now
Layered defense, not relying on a single rule. Device fingerprint+behavior analysis+real-time anomaly detection+machine learning model+manual review as a backup.

Give a 'trusted agent' the right path. Open controlled APIs or data summaries (price/inventory/timeliness) to verified agents/partner platforms, and downgrade or sandboxe unverified sources.

Dynamic speed/purchase restrictions and session integrity. Adjust frequency limits, queue purchases, and secondary verification based on user reputation and behavior for rush buying/promotions/new listings.

Add "anti brush armor" to the discount tool. Set threshold for coupon/discount code, whitelist, device/account binding, risk control trigger deactivation, and monitor abnormal write off paths.

Traffic "physical examination". Label LLM/proxy sources, clean up synthesized sessions/abnormal clicks, avoid contaminating advertising attribution and increasing customer acquisition costs.

The "anti arbitrage" strategy of inventory and price. Set dynamic quotas for popular products, release inventory on a time-sharing basis, and diversify across warehouses to reduce the risk of being instantly cleared.

High risk action reinforcement verification. Large orders, abnormal address/payment method changes, frequent returns and exchanges trigger KYC/two factor verification/delayed performance+manual verification.

SOP and Event Response. Pre set emergency procedures for scenarios such as "abnormal peak/promotion being brushed/inventory being cleared", and quickly collaborate with platform/risk control partners.

Core idea: "Don't block AI across the board", but make "good agents smoother and bad agents harder".

Special Reminder for Cross border Sellers
Cross border customs clearance process: Abnormal orders may lead to uncertainty in tax declaration and compliance review; Suggest collaborating with risk control/logistics partners to handle risks in a tiered manner (such as delayed performance and review of high-risk orders).

Warehouse allocation and capacity flexibility: During promotional/hot periods, AI agents will amplify fluctuations, so it is necessary to prepare flexible solutions for warehousing, transportation, and multiple express channels to avoid SLA collapse.

Customer trust: In addition to necessary security checks, maintain process transparency and explanation friendliness - while blocking bad orders, do not force real customers to leave.