Journal / Kirill Naumov
AI in retail loss prevention: find the process signal, not a suspect
A responsible approach to retail loss analytics: start with a specific process problem, verify alerts and keep consequential decisions with people.
By Kirill Naumov ·
‘Shrink’ sounds like one problem. In a real store, it can cover counting errors, damaged stock, receiving discrepancies, process gaps, theft, fraud and data that arrived late. If an AI system is trained on that mixed label, it may become very confident about the wrong question. That matters because the response can affect a customer, a cashier or an entire store team.
The National Retail Federation’s 2026 report illustrates why it helps to separate categories: survey respondents reported fewer shoplifting and merchandise-theft incidents in 2025 than in 2024, while reporting increases in several other forms of theft, fraud and scams. ‘Shrink’ is not a synonym for shoplifting. The right use of AI begins by defining the operational event the team wants to understand.
Start with the loss event, not the model
Choose one question: which deliveries repeatedly arrive with a quantity mismatch? Which stock adjustments lack a reason code or second check? Which refund workflow creates unusual rework? Where do inventory counts diverge after a process change? These are questions about transactions and processes, not a person’s face, clothing or perceived identity.
A predictive model may not be the best first move. A clean exception report, clearer reason codes or a rule that flags repeated mismatches could give the team a better starting point. AI earns its place when it helps prioritise a large review queue, summarise approved records or surface a pattern a person can verify. If a simple report answers the question, the elaborate tool has not earned its keep.
Treat an alert as a question
A useful alert explains what the system observed, what records it used and what a reviewer should check next. For example: ‘This delivery shows a repeated quantity variance against its receiving history; compare the invoice, scan log and adjustment record.’ That is more actionable than ‘high-risk colleague’ or ‘likely theft.’
Keep evidence close to the alert. Reviewers should be able to inspect the source records, see the time period and threshold, and record why they agreed or disagreed. If a system cannot show where a consequential signal came from, it should not trigger a consequential action.
Prioritising a stock count is different from confronting a shopper. AI may help decide which shelf, delivery or workflow to audit first. It should not on its own trigger a stop, search, accusation, disciplinary action or denial of service. Those steps need a defined human process and reliable evidence.
Do not confuse correlation with misconduct
A promotion, delivery delay, substitution, new colleague learning a process, system outage or local event can all create unusual patterns. Yesterday’s ‘normal’ may flag today’s sensible exception. When an alert is wrong, record why; do not quietly erase the case. Otherwise the team cannot improve the data, rule or workflow.
Use special care with biometrics and video analytics. For facial recognition in the UK, the Information Commissioner’s Office asks organisations to assess lawfulness and necessity, document the justification and consider less intrusive alternatives; its guidance is under review after legislative changes. Check current local requirements for each use case. This article is operational guidance, not legal advice.
Design the human review before the pilot
The NIST AI Risk Management Framework emphasises governance, mapping, measurement and management throughout the AI lifecycle. For a store pilot, translate that into practical controls: name an accountable owner; state what the tool may prioritise and which decisions stay with a trained person; log evidence, overrides, complaints and errors; review false alerts across store types, shifts and roles; and define when to pause the test.
These controls are not paperwork for its own sake. They tell you whether the system improves the operation or merely gives extra work a futuristic name. A reviewer must be able to explain the signal and meaningfully influence what happens next.
Measure the operation, not the model’s confidence
Agree on a baseline first. Depending on the use case, track time to resolve an inventory discrepancy, how many alerts reviewers confirm, whether the same process issue returns, how often alerts are overturned and whether employees or customers report harm. Count review costs as well as any reduction in loss; a dashboard full of alerts is not a business result.
Compare like with like, account for seasonality and promotions, and review a sample of unflagged cases too. Otherwise the system can look accurate because no one checked what it missed. Revisit the categories and assumptions when the operation changes.
The best first AI experiment in loss prevention may be unglamorous: a cleaner exception queue, a better handoff between receiving and inventory teams or a report that helps a manager choose which discrepancy to check before the next shift. Start there. If the system cannot help people investigate a process fairly and explainably, it is not ready to judge people. We explore practical AI in retail, sales, learning and loyalty at the AI for Retail workshop.
Explore the AI for Retail workshop → · Start a practical AI diagnostic
Further reading: NRF’s 2026 retail theft and violence report · NIST AI Risk Management Framework · UK ICO guidance on facial recognition.