Half of ordered items return, yet most can be offered for sale again

European fashion platform Zalando describes a shopping model in which customers order several sizes, try them on at home and send back what they do not keep. The company says that on average around 50% of ordered items are returned. It also says about 98% of returned fashion items can be offered again on the Zalando shop or Lounge by Zalando.

That 98% is not the share that was ultimately sold. Zalando adds that more than 96% arrive in ideal condition and can re-enter its offering immediately. These are published descriptions of existing operations, not improvements created by robots installed in October 2026. If a high proportion is already suitable to offer again, the next question is how quickly and at what full cost those items return to inventory.

A return is not simply outbound fulfillment in reverse

Zalando describes checking returned items against the original order, inspecting condition, deciding whether cleaning or repairs are needed and selecting a facility where the item is most likely to sell again. Its inventory placement incorporates recent demand and stock availability across locations.

Returned garments are not uniform industrial parts. They can arrive creased or loosely packed, with or without tags, in different states of wear. Fashion also has seasons, discount windows and value decay. Completing a return should therefore mean something more useful to the retailer than physically moving a garment: making inventory available for another sale.

Dual-arm robots entered live returns operations on October 7

On October 7, 2026, Zalando, CEVA Logistics and German robotics company Sereact announced live dual-arm robot deployments for fashion returns at CEVA locations in Greven, Germany, and Świebodzin, Poland. Sereact says its Cortex physical AI system can identify, grasp and sort unfamiliar products without per-item model training or a fixed SKU catalog.

These stations operate under a Robotics-as-a-Service arrangement. The companies say the robots take on repetitive handling while employees focus on supervision, exceptions and quality control. Live operation does not mean the entire workflow is unmanned. The public release does not establish how much condition grading, refurbishment or final resale eligibility assessment is actually performed without people.

Units picked per hour can hide the economic result

CEVA identifies the distinctive costs of returns: unpredictable flows, manual quality inspection, seasonal demand, lost value over time and high unit handling costs. On that basis, a high robotic picking rate is not sufficient evidence of economic improvement. Incorrect sorting and repeat inspection can shift costs elsewhere rather than remove them.

Banseog's proposed measurements start with elapsed time from receipt to saleable inventory, fully loaded cost per completed return, and the shares of exceptions, damage and rework. Offering an item again is also different from selling it to a customer. None of the partners has disclosed a robust before-and-after comparison for those business outcomes, so the deployment should not be described as already having proven higher margins or faster resale.

The October 22 announcement is about teaching the robot from failures

On October 22, Sereact, NVIDIA, Nebius, CEVA and Zalando announced a feedback loop for robotic returns. Video and robot movement data from live stations feed into a shared storage and compute environment. NVIDIA Isaac Lab and Cosmos generate variations in packaging, garment appearance and lighting for further Cortex training, after which new policies are scored on held-out cases before live testing.

The partners describe moving a retraining and validation cycle that previously took weeks toward updates within days. This is the suppliers' account of a system and its intended speed, not a published controlled measurement of Zalando's operating economics. The strategic question is whether new, unpredictable returns can be handled with less repeated engineering and less expensive human intervention over time.

The retailer, logistics operator and robotics supplier face different incentives

Zalando values saleable inventory and the performance of its commerce and partner infrastructure. CEVA runs contract logistics operations and must manage cost and service quality. Sereact supplies the robot intelligence and service layer, while Universal Robots, NVIDIA and Nebius provide other parts of the physical AI and computing stack.

Those interests lead to a practical contract question: if a robot sorts the wrong item and an employee must correct it, who bears the cost? The term Robotics-as-a-Service alone does not say who guarantees throughput, pays for exceptions or handles underperformance. The companies have not published the relevant contractual detail. These are decision questions, not allegations of a defective agreement.

The EU destruction ban adds context, not proof of causation

From July 19, 2026, large companies across the EU have generally been prohibited from destroying unsold apparel, clothing accessories and footwear. There are specified exemptions, including some unsafe or damaged items, and medium-sized companies become subject to the rules in 2030. EU guidance instead emphasizes resale, repair, donation and reuse.

This change increases the importance of decisions about unsold products, but it does not establish that the regulation caused Zalando to select Sereact. Zalando already operated resale, refurbishment and alternative outlet channels. The proper question is how companies can retain product value while controlling inspection, handling, storage and recovery costs under tighter circularity expectations.

An internal logistics capability can become a service sold to brands

Zalando is also an infrastructure supplier to other brands and retailers. It says more than 60% of items sold through its Partner Program use its returns infrastructure, while ZEOS offers fulfillment, shipping and returns services across multiple sales channels.

If robotics eventually lowers time-to-saleable inventory and fully loaded costs, that capability could improve the competitiveness of services offered to partner brands. But no evidence shows the new Sereact system is already available to all ZEOS customers or has delivered a quantified partner-level benefit. The potential business asset is the ability to complete difficult returns at predictable service levels, not simply ownership of robotic arms.

BANSEOG VIEW

BANSEOG VIEW | Measure completed work, not the robot's individual moves

Zalando already describes approximately 98% of returned fashion items as suitable to be offered again. That baseline must not be repackaged as an automation achievement. More informative metrics are the time and total cost of turning a returned item into saleable inventory.

The October 7 announcement establishes live robotic handling; the October 22 release describes the new real-world-data retraining loop. Neither release supplies independently verified before-and-after figures for resale completion time, cost per return or realized sell-through.

An enterprise buyer should measure physical AI at the unit of completed business work, including inspection, exceptions, damage and human supervision. Only once that performance is demonstrated can repeated deployments and B2B reuse be evaluated as an operating advantage.

SOURCES

Primary sources and references

Prepared October 23, 2026. Zalando's average ~50% item return rate, ~98% eligible for offering via resale channels, >96% ideal condition and >60% Partner Program returns infrastructure use are published company descriptions, not new operating achievements attributed to October robots. CEVA's Oct 7 release documents live Greven/Świebodzin deployment, Cortex identify/grasp/sort capability and RaaS alongside human quality roles. Oct 22 release describes the Sereact/NVIDIA/Nebius training architecture and weeks-to-days vendor claim; no independent measured ROI, reduced resale latency or cost savings are available. EU disposal ban from July 19 covers large businesses with exemptions and is not established as causing the investment.