AI agent · Retail & e-commerce · Score & detect
Return fraud patterns
The agent scores each return against the history of the customer, the item, the store and the channel. Normal returns are refunded as usual; the outliers reach a loss-prevention analyst every day, with the reasons they stood out.
Typical volumes for this process, not a client figure.
Patterns only spotted when someone happens to notice.
Every return scored against history; the outliers surfaced daily.
Where the time goes today
Most returns are honest: the wrong size, a change of mind, a faulty item. A few are not. Clothing worn to an event and brought back, a receipt reused for a second item, stolen goods returned for a refund, an empty box sent back to the warehouse, a refund paid to a different card, a till operator refunding friends. Each looks ordinary on its own. The pattern only appears across many returns: one customer working through several stores, the same item coming back again and again, one till with an unusual run of no-receipt refunds.
Today these patterns are found by chance: a store manager who recognises a face, an analyst who runs a report after a stock count looks wrong. Till, online order and warehouse data sit in separate systems, so the reports are assembled by hand, monthly at best, long after the refund has been paid. Blanket rules such as no refund without a receipt catch some abuse but irritate the honest majority and generate complaints.
How the agent works
- Gather each returnThe agent collects every return from tills, the online returns portal and warehouse receiving: item, original sale, tender used, refund method, store, till operator, stated reason and the condition recorded on receipt.
- Link the historyIt links the return to its original sale and to earlier returns by the same customer, card, address or loyalty account, and to the item's return history across channels.
- Score against indicatorsIt computes a score from indicators your loss-prevention team has written down, such as refunds to a different tender, frequent no-receipt returns, or a condition mismatch at the warehouse. Each indicator's contribution is visible.
- Refund or holdReturns below the threshold follow your refund policy with no delay. Above it, a till refund is reviewed afterwards; an online refund can be held for a limited time where your policy allows.
- Explain the outlierEach surfaced case arrives with the indicators that fired, the linked returns and what the analyst should check first.
- Record the outcomeThe analyst's conclusion is stored, and indicators are recalibrated periodically against confirmed cases, with loss-prevention sign-off.
What stays with a person
A score is not an accusation. An analyst decides whether an outlier is abuse, an error, or an honest customer with an unusual pattern, such as a parent returning items for several children. Refusing a refund, closing an account, banning someone from returns, and anything involving a member of staff are decided by people, under your policy and the consumer and employment rules that apply where you trade.
Loss prevention also owns the indicators and the threshold. The agent does not invent new signals; a proposed change is tested on past returns and approved before use. Some indicators can act as proxies for characteristics you must not use, and only a person can check that.
What it reads, what it produces
| It reads | It produces |
|---|---|
| Till return transactions, with operator and store | A score for every return, with the indicators behind it |
| Online orders and returns portal records | A daily list of outliers, each with its linked history |
| Warehouse returns receiving, including condition grading | Held online refunds awaiting an analyst's decision |
| Customer, loyalty and payment token data, as your privacy rules permit | A periodic report on which indicators confirmed and which misled |
| Your returns policy and past loss-prevention case records |
Controls that come with it
- The threshold is set by arithmetic: expected loss from a missed abusive return against the cost of review and of delaying an honest refund.
- Holds are time-limited; if no analyst acts, the refund goes through.
- Indicators are reviewed for fairness and data protection before use, and exclude protected characteristics and their obvious proxies.
- Cases involving staff go to a restricted team, not the general queue.
- Every score can be reproduced from the recorded data, indicators and weights.
- A sample of returns below the threshold is reviewed periodically to see what is being missed.
How you know it works
- Share of surfaced outliers that analysts confirm as abuse
- Confirmed abuse found later that the agent had scored below the threshold
- Time from return to analyst review
- Complaints from customers whose refund was held
Is your process ready?
- The indicators of abuse are written down and agreed by loss prevention, not held as instinct by one investigator.
- Till, online and warehouse systems export returns with enough identifiers to link them, and your licences and data protection rules allow that linking for this purpose.
- An analyst can confirm or clear an outlier from its case file in minutes, so the outcome is cheap to check.
- There are too many returns for anyone to look at each one, which is the reason to score them.
- Stores, online and loss prevention agree on where generous policy ends and abuse begins.
The five candidacy checks are explained, with an exam, in the free Module 01.
What goes wrong
- Customers cannot be linked across channels because of guest checkouts and cash, so patterns fragment. Start where linking is reliable.
- The threshold is tuned on too few confirmed cases, analysts are flooded with false alarms, and they stop reading the list.
- After peak gifting periods, normal behaviour shifts. Indicators that ignore the season flag honest customers in bulk.
- Holding refunds on a score alone harms honest customers and the brand. Keep holds rare, short and reviewed.
Questions we get
Will honest customers be refused refunds?
The agent refuses nothing. Returns below the threshold follow your policy unchanged. At the till, the refund is made and any outlier is reviewed afterwards. Online, a refund can be held for a limited period where your policy allows, and it goes through if nobody acts. Any refusal is a person's decision, made on the case file.
Is the score a black box?
No. It is built from indicators your team has written down and agreed, and each contributes a visible amount. The analyst sees which ones fired and on what data. Weights may be fitted to past confirmed cases, but the indicators stay readable, and any score can be reproduced from what was recorded at the time.
What data do we need to start?
Returns linked to their original sales, at least in the channels where you suspect abuse, with some way to connect a customer or payment method across transactions. You also need past cases your team has already confirmed or cleared. Without them you cannot measure how often the outliers are right, so assembling that set is often the first piece of work.
Can it detect collusion by staff?
It can surface patterns at till and operator level: refunds to the same card, refunds outside trading hours, no-receipt refunds well above the store's norm. These cases are sensitive. They go to a restricted team and are handled under your investigation procedure and the employment rules that apply. The agent reports the pattern; people investigate it.
Want this agent on your process?
Tell us about your version of this process — volumes, systems, what goes wrong. A person answers with an approach and a price, usually within two working days, or tells you it is the wrong project.