Inspect Returns with AI: Document Condition and Compensation

Once Black Week and Cyber Monday are over, returns pile up at goods-in, and every parcel raises the same question: does the item go back into stock unchanged, into sale as B grade, or back to the customer with a justified claim for compensation? If you inspect returns with AI, this becomes a fixed process: photos at the inspection desk, a model that detects marks and missing accessories, a suggested grade and a draft justification. A staff member makes the decision. This article shows where the law draws the line for compensation for diminished value, why the burden of evidence lies with the retailer and how the pre-check connects to the shop and merchandise management. Technically, this is a case for AI automation.

Why goods-in becomes a bottleneck after sales events

Returns are part of online retail. On average, online shoppers in Germany send back 11 percent of their online purchases (Bitkom); in fall 2023, 1,050 online shoppers aged 16 and over were surveyed. In an earlier survey, retailers themselves estimated the share of returned goods at 15 percent (Bitkom 2021), and 44 percent of online retailers said that returns were a burden on their business (Bitkom 2021). After sales events with many orders, the parcels come back in bulk. The inspection desk in the warehouse then becomes a bottleneck, because every return has to be opened, examined, graded and booked before the refund can go out. How to clear the backlog on the order side is described in our article on the order backlog after sales events.

A second result of the same survey matters for the inspection: 56 percent cited a faulty or damaged product as a reason for returns (Bitkom). So not every mark noticed at the inspection desk is caused by how the customer handled the item. The inspection has to distinguish between defects that already existed at dispatch, damage on the way back and marks that occurred at the customer.

At the same time, the clock is running. After a withdrawal, the goods and payments received must be returned no later than after 14 days (BGB), and for the retailer the period starts when the notice of withdrawal is received. In a consumer sale, the retailer may withhold the refund until the goods are back or the customer proves that they were sent (BGB). Customers pay close attention to this: for 94 percent of online shoppers, a quick refund of the purchase price is important (Bitkom). A thorough inspection must therefore not turn into a queue.

What the law means by compensation for diminished value

Compensation for returned goods is governed by Section 357a(1) of the German Civil Code (BGB). The consumer owes it when two conditions are met together: the loss in value results from handling the goods in a way that was not necessary to examine their nature, characteristics and functioning, and the retailer informed the consumer about the right of withdrawal as required by the Introductory Act to the Civil Code (BGB). If either condition is missing, there is no compensation. The European basis puts it even more clearly: the consumer is in any event not liable for diminished value if the trader failed to provide notice of the right of withdrawal (Directive 2011/83/EU).

What "examination" means is explained in recital 47 of the directive with a comparison that suits any inspection desk: the consumer should only handle and inspect the goods in the same manner as he would be allowed to do in a shop (Directive 2011/83/EU).

For example, the consumer should only try on a garment and should not be allowed to wear it.

Directive 2011/83/EU on consumer rights, recital 47

The European Commission has made this shop standard more concrete in its guidance on the directive. According to the guidance, diminished value can consist in particular of cleaning and repair costs and, if the goods can no longer be sold as new, of an objectively justified loss of income when the trader disposes of them as second-hand goods (European Commission). Whether a test went beyond what was necessary has to be assessed case by case in the event of a dispute. The guidance gives examples that translate directly into inspection rules:

What the retailer has to prove

If you claim compensation, you need your own evidence. In the Messner case (C-489/07), the Court of Justice of the European Union ruled that a national rule placing on the consumer the burden of proving that the goods were not used beyond what was necessary would adversely affect the effectiveness of the right of withdrawal; the Commission refers to this in its guidance (European Commission). In practice, this means that the retailer generally has to show the loss in value and its cause. A photo taken on arrival is often not enough. It shows a scratch, but not when it occurred.

There is another rule that many inspection desks overlook: in the event of withdrawal, the trader bears the risk of the goods being sent back (BGB). Damage that occurred on the way back is not a case for compensation. It is a matter for the carrier. That is why the condition of the parcel on arrival belongs in the documentation before it is opened.

Whether the retailer deducts a justified claim from the refund or claims it separately is not regulated by the directive, but by national contract and procedural law (European Commission). Either way, the inspection process has to produce reliable records. A complete inspection log usually contains:

• Return number, order and time of arrival, linked to the acknowledgment of receipt of the withdrawal
• Photos of the parcel before opening, so that damage on the way back remains visible
• Photo series of the item from fixed angles under fixed lighting, plus detail photos of each finding
• Dispatch photos for high-value items, so that the condition before shipping can be compared
• The version of the withdrawal notice the customer received when ordering
• Finding, grade, justification and name of the person who approved

How an AI-assisted pre-check works

It all starts with an inspection desk that photographs each return the same way: a camera on a fixed arm, even lighting, a mat with a scale. The staff member scans the return label, and the system loads the order, the item data and the product images from the catalog. The comparison is what matters: the model sees not only the return photo, but also what the item looks like new and what the item master lists as the scope of delivery.

Each finding receives a confidence value. If it is below a threshold that you define, the system flags the case for a close visual check. If it is above, the staff member sees the suggestion and confirms it with one click. How such thresholds and approvals are set up is described on the page on automated decision-making. The principle is the same as with AI order entry: the machine prepares, a person approves.

A, B or C grade: grading with clear rules

Grading answers an operational question: how can the item be sold again? Each product group needs its own rules, which the staff member applies in the same way as the model. A common split looks like this:

A second rule from the guidance applies to resale: the trader must not mislead a new buyer as to whether refurbished goods are new or used (European Commission). The grade from the inspection log should therefore be transferred directly into the item condition and product text in the shop instead of being entered by hand.

Grading also has an environmental side. Since July 19, 2026, large companies have been prohibited from destroying unsold apparel, clothing accessories and footwear (Ecodesign Regulation). Unsold explicitly includes products that a consumer returned under the right of withdrawal (Ecodesign Regulation). For medium-sized companies, the ban applies from July 19, 2030 (Ecodesign Regulation). In addition, the German Circular Economy Act requires distribution to keep products usable (KrWG). Clean separation into A, B and C grade opens a second sales channel for more items.

Justify compensation instead of deducting a flat amount

If a finding may justify compensation, the staff member checks three points before an amount is set. First, the notice: did the customer receive a correct withdrawal notice when ordering? If it is missing or does not meet the legal requirements, there is no compensation, and the withdrawal period can be extended to up to twelve months and 14 days (BGB). Second, the cause: does the mark go beyond what would be allowed in a shop? Third, the loss in value: what cleaning or repair costs arise, or what loss of income occurs when it is sold as used?

The model drafts the text for the customer from the inspection log. It names the finding, refers to the photo and explains the standard in plain words. The staff member sets the amount. The German model withdrawal notice phrases the standard in a way customers can know: they only have to pay for any loss in value if it results from handling that was not necessary for examining the goods (EGBGB). A justification that picks up this sentence and backs it with a photo is understandable for the customer.

{
  "return_id": "RT-104233",
  "order_id": "SW-58817",
  "received_at": "2026-12-02T09:14:00+01:00",
  "parcel_photo": "parcel_arrival.jpg",
  "withdrawal_notice_version": "Withdrawal notice, version June 2026",
  "photos": ["front.jpg", "back.jpg", "display_detail.jpg", "contents.jpg"],
  "findings": [
    {"id": 1, "type": "scratch", "location": "display, top right", "confidence": "high"},
    {"id": 2, "type": "missing_accessory", "item": "power adapter", "confidence": "high"}
  ],
  "grade_suggestion": "B",
  "compensation_draft": "Scratch on the display, not visible in the dispatch photo. Standard: handling beyond examination.",
  "decision": {"grade": "B", "compensation": "Ask about the adapter, review the scratch", "assigned_to": "returns-team", "approved_by": null, "approved_at": null}
}

People decide, the model prepares

A reduced refund affects the customer directly. The GDPR gives every data subject the right not to be subject to a decision based solely on automated processing which produces legal effects concerning him or her or similarly significantly affects him or her (GDPR). Whether an individual deduction falls under this can be left open if the inspection process is designed from the outset so that a staff member decides and the model only suggests.

This very uncertainty keeps many companies away from AI. Slightly more than a quarter of companies in Germany with 10 or more employees (26 percent) used artificial intelligence in 2025 (Destatis). Companies that had considered using it but had not done so cited, among other reasons, uncertainty about the legal consequences (62 percent) and data protection concerns (60 percent) (Destatis). A pre-check with clear approval by people addresses both points better than a system that deducts amounts on its own.

• No automatic deduction: the model suggests, a staff member sets the grade and the amount.
• Make thresholds visible: uncertain findings go to a visual check instead of being waved through silently.
• Log deviations: if the staff member deviates from the suggestion, the reason is stored. This shows where the model needs retraining.
• Take samples: approved cases without findings are also checked again regularly.
• Involve customers: if you claim compensation, send the photo and justification along and name a contact for questions.

Photos, data and operation in-house

Return photos show more than the goods. Address labels are stuck to the parcel, notes lie in the box, and returned devices may still hold the customer’s accounts, photos or messages. An inspection desk should cover labels when taking photos or blur them automatically, and record devices only after their data has been deleted if the screen is part of the inspection. Under the GDPR, personal data may only be kept in a form that permits identification for as long as necessary for the purpose (GDPR).

A simple distinction helps with retaining the photos. Returns without findings usually no longer need photos once the refund is complete. For cases with compensation, the period follows the possible clarification: the regular limitation period is three years (BGB) and starts at the end of the year in which the claim arose and the creditor became aware of it or would have become aware of it but for gross negligence (BGB). The specific period is set by your data protection concept.

Because return photos can contain personal data, it makes sense to run the model in-house rather than with a third-party service. Image recognition for a clearly defined purpose can run on your own hardware or in a hosted environment without images going to third-party AI services. Our article on AI in the shop without data transfer describes the operating models. On request, we take over operation through our hosting.

Connecting shop, merchandise management and refund

A pre-check only pays off when its result flows on without retyping. The process starts with the withdrawal: via the electronic withdrawal function, the customer receives an acknowledgment of receipt with date and time (BGB), and this acknowledgment belongs to the return. How the function is implemented in the shop is described in our article on the withdrawal button in Shopware. This is followed by the steps at the inspection desk and in merchandise management:

1. Register the return, create the label, link withdrawal and order
2. Scan the parcel at goods-in and photograph it before opening
3. Photo series at the inspection desk, analysis by the model, suggested grade and justification
4. Approval by the staff member, storage of the inspection log
5. Stock booking by grade: new goods, B grade item or repair stock
6. Trigger the refund or credit note in the shop and in accounting, with a justification to the customer where compensation applies

Booking, credit note and return number live in the shop and merchandise management; the inspection process only supplies the decision. The connection is made via the existing integrations, for example with JTL-Wawi or another ERP. In Shopware 6, the connection runs through the Admin API, which can set stock levels and order status, for example, so the inspection desk does not need its own interface in the shop. How to make return processes leaner overall is covered in our article on returns management.

Introduction in four steps

AI for logistics is still rare in companies. In the EU, 19.95 percent of enterprises with 10 or more employees used AI technologies in 2025 (Eurostat). Of the enterprises using AI, 34.70 percent used it for marketing or sales (Eurostat), but only 6.08 percent for logistics (Eurostat). If you introduce AI-assisted returns inspection, you are not on well-trodden ground and should start small:

1. Take stock: which product groups come back most, which findings occur, how long does an inspection take today?
2. Define rules: describe the shop standard for each product group: what may the customer do to examine the item, what goes beyond that? Add the criteria for A, B and C grade.
3. Pilot at one inspection desk: the model runs alongside the current process and only shows suggestions. After a few weeks, compare where suggestion and staff judgment differ.
4. Connect and roll out: only once the suggestions hold up are stock booking and refund connected and further inspection desks equipped.

What we take on for AI-assisted returns inspection

We implement AI-assisted returns inspection as an individual project, tailored to your product range, warehouse and existing systems. This includes the inspection rules per product group, the model for detecting marks and checking completeness, the interface at the inspection desk with approval by your team, and the connection to the shop and merchandise management. On request, documents that come with the goods, such as delivery notes or return slips, are read by our AI document processing. The connection to Shopware is handled by the team of our Shopware agency.

The decision on grade and compensation stays with your staff. The page on AI automation shows further use cases.