US retail shrink exceeded $112 billion in the most recent NRF survey. Employee theft and organized retail crime account for the majority of that figure — but most of it is discovered at inventory count, weeks or months after it happened. By then, evidence is gone and investigation is impossible.
The stores that catch shrink in real time have measurably lower rates. The difference isn't luck — it's the tools. AI loss prevention gives loss prevention teams visibility between incidents and inventory counts. That window is where most shrink becomes irrecoverable.
The shrink problem in retail
The core challenge isn't that retailers don't know shrink is happening. It's that they find out far too late. The detection gap — the time between when a loss event occurs and when it's discovered — is where most shrink becomes unrecoverable. By the time an inventory discrepancy is flagged, there is no footage worth reviewing, no transaction trail worth investigating, and no employee behavior pattern worth acting on.
Most loss occurs not in dramatic theft events but in accumulation: refund fraud processed incrementally across hundreds of transactions, scan avoidance repeated across dozens of shifts, after-hours access that goes unquestioned because nobody is watching. These patterns are only visible in aggregate — and only if someone is looking continuously, not retroactively.
Stores that catch shrink in real time have measurably lower rates. But most don't have the tools to do it. Security cameras record everything and surface nothing. POS data captures transactions but can't tell you what was happening at the register when the transaction was processed. Loss prevention teams work cold leads by default.
Why traditional loss prevention misses most shrink
Traditional loss prevention operates across three channels — none of which are continuous or scalable across a large store estate.
- Security staff on the floor: Expensive to staff at meaningful levels, inconsistent across locations and shifts, and actively harmful to customer experience when overt. Even the best LP associate can only observe one area at a time and cannot be everywhere.
- Post-incident CCTV review: Labor-intensive, requires knowing exactly what to look for and approximately when the incident occurred, and provides no coverage for events that haven't yet been identified as incidents. Most theft events are never reviewed at all — they become inventory variance.
- Exception-based POS reporting: Catches statistical anomalies in transaction data — suspiciously high refund rates, repeated voids, unusual price lookup frequency — but delivers statistical flags, not video evidence. Without the visual context of what was happening at the register, these flags are difficult to act on for prosecution or internal disciplinary action.
None of these approaches are continuous, scalable, or capable of correlating behavioral evidence with transaction evidence in real time. The result is that most loss prevention is reactive by design — reviewing cold evidence for incidents that were discovered by accident, not by detection.
How AI loss prevention works
AI loss prevention platforms analyze camera feeds and POS transaction data simultaneously, in real time. Computer vision models trained on retail-specific behaviors can detect signals that human observers and exception-reporting systems miss individually but that become unmistakable in combination.
The behaviors these models are trained on include: refund patterns at the POS that correlate with historical shrink data, after-hours motion in secured stockroom or back-of-house areas, spatial movements that precede or accompany shoplifting incidents, and scan avoidance behavior at self-checkout terminals. The models do not look for a specific individual — they look for a pattern that matches known shrink behavior signatures.
When a pattern is detected, the loss prevention team is alerted immediately — not with a statistical flag that requires manual investigation, but with a footage clip and contextual data. The clip shows what happened. The context shows when, where, and what transaction (if any) was in progress. LP teams can review the alert in under two minutes and decide whether to escalate, investigate further, or dismiss.
The operational shift AI loss prevention enables is from cold-lead investigation to real-time detection. Evidence is fresh. The transaction is still in the system. The employee is still on shift. This is the window where intervention is possible — and where most traditional LP programs have no visibility at all.
AI loss prevention also changes the economics of scale. A single loss prevention manager reviewing AI-generated alerts can effectively cover a store estate that would previously require a dedicated LP associate in each location. Coverage becomes continuous, not staffing-dependent.
POS exception detection: the highest-ROI use case
Of all the detection capabilities AI loss prevention platforms offer, POS exception detection combined with camera correlation delivers the fastest and most measurable ROI. The reason is straightforward: it converts statistical flags that were previously difficult to act on into actionable visual evidence.
The mechanics work as follows. The AI loss prevention platform ingests transaction data from the POS system in real time. When a transaction-level anomaly is detected — a suspicious refund, a price lookup override on a high-value item, a repeated scan void — the platform automatically pulls the camera footage of the register at the relevant timestamp. The LP manager receives the alert with both the transaction detail and the video clip already linked.
What previously required 30 minutes of footage scrubbing to verify a single suspicious transaction now takes under 2 minutes. Across a multi-location estate, this multiplier effect is significant. Managers can review a full day's exception queue in the time it previously took to investigate one event.
"68% of refund fraud is caught too late under traditional post-incident review models." POS exception detection with camera correlation changes that — turning transaction anomalies into timestamped visual evidence while the event is still fresh.
For retailers with high refund volumes — fashion, electronics, home goods — POS exception detection is typically the first AI loss prevention module deployed, because the ROI is demonstrable within the first inventory cycle.
→ See EdgeRetail Guard — AI Loss Prevention & Shrink Detection
BIPA and privacy compliance
Privacy compliance is the most consequential evaluation criterion for AI loss prevention, and the one most frequently mishandled in vendor selection. The Illinois Biometric Information Privacy Act (BIPA) carries statutory damages of $1,000 to $5,000 per violation per person for unauthorized collection of biometric identifiers — and plaintiffs' attorneys actively pursue BIPA claims in retail. This is a material legal risk, not a theoretical one.
The key compliance question for AI loss prevention platforms is whether the platform collects or processes biometric identifiers — specifically, facial geometry data from facial recognition systems. Many loss prevention platforms that claim BIPA compliance are actually processing facial recognition data in states where it is not yet regulated, and exposure exists the moment those platforms expand into Illinois or to states with analogous legislation.
EdgeRetail Guard does not collect biometric identifiers. The platform analyzes behavior patterns and spatial movements — it does not identify individuals by facial geometry or store any biometric data. This architectural decision is the basis for its explicit compliance with BIPA, CCPA, GDPR, SOC 2, and ISO 27001.
When evaluating any AI loss prevention vendor, request explicit compliance documentation — not a general privacy policy, but specific certification for the jurisdictions where your stores operate. Full security and compliance documentation for EdgeRetail is published at trust.edgesignal.ai.
Evaluating AI loss prevention platforms
The AI loss prevention market includes a range of vendors from enterprise surveillance providers adding AI features to purpose-built retail analytics platforms. The following criteria distinguish production-ready platforms from marketing-stage products.
| Criterion | What to verify |
|---|---|
| Detection types covered | Does the platform detect refund anomalies, after-hours access events, and behavioral shrink patterns? Verify each is in production, not on the roadmap. |
| POS integration | Does the platform ingest POS transaction data in real time? Which POS systems are supported natively, and what is the integration timeline for your vendor? |
| Biometric data policy | Does the platform use facial recognition or store facial geometry data? Request a written statement of biometric data policy before proceeding. |
| BIPA / GDPR certification | Ask for explicit compliance documentation — not a checkbox on a marketing page, but a published trust center or auditor letter covering BIPA, CCPA, GDPR, and SOC 2. |
| False positive rate | High false positive rates destroy LP team trust in the system within weeks. Ask for false positive benchmarks from comparable retail deployments and request a tuning SLA. |
| Alert routing | How are alerts delivered — email, mobile push, dashboard? Can alerts be routed by store, region, or incident type? Will your LP team actually see and respond to them? |
Before committing to a full deployment, run a structured pilot in 2–3 stores with different operational profiles. Measure shrink rates, alert response times, and LP team engagement over a full inventory cycle. If the platform cannot demonstrate measurable detection improvement in a pilot, the full estate deployment will not perform differently.
EdgeRetail Guard is the loss prevention and shrink detection module of the EdgeRetail platform, built on EdgeSignal by Redesign Business LLC. It connects to existing cameras, requires no biometric data collection, and is certified for BIPA, CCPA, GDPR, SOC 2, and ISO 27001. Learn more →