Japanese pilot uses AI to cut packaging-related returns
Examples of damaged cardboard packaging resulting in products being returned during delivery (image:Fujitsu)
Damaged packaging creates costly complications for supply chains in Japan and around the world. A crushed corner or dented box can trigger inspections, rejected deliveries and replacement journeys, even when the product inside remains intact. The challenge is deciding which damage warrants further investigation and which is merely cosmetic.
In Japan, Fujitsu’s Kozuchi AI platform will supply image-recognition technology for a pilot designed to make that judgement more consistent. Fujitsu and SoftBank Robotics are working with the Home Appliance Supply Chain Council and the Distribution Economics Institute of Japan, combining their AI tools with assessment criteria developed across the industry.
The pilot began on 2 November 2026 and runs until 15 January 2027. It covers deliveries from manufacturers to mass retailers and will test the accuracy and practical usefulness of AI-assisted packaging assessments in logistics operations.
Why damaged packaging leads to unnecessary returns
The immediate issue is inconsistent judgement. According to the partners, companies apply different acceptance criteria, while decisions also depend on staff experience. A scuff or dent can therefore receive different treatment depending on who inspects it and where the delivery arrives.
The consequences extend beyond moving a box back to its sender. Inspection, storage, claims administration and replacement deliveries all require resources, as monitoring supplier SpotSee explains in its white paper on supply-chain damage. Avoiding an unnecessary return can remove several tasks from an already busy logistics operation.
In Japan, those demands fall on an industry facing workforce pressures. A 2026 transport ministry presentation describes relatively few younger trucking workers, many aged 50 or above and more than two vacancies per jobseeker. Further retirements threaten to deepen recruitment difficulties.
Since April 2024, truck drivers have also been subject to an annual overtime ceiling of 960 hours, introduced to address excessive working hours. Government guidance warned that transport capacity could fall around 34% short of requirements by 2030 without efficiency measures. The pilot announcement cites labour shortages and rising delivery costs as its background; the workforce profile and working-hour reforms explain the wider pressure to use driver time more efficiently.
Damage-related returns also concern international markets. DHL’s 2026 E-Commerce Trends Report found that goods lost or damaged in transit were among the main return reasons cited by 39% of shoppers and 42% of businesses. Its surveys covered 29,000 shoppers across 29 countries and 5,800 businesses across 28.
The packaging itself can be a misleading guide. In Stord’s US mystery-shopping study of more than 250 transactions, 6% of orders arrived with visibly damaged packaging, but most of those products were undamaged. Some damaged products arrived in intact packaging. Although the study covered clothing, beauty, nutrition and food products rather than appliances, it illustrates the limits of judging contents by the condition of their packaging.
\
How Fujitsu Kozuchi and SoftBank Robotics will assess packaging damage
The two industry organisations are conducting the project through their Council for Sustainable Home Appliance Logistics. They will plan the trial, provide packaging images and operational expertise, and develop common criteria from rules already used by manufacturers, retailers and logistics providers. The project has been selected under a transport ministry subsidy programme supporting changes in logistics practices through digital technology.
Fujitsu Kozuchi provides access to AI technologies and software components that companies can test and apply to specific business tasks. In this pilot, the relevant technology is Amalgamation AI for Vision, which the announcement identifies as Kozuchi’s image-recognition technology. Kozuchi supplies part of the technical foundation for the trial; the project also relies on SoftBank Robotics’ solution and the industry organisations’ data and assessment rules.
Amalgamation AI for Vision is designed to assemble a workflow of vision tools from natural-language instructions and a small number of example images. Fujitsu says users can validate performance on real images, adjust sensitivity and correct false alarms. These capabilities fit the pilot’s need to turn agreed packaging criteria into a visual assessment and refine it during testing.
SoftBank Robotics contributes a solution that analyses smartphone photographs of cardboard packaging against predefined criteria and records the results. Photograph-based assessment offers a way to capture evidence at the delivery point, while the stored findings allow decisions to be reviewed later. The announcement provides less technical detail about this solution and does not specify how the two companies’ systems will be integrated.
The partners will evaluate accuracy, consistency of the criteria, operational feasibility and effectiveness in reducing returns and redeliveries. A photograph remains evidence about the visible packaging, rather than a complete inspection of the appliance inside. No quantified return-reduction target or measured environmental saving has yet been disclosed.
How AI and shipment monitoring address damage worldwide
Elsewhere, companies are applying image analysis at other stages of the journey. Amazon’s Project P.I. (Private Investigator), announced in 2024, uses imaging tunnels to identify imperfect products before dispatch. Employees review flagged items and decide their next destination, helping prevent defective goods from reaching customers.
Amazon’s researchers have also described the challenge of distinguishing tolerable packaging damage from significant product damage. A threshold that flags damage too readily can divert undamaged goods for inspection, increasing workload. AI therefore needs to support the appropriate judgement, as well as recognise visual abnormalities.
Transport monitoring supplies another kind of evidence. SpotSee’s devices record impacts, tilt and temperature excursions, helping companies investigate handling incidents and decide where further inspection is warranted. Such records describe conditions during the journey; photographs document visible condition at a particular point.
Alongside these tools, the International Safe Transit Association’s packaging-test guidance calls for stakeholders to agree acceptable product damage and packaging degradation. Although this concerns testing rather than AI delivery inspection, it reinforces the importance of an agreed basis for decisions.
Taken together, these examples suggest a practical role for AI: making visual checks more repeatable, directing attention to cases requiring investigation and preserving evidence for review. The value depends on connecting those assessments to sound acceptance rules.
The Fujitsu and SoftBank Robotics pilot will test that connection in appliance logistics. The partners intend to publish the criteria and explore use beyond the sector, potentially giving other industries a clearer basis for accepting deliveries with confidence and avoiding unnecessary return journeys.
Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.