AI in reverse logistics and returns: what actually gets automated (and what is marketing)

The real problem: a return costs you twice

A return is not a sale in reverse. It is a cost. The product travels back, someone receives it, inspects it, decides what to do with it and, meanwhile, that money is tied up. In Spain the problem is one of the biggest in Europe: according to the 2025 Annual Returns Benchmark Report by ZigZag and Retail Economics, 23.7% of non-food online orders get returned, and the returned volume will approach 13.3 billion euros in 2025.

The picture gets worse exactly where online sales are highest. Clothing gets returned at 31%, footwear at 27% and electronics at 23%. This is not a one-off bad quarter: European B2C e-commerce grew 7% in 2024, going from 784 billion to 842 billion euros (EuroCommerce). The return flow grows right alongside sales.

A note to clear the smoke. The 890 billion dollar figure you will see repeated on LinkedIn is for the United States (NRF and Happy Returns, 2024), not Spain or Europe. It is useful to size the global problem, not your warehouse. What affects you is your own return rate and what it costs you to process each return, the same way it pays to work out what it really costs to process an invoice by hand before deciding where to put technology.

• Return shipping and reverse logistics.
• Staff time to receive and inspect each item.
• Reconditioning, repackaging and relabeling.
• Loss of value: what it is worth less for having left the warehouse.
• Refund paid to the client before the product is resold.

What AI actually automates today (and what it does not)

Reverse logistics has a chain of decisions. You receive a package, identify what it is and what condition it arrives in, decide its destination (resell as new, liquidate, repair, recycle), price it and notify someone. AI is good at the parts of that chain that are classification, prediction and text. It is bad, or flatly useless, at the parts that are physical handling.

This table separates what is real from what is aspirational. The column that matters most is the last one: what stays human is not a failure of the technology, it is its honest limit as of 18 August 2026.

1. Classification and triage of what comes back

This is the most mature application and the one that pays back fastest. When a package comes back, someone has to open it, look at the product, compare it with the delivery note, decide whether it is fit for resale and enter it into the system. That triage usually depends on one or two experienced people, and it becomes a bottleneck on Mondays after a weekend of sales.

A vision model looks at the photo of the item and its label, and proposes a category, a reference and a first assessment of condition (new and sealed, opened but unused, showing signs of use, damaged). It does not replace the final judgment, but it removes the work of identifying and typing. The person goes from looking at 100% to reviewing only what the model flags as doubtful.

What you need and what you do not

You do not need a vision cell costing tens of thousands of euros to get started. Part of the triage is solved just by reading the return reason the client writes and the documents that accompany the return. There, the technology is the same one we use for AI invoice OCR: extracting data from a text or an image and placing it in your system without anyone typing it.

At a logistics operator we work with, triaging every return depended on two people with years of experience. Once the first classification was handed to a model that read the photo and the delivery note, those two people stopped identifying every item and started validating only the cases the system was not sure about. The know-how stayed theirs; what disappeared was the repetitive typing.

2. Dynamic liquidation pricing and buyer matching

A returned product that does not go back into the main sales channel ends up in liquidation. There are two decisions AI handles well here: how much you sell it for and to whom. Setting the clearance price by eye is slow and usually leaves money on the table, on one side or the other. A model cross-references the item's condition, the accumulated stock for that reference and recent demand, and proposes a price. The person sets the floor (we do not sell below this) and approves.

Matching is the least known step and the most interesting for anyone moving volume. Instead of listing a lot and waiting, the system cross-references what you have to liquidate against the history of who buys that type of stock: which categories, what lot sizes, at what price they usually close. It turns a list of leftovers into offers targeted at the B2B buyers who actually want them.

All of this is, at bottom, sales automation applied to reverse logistics. The same logic that organizes a sales funnel serves to organize the exit of returned stock. If you already have a CRM, it fits with what we cover in sales automation and CRM: getting client and buyer data to work on its own instead of living in a spreadsheet.

3. Communication with B2B buyers and with the end client

In reverse logistics there are two conversations that repeat themselves thousands of times. With the end client: where is my refund, when does the label arrive, why did you deduct money from me. With the liquidation buyer: what do you have available, in what condition, at what price, when can you ship it. Both are perfect to automate because 80% of the messages are variations on the same questions.

An assistant connected to your system answers the status of a return or a liquidation order instantly, without anyone on the team interrupting what they are doing. When the query goes off script (a complaint, a serious negotiation), it hands it to a person with the context already gathered. For the channel, a WhatsApp chatbot for companies is the most natural fit in Spain, because it is where the client and the small buyer already write.

The goal is not for nobody to talk to anybody. It is for your team not to spend the morning copying tracking numbers. The real commercial relationship, the one that closes a recurring liquidation deal, is still handled by a person.

When it does NOT pay to automate your reverse logistics

This is not for everyone, and promising it would be exactly the smoke you are tired of. There are clear cases where building AI into reverse logistics is spending money just to show off a project.

The rule is simple: if you do not have volume or repetition, there is nothing to automate. AI pays off when the same decision gets made hundreds of times. If you get twenty returns a month and they are all different, a person with a good template is faster and costs you less.

• Few returns a day and highly varied: there is no pattern to learn, the cost does not add up.
• Unique or high-value product where every return gets assessed by hand anyway.
• Dirty data: if your inventory and your return reasons are not at least minimally organized, that gets cleaned up first.
• No defined process: if nobody today knows who decides a return's destination, automating the chaos only speeds it up.
• Expecting AI to recondition items: cleaning, repairing or repackaging is physical work and will stay that way.

How to start without building a six-month project

The usual mistake is wanting to automate the whole chain at once. What works is picking the most expensive bottleneck and tackling it first, measuring, and then moving on. In most operations that bottleneck is triage or repetitive communication, not pricing.

An honest pilot gets built in weeks, not quarters, and it runs on a bounded scope: one product family, one return reason, one message channel. It shows you the real number for your operation before you invest more. If you move logistics volume, the useful starting point is this list of processes a logistics operator can already automate, to place reverse logistics within the rest.

We roll it out measuring before and after, with no promises of a magic percentage. If you want a read on your specific case, that is what our AI for logistics work is for, and a first conversation of AI consulting in Barcelona to decide whether your volume justifies taking the step.

Frequently asked questions

Sources

• Informe Benchmark Anual de Devoluciones España 2025 (ZigZag Global y Retail Economics)
• Las devoluciones online en España superan los 13.300 millones en 2025 (Marketing4eCommerce)
• European E-commerce Report 2025 (EuroCommerce y Ecommerce Europe)
• NRF and Happy Returns Report: 2024 Retail Returns to Total $890 Billion (National Retail Federation)

Keep reading