Where AI Actually Pays Off in Retail Operations

Retail has no shortage of AI pilots aimed at personalizing the shopping experience. But some of the strongest business cases for AI sit elsewhere: in the operational problems quietly reducing margins every day.

Return fraud, inaccurate inventory, and shrinkage all have something important in common. They are existing processes with measurable costs, large volumes of operational data, and outcomes that can be compared against a clear baseline.

That makes them particularly suitable for AI projects where the objective is not simply to introduce new technology, but to improve a specific operational metric.

Here are three areas where that approach can make a measurable difference

1. Detecting fraudulent returns without adding friction for legitimate customers 

Returns are a major operational cost for retailers, and fraud adds another layer to that problem.

According to the National Retail Federation's 2025 Retail Returns Landscape, retailers expected approximately $850 billion worth of merchandise to be returned during the year, with 9% of all returns estimated to be fraudulent. Retailers tracking specific forms of fraud also reported increases in practices including overstated return quantities, empty-box returns and counterfeit or decoy products.

Traditional return controls typically rely on fixed rules: return windows, receipt requirements, transaction limits or manual checks by store employees. These controls remain useful, but they can struggle to identify patterns distributed across multiple transactions, channels and customers.

Machine learning can add a behavioral risk layer to the existing return process.

A fraud detection system can analyse signals such as:

• return frequency and timing;
• purchase and return history;
• product categories and values;
• discrepancies between purchased and returned items;
• unusual patterns across stores or channels;
• previously confirmed fraud cases.

Each return can then be assigned a risk score. Low-risk transactions continue through the normal process, while unusual cases can be routed for additional verification.

The objective should not be to make every return harder. It should be to identify the small proportion of transactions that justify additional attention while keeping the experience frictionless for legitimate customers.

What the AI actually does

Transaction and return data → behavioral features → risk scoring → intervention or normal return flow

This also creates an important measurement discipline.

A successful system should track both fraud losses prevented and the false-positive rate. Reducing fraud at the cost of incorrectly blocking legitimate customers simply moves the financial impact somewhere else.

2. Closing the gap between recorded inventory and physical inventory

Omnichannel retail depends heavily on one promise: if a product appears available online, it should actually be available when the customer tries to buy or collect it.

That promise becomes difficult when inventory systems and physical stock fall out of sync.

The problem can originate in receiving errors, misplaced products, inaccurate counts, unrecorded shrinkage, returns, transfers between locations or simply delays in updating systems.

This is different from demand forecasting.

A retailer may predict demand accurately and still fail an online order if the inventory system believes an item is available in a store when the product cannot actually be found.

Technologies such as RFID can significantly improve the frequency and accuracy of physical inventory signals. Retail deployments have demonstrated inventory accuracy improvements from around 65–70% to approximately 95% in specific environments.

AI and analytics can then operate on top of those signals, together with POS, warehouse and ecommerce data, to detect discrepancies and determine where operational intervention is needed.

What the system actually does

POS + inventory system + RFID/computer vision + location data → inventory reconciliation → anomaly detection → corrective action

For example, the system might identify that:

• an item repeatedly appears in inventory but cannot be fulfilled;
• stock is recorded in the wrong location;
• a store shows an unusual pattern of inventory adjustments;
• expected stock movement does not match sales or transfer activity;
• a particular SKU requires a physical count.

The important distinction is that RFID itself is not AI. It creates better inventory signals. Machine learning or analytical systems can use those signals to detect anomalies, prioritize action and improve decision-making.

The most relevant KPIs are therefore operational rather than technological:

• inventory accuracy;
• failed pickup or fulfilment rate;
• phantom stockouts;
• time spent on manual inventory checks;
• product availability.

Improving the prediction model is useful only if it ultimately improves one of these outcomes.

3. Reducing shrinkage with computer vision and transaction data

Shrinkage is often discussed as if it were synonymous with shoplifting. In reality, it is broader: it represents the difference between the inventory a retailer expects to have and what is physically present.

The causes can include theft, checkout errors, process failures, administrative mistakes and other forms of inventory loss.

AI cannot address every source of shrinkage. But computer vision can be particularly effective where the loss is associated with observable store or checkout events.

Self-checkout is a good example.

A vision system can analyse defined checkout interactions and cross-reference them with point-of-sale transactions. Instead of reviewing video after an inventory discrepancy is discovered, the system can identify unusual events while the transaction is taking place.

Examples include:

• an item entering the bagging area without a corresponding scan;
• the scanned product not matching the observed item;
• repeated unusual checkout patterns;
• transaction events that differ significantly from normal behavior.

What the AI actually does

Video event + POS transaction → event matching → anomaly detection → real-time alert or later review

The objective is not autonomous enforcement.

Well-designed systems use AI to identify a much smaller set of events that deserve attention, allowing employees or loss-prevention teams to focus their review where it is most useful.

Computer vision can also create operational signals beyond loss prevention. Depending on the system architecture and privacy model, the same infrastructure can support analysis of traffic patterns, queue behavior, store utilization and other operational events.

But these additional applications should be designed intentionally rather than treated as an automatic consequence of installing cameras.

Privacy also needs to be part of the architecture from the beginning. Systems can often be designed to detect events and behavioral patterns without identifying individual shoppers, with clear rules around data access, retention and processing.

The common architecture behind all three use cases

Return fraud, inventory accuracy and shrinkage initially look like three separate retail problems.

Technically, however, they share a common pattern. Retail systems describe what should have happened. Operational signals provide evidence of what actually happened.

AI becomes useful when it can identify meaningful differences between the two.

• Transactional systems + physical-world signals
• Unified operational data
• AI / anomaly detection
• Decision or intervention
• Measured business outcome

For returns, the discrepancy may be between expected customer behavior and an unusual return pattern.

For inventory, it is the difference between recorded stock and physical availability.

For shrinkage, it may be the difference between what the POS recorded and what occurred during the checkout interaction.

This common architecture also creates opportunities beyond individual use cases. Once retailers can reliably combine transaction, inventory and physical-event data, the same data layer can support multiple operational applications.

Before AI: is the operational data entry?

A strong AI use case does not necessarily require perfect data before the project starts.

It does require understanding what data exists, how reliable it is and whether different systems can be connected.

Return fraud detection may depend on transaction history, returns records and product data.

Inventory reconciliation may require POS, warehouse, ecommerce, store-location and physical inventory signals.

Computer vision for checkout events requires accurate synchronization between camera events and POS transactions.

In many projects, building or improving this operational data layer is part of the AI initiative itself.

Before choosing a model or platform, retailers should therefore answer a more fundamental set of questions:

• What business event are we trying to detect or predict?
• Which systems currently record that event?
• Is there enough historical data to establish a baseline?
• How will an AI output change an operational decision?
• Who acts when the system identifies an anomaly?
• Which KPI determines whether the project succeeded?

If those questions cannot be answered, the project is probably not ready for model selection yet.

How Exaud aproaches operational AI projects in retail

At Exaud, operational AI projects start with the business process rather than the model.

The first step is to understand where the measurable loss or inefficiency occurs and establish the baseline against which improvement will be evaluated.

From there, a typical project can involve:

1. Baseline the operational problem
Quantify the current fraud rate, inventory discrepancy, shrinkage pattern or other target metric.

2. Audit the available data
Identify the POS, ecommerce, ERP, WMS, RFID, camera or other operational data that can support the use case.

3. Build the required integrations
Create the data flows needed to connect systems that were not originally designed to work together.

4. Prototype against historical data
Test whether useful patterns can be detected before introducing the system into live operations.

5. Measure accuracy and operational impact
Evaluate precision, false positives and the effect on the business KPI defined at the start.

6. Integrate AI into the workflow
Connect predictions or alerts to the people and systems responsible for acting on them.

7. Monitor business performance after deployment
Track whether improvements continue once the system operates under real-world conditions.

This is what separates an interesting AI experiment from an operational system with a defensible return.

For retailers evaluating where to start, the best opportunity is often not the most ambitious AI idea. It is the operational problem with a meaningful cost, sufficient data and a clear action that can be taken when the system detects something important.

Frequently asked questions

Which of these three retail AI investments delivers the fastest return?

Loss prevention and returns fraud detection typically show measurable results fastest, often within three to six months of deployment, because both problems generate a constant stream of transaction-level data to train and validate the system against. Inventory accuracy improvements usually take longer to show up in the top-line numbers, since the benefit compounds through fewer failed promises over time rather than an immediate dollar figure.

Do these AI projects require replacing existing point-of-sale or inventory systems?

No. The strongest results come from AI layered on top of existing point-of-sale, camera, and inventory systems, cross-referencing data that already exists rather than replacing the systems that generate it. A full systems replacement is a much larger, higher-risk project, and it is rarely necessary to capture the gains described here.

How do you measure wether a returns fraud AI project os working?

Track fraud dollars prevented against a pre-project baseline, alongside the false-positive rate for legitimate customers. A project that reduces fraud losses but also increases complaints or lost repeat customers from wrongly flagged returns has not actually solved the problem, only moved the cost somewhere else.

Is inventory accuracy mainly a technology problem or a process problem?

Both, but process gaps are usually the larger contributor. Manual counting errors, receiving mistakes, and unrecorded shrinkage all degrade the accuracy of the underlying data before any AI system touches it. The most effective projects pair a technology layer, such as computer vision or RFID, with a review of where the recorded data is falling out of sync with reality in the first place.

Does reducing shrinkage with computer vision raise private concers?

It can, and the systems that hold up best are built to detect behaviors and patterns rather than to identify individual shoppers, with clear policies on what is recorded and how long footage is retained. This is a design decision made early in the project, not something added after deployment, and it materially affects which vendors and architectures are appropriate for a given retailer's regulatory environment.