Rules check eligibility, create labels and refund, AI sorts customer reasons and spots patterns, and people inspect goods and decide exceptions and policy.
Partly, and mostly through rules. Handling returns is a chain of decisions that your return policy has mostly made already: is the item eligible, which label does the customer get, how much is refunded. Returns management software applies that policy reliably. AI adds value around it, by reading what customers write and by spotting patterns in the reasons, while a person judges damaged goods, exceptions and goodwill.
The process: the customer asks to return something, you check eligibility, issue a label, receive and inspect the parcel, decide the outcome, refund and update stock, and afterwards analyse why items came back. Many shops look for returns management software to take this over. Good software removes the email ping-pong and the manual refunds. It does not tell you whether a stained jacket is the customer's fault or whether to make an exception for someone who has bought from you for years. If returns still run through a shared inbox and a spreadsheet, the first gain is one portal with fixed rules. AI is an addition on top of that, and it should suggest, not refuse.
What each step needs
01● Standard automation
Take in the return request and check eligibility
The customer enters an order number or opens a returns portal, picks the items and gives a reason. The system checks the order against your policy: is it within the return window, was the item bought at full price, is it in a category you do not take back. In the United States there is no general legal right to return an online purchase, so the window and the conditions are your own policy, and you must apply them consistently and state them clearly. A portal that does this check at once removes the back and forth by email. The answer is a yes or a no based on rules you wrote down, so no AI is needed.
⛨Works when the policy is written as clear rules (days, categories, conditions) and order data is available to the portal. Unclear cases go to a person instead of being refused automatically.
02● Standard automation
Issue the label and the instructions
Once the return is accepted, the customer gets a prepaid or paper-free label, a drop-off point and a short instruction on how to pack the item. The carrier integration creates the label, the order status changes to 'return announced' and the warehouse sees what is coming. Fixed automation does this reliably, including reminders when the parcel has not been sent after a few days.
⛨Works when you use a small number of carriers and know who pays for the return shipping. Heavy or oversized items may need a booked collection by a person.
03● AI candidate
Read the reason and the customer's message
Reasons arrive as a dropdown choice, a free-text box or an email. AI can read the free text, sort it into reasons such as wrong size, damaged on arrival, not as described or changed mind, and flag messages that mention a safety problem, a complaint or a threat to escalate. It can also draft a reply in the customer's language. The sorted reason feeds the next steps and the analysis later on. A person reads anything flagged as a safety issue or a complaint.
⛨Works when the reason categories are limited and agreed. Check a sample of sorted messages every month, because a wrong category skews the later analysis.
04● Human review
Inspect the returned goods
When the parcel arrives, someone opens it and checks what is inside: is it the right product, is it complete, is it used, damaged or unsellable. Photos can help with a first look, and an AI image check can flag obvious differences, but the grade of the item (as new, open box, refurbish, scrap) is a physical judgement. It decides how much you refund and whether the item can be sold again. Hygiene, safety and fraud checks, such as a different item in the box, need a person who can handle the goods.
⛨Needed for every parcel that is not a straight match. Use a short grading list with photos so that two people grade the same item the same way.
05● Human review
Decide on refund, exchange or credit in exceptional cases
Most returns follow the policy: a full refund for a good item, a partial refund for a used one. The exceptions need a decision: a loyal customer returning something after the window, a damaged item that is not covered, a dispute about who damaged it, or a suspicious pattern of returns. Software can show the order history and the policy, and a model can suggest an outcome with the reasons. A person with a clear mandate decides, because the decision affects the relationship with the customer and sometimes the law.
⛨Works when there is a set limit for what staff may grant without approval, and a record of why an exception was made. Never let an AI refuse a refund on its own in a disputed case.
06● Standard automation
Pay the refund and update stock and accounts
After approval or inspection, the refund goes back to the original payment method, the stock count is updated, the item is booked as sellable, in repair or scrap, and the credit note reaches the accounts. Refunds should reach the customer within the time you promise, and the FTC rule on orders by mail, internet or phone requires you to ship on time or offer a refund, so keep an eye on delays. These are fixed actions between systems. A rules-based workflow does them every time and leaves a trail, which is what your finance team needs at month end.
⛨Works when the webshop, the payment provider, the warehouse system and the accounts are connected, or at least export to each other. Without that link, refunds and stock drift apart.
07● AI candidate
Find out why items come back
Every return carries information: which products, sizes and suppliers come back most, and for which reasons. AI can read the free text and the inspection notes together and point to patterns, such as a size chart that is wrong, a product page that promises too much or a supplier with a defect. The result is a short list of causes that the product and purchasing teams can act on. The model finds the patterns, people decide what to change.
⛨Works with a few months of return data and consistent reason categories. A small shop with a few returns per week can do this by hand in half an hour.
08● Keep human
Set the return policy and the exceptions
How long the window is, who pays the shipping, what you do with used items and how generous you are to good customers are business decisions with a cost and a brand effect. In the United States the policy is largely yours to set, within state consumer law, and it should be clear before the customer buys. Software applies the policy, it does not set it. Review it twice a year using your real return rate, the cost per return and customer feedback, and write the changes down so that the portal and staff use the same version.
⛨Always applies. The more of the process is automated, the more important it is that the written policy is correct and that customers can reach a person.
A sensible first experiment
Take one product group and handle returns as usual for a month, while recording the time per return, the days between the parcel arriving and the refund, the number of emails per return and the share of returns that went back into stock. Then switch on a returns portal with automatic eligibility checks for that group and compare the same numbers for a month. Add the AI step for sorting reasons last, and check a sample of 50 sorted messages by hand. Keep each step only if the numbers improve and complaints about returns do not go up.
The trap to avoid
The usual mistake is automating the refund before the inspection. Money goes out for a parcel that contains the wrong item or nothing at all, and the loss is only visible later in the accounts. The second mistake is letting a model decide disputed cases, which creates unfair refusals and complaints. Keep the refund tied to a clear trigger (carrier scan or inspection) and let a person decide anything that is not in the policy.
Questions teams ask
Do I need AI for returns management?
No. Most of the work follows fixed rules: checking eligibility, creating labels, refunding and updating stock. Returns management software does that without AI. AI is useful for reading free-text reasons and spotting patterns, which is worth adding once the basics run.
Can AI decide whether to refund a customer?
For clear cases the policy decides, and software can apply it. For exceptions and disputes a person should decide. A model can suggest an outcome with the facts, but it should not refuse a refund on its own.
Is it legal to automate refunds?
Yes, as long as the outcome follows your published policy and the law that applies to your sales. Keep a record of each decision, let a person handle disputes, and check which consumer rules apply in each state or country where you sell.
How can I reduce the number of returns?
Look at why items come back. Most reduction comes from better size information, accurate photos and descriptions, and fixing products with defects. AI helps to find the patterns in the reasons, but the fixes are done by your product and purchasing teams.
Illustrative workflow guidance by Arcgent. Each business needs its own assessment. No integration or savings claim has been verified for your systems.