Start with the number, because the number is the first thing that's wrong. The widely repeated claim that organized retail crime costs tens of billions of dollars a year traces back to a 2023 National Retail Federation report asserting that ORC accounted for nearly half of the $94.5 billion in 2021 retail shrink. The NRF retracted that claim in December 2023, after reporting traced it to Senate testimony that had actually described total shrink rather than ORC. There is no reliable public estimate of what ORC costs nationally.
What there is: shrink totals that retailers do report, and case-level evidence that some portion of it is coordinated rather than opportunistic. Operations that target specific products, cycle locations to avoid pattern detection, and move merchandise through supply chains spanning multiple states are real and well documented in individual prosecutions. If you are building a business case on a headline ORC dollar figure, find out where that figure came from before you put it in a deck. Mine came apart when I checked it.
AI claims to solve ORC are everywhere. Every security technology vendor has an AI angle. Detection, prediction, network analysis, perimeter intelligence. The slide decks are sophisticated and the demos are compelling.
After spending time with security leaders who work at the intersection of retail loss prevention and ORC intelligence, people running investigations, coordinating with law enforcement, and actually building or evaluating these systems, here's what's working versus what's theater.
What's actually working
Network visualization for known actors. This is where AI adds genuine value. ORC rings operate as networks: a fence, a booster, a driver, a storage location, a buyer. Individual incidents look like isolated shoplifts. Network analysis connects them into an organized operation with identifiable members.
The AI application that works is not prediction. It's connection. Taking incident data from multiple locations, matching descriptors, correlating timing with travel patterns, and surfacing the network structure that individual-incident review would never find. Human analysts then validate the connections and build the case. AI finds the signal in the noise.
Image-based search across large evidence sets. ORC investigations generate enormous evidence volumes: hours of CCTV footage, thousands of still images, vehicle records from multiple jurisdictions. Manual review at this scale is impractical. AI-assisted image search, match this face, find this vehicle, identify this product pattern, is genuinely useful when the matching is done as investigative assistance rather than autonomous identification.
The distinction matters: a detective using AI to find similar images in a database is different from a system automatically identifying suspects and triggering actions. The first is a productivity tool. The second is a liability the moment it gets something wrong, which it will.
Pattern analysis across incident reports. The best ORC applications I've seen ingest structured incident data from multiple retailers in a region, normalize the formats, and identify shared patterns: same target products, overlapping time windows, same booster descriptor, similar modus operandi. This requires data sharing agreements between retailers. The technical problem is easy; the legal and competitive reluctance is hard. When the data sharing exists, the pattern analysis produces actionable intelligence.
What's theater
Predictive prevention. The claim that AI can predict where ORC activity will occur before it happens is mostly marketing. ORC rings are adaptive. They change locations, timing, and methods specifically to avoid patterns. Static prediction models get trained on historical data that the rings have already moved past.
What some vendors are actually doing, and what's useful, is forecasting based on leading indicators: changes in specific product pricing, shifts in resale market demand, law enforcement activity patterns that move operations to adjacent areas. This is more like economic trend analysis than crime prediction. It's useful but different from what's being sold.
Real-time facial recognition for prevention. In my view the accuracy claims and the liability exposure don't line up. NIST's demographic-effects evaluations of face recognition algorithms have documented error rates that vary by demographic group, and a false match at a store entrance is an incident with its own cost. Biometric privacy statutes in several states, Illinois' BIPA most prominently, add litigation exposure on top of that. The organizations using image technology effectively are using it retrospectively in investigations, not prospectively at points of entry.
AI-generated prosecution packages. I've seen demos of systems that automatically generate law enforcement referral packages from incident data. The packages look professional. My objection is structural rather than empirical: prosecution depends on chain-of-custody documentation, witness statements, and evidence handling that a generated document cannot itself supply. A model can organize and draft; it cannot create the evidentiary foundation underneath. A package whose provenance an analyst cannot walk through under cross-examination is a package I would expect to be challenged. I know of no published data on outcomes either way, and I am not claiming any.
The coordination problem that AI doesn't solve
The biggest opportunity in ORC intelligence isn't AI. It's data sharing. Most retailers have incident data that would be valuable to adjacent retailers in the same region. Most of them don't share it because of competitive concerns, liability concerns, and the absence of a trusted intermediary.
The organizations making the most progress on ORC intelligence are the ones that have solved the coordination problem: regional coalitions with formal data sharing agreements, law enforcement liaisons, and shared intelligence platforms. The AI layer on top of shared data is an accelerant. The AI layer on top of siloed data produces siloed intelligence.
The conferences I've attended on ORC and AI security have the same productive conversation: not "what can AI do for ORC?" but "how do we get the data that makes AI useful?" The technology is further along than the institutions.
What this means for security leaders
If you're evaluating AI for ORC, the questions that matter are:
- What data will this system have access to, and is that data high enough quality to produce signal rather than noise?
- Is this tool positioned as investigative assistance for trained analysts, or as autonomous action without human review?
- What's the liability exposure if the system produces a false identification that leads to an incident?
- Does this product require data sharing with other retailers, and if so, how is that governed?
The vendors who can answer all four questions clearly are building real products. The ones who pivot to the demo when you ask about data sources are building theater.
PurviewX builds intelligence platforms for security-intensive industries. Start a conversation.