How AI can handle fraudulent returns
12 Jan 2026
6.5 mins
Fraud is a major problem for ecommerce brands. There are a number of different ways that people can try to commit fraud, and not all of them are easily preventable.
One example could be someone returning the wrong item in a package and then pushing for a refund before a return has been processed. Most brands don’t want to keep loyal customers waiting for their refund and can approve refunds early. This desire to please customers plays into the hands of fraudsters.
On the other hand, waiting until the item has been received and checked means lots of genuine customers are left waiting for their money. Plus sometimes it’s just not practical to wait until processing.
So brands need to strike a balance, and build in some nuance to the process. This is perhaps easier to do with experienced customer service operatives who may be able to sense when something is wrong. But on the face of it, it looks difficult with AI agents, who are more rules-based. This can create a “computer says no” set up that annoys loyal customers.
Modern AI agents can go the other way. Large Language Models (LLMs) are generally eager to please the person they are talking to. So these AI agents could bypass any rules to approve fraudulent returns.
This is where a hybrid approach works. Having an LLM-based approach that can be personable and explain decisions, while being tied to rigid rules means that brands can process returns and refunds for legitimate customers while cutting down on fraudulent claims.
Here’s how we approach preventing fraud in these situations
Validating customers’ identities
The first and most important thing is to ensure that the customer is who they say they are. This means gathering personal information to go through a security check. This prevents situations where a fraudster tries to redirect packages to a different address, away from the original customer, for instance.
The level of checks will vary depending on the brand and its preferences. Some are happy with an order number and a postcode. Others will need the customer to be logged in, or to validate the email address on the order with a code.
This is something that DigitalGenius can build into an agent workflow. DigitalGenius doesn’t store this information, per our data processing, but can connect with the systems you use to validate customer identities.
Checking the return window
Another way customers can try to deceive brands is by returning items well outside the return window and claiming a refund for an item that has been well used. There are also situations where the item sold was sold as a final-sale item and is not eligible for a refund.
Some brands will operate a strict returns window of 30 days. While others may publicly say 30 days, but allow a little extra as grace. That’s sometimes wise because if you give customers 30 days from ordering, and the package is delayed on the way, it eats into the return window.
Then there are cases where the customer gives what feels like a legitimate reason why they were unable to return the item. A hospital stay, or a sudden family emergency, for example. Given the sensitive nature of these situations and to avoid a bad feeling, brands may not want to outright refuse someone who returns something after 33 days, say. This is something that an AI agent can detect.
At the same time, you don’t want to throw your return window out of the window (so to speak!) just because someone has a sob story. Fraudsters are not known for their honesty. So you may want to create a rule for your AI agent where the request gets escalated to a human to review if it’s just outside of your window.
Check whether the return has actually been received
Because DigitalGenius can integrate with your warehouse or reverse logistics provider, we can check to see whether a return package has actually been received.
If a customer is pushing for a refund, but the package is still in transit, then you can build in logic to wait until you have confirmation from the warehouse that the item has been returned. You could even add a check for whether the item has been reviewed.
Check for other return claims
Some people will try to claim multiple times on the same return, in the hope of getting multiple refunds for the same item.
Once again, this is something that an AI Agent can check with your returns platform and see if a refund has been approved and sent.
AI Agents that aren’t integrated with the right platforms will have no way of detecting this information and could therefore miss a previous claim and issue multiple refund requests.
Review customer profiles
In a lot of cases, people committing returns fraud do follow certain patterns. Often, they are first-time buyers who have made a particularly large-value purchase and are looking to claim a return while keeping the item.
For some retailers, there are territories where the physical process of returning the item takes a long time. This can mean that no one has checked the item for weeks or months until after it has been ‘returned’, at which point they find that the product isn’t the right product or is not in a re-sellable state. But of course, they have issued a refund to the fraudster who has then moved on to a different fake profile.
The opposite is also true: that long-term loyal customers do follow certain patterns. Repeated purchases with low numbers of returns, high LTV, VIP membership and so on. You would therefore want to treat the two types of customers differently.
For example, for loyal customers, you could offer expedited refunds – like issuing them when the carrier has scanned the return label. Whereas for ones that throw up red flags, you could have a more stringent process to review the return before issuing the refund.
This can take a little bit of work in your CRM to configure for an AI Agent. You can use specific fields, or even just check the number of orders the customer has made and Lifetime Value. If you have tiers you use internally, you can also check them.
You may even have membership tiers, meaning that “Gold” customers get a different experience to “Bronze” for example.
These are all criteria that DigitalGenius can check before going down the appropriate route. This means that long-term loyal customers get refunds quickly, while potential fraudsters have to wait a little longer.
Add legal disclaimers
Another way to deter fraud is to use legal disclaimers in your messages. This has been used outside of the returns process, when items have been marked as “Delivered” but the customer claims not to have received them. In that situation, before your AI Agent approves a replacement, it can add an affidavit or other legal statement which the customer agrees to in order to receive their replacement.
This approach does deter some potential fraudsters who choose not to risk it. You could also ask customers to validate that the returned items they are sending are genuine to deter any casual fraud in this area.
The advantages of a hybrid approach to fraud
When using AI to approve returns, refunds and replacements, brands have to be careful that their AI agent doesn’t approve everything, but also provides quick and accurate resolutions to legitimate customers.
An approach that relies solely on LLMs or solely on workflows will not work: brands have to be able to blend them correctly. That’s why a hybrid approach is so important. You can have nuance and flexibility, but still apply hard and fast rules when it matters.
An LLM-only approach can lead to your AI Agent skipping key steps such as validation, or making a judgment call based on information in the CRM that may not be accurate.
But, when approached correctly, with an AI provider that understands the complexity of the ecommerce space, brands can automatically escalate incidents of fraud for manual review, and deter fraudulent returns.
To find out more about how we approach returns, speak to the team today.
