The Retail Shrink Measurement Gap: A 2026 ROI Framework on Verified Detection Data, the NRF Reporting Gap, and the AI Video Analytics Business Case
NRF says shoplifting incidents rose 18 percent in 2024. The FBI's own two crime-reporting systems disagree by 93 percentage points on the same underlying trend. A primary-source economic model for loss-prevention leaders who need a defensible number instead of a contested one.
Every retail loss prevention budget request in 2026 leans on a statistic from somewhere: a National Retail Federation survey, an FBI crime report, a vendor's shrink calculator. Few of the people citing those numbers have checked whether the sources agree with each other. They do not. The retail industry's own trade association reports shoplifting incidents climbing double digits year over year, while the federal government's two parallel crime-reporting systems disagree by more than ninety percentage points on the same underlying trend, and roughly two-thirds of retailers admit they never send most of their theft incidents to police in the first place. This report builds a primary-source case for a different kind of baseline, one generated by a facility's own AI-verified detection data rather than borrowed from a national statistic nobody can fully reconcile, and shows how to translate that baseline into an ROI model a finance committee will not send back with questions.
Retail loss prevention is one of the few corners of physical security where the buyer already believes the threat is real. Shrink is a board-level line item. Organized retail crime has its own federal legislation working through Congress. The problem is not persuading a retail security director that theft and violence are increasing. The problem is that the three sources retailers cite to prove it, law enforcement statistics, their own industry association's incident survey, and informal internal counts, do not describe the same reality, and a capital request built on the wrong one will not survive scrutiny.
This report does three things. First, it lays out what the National Retail Federation and the Loss Prevention Research Council are actually hearing from retailers in 2025, separated from the shrink-dollar figures the industry stopped publishing in a comparable format. Second, it walks through why the federal government's own two crime-reporting systems, the Summary Reporting System and the National Incident-Based Reporting System, tell contradictory stories about the same shoplifting trend, and why that contradiction is not a rounding error. Third, it builds a four-input ROI framework that lets a loss prevention leader construct a defensible business case for AI video analytics using a baseline the organization controls, instead of one it has to hope survives a fact-check.
The Retail Shrink Reporting Gap: Why the Industry's Best Number Might Be Off by Half
The National Retail Federation and the Loss Prevention Research Council released The Impact of Retail Theft & Violence 2025 on October 28, 2025, based on a survey of 70 retail companies representing 168 brands, $1.3 trillion in 2024 fiscal-year sales, and 25.1 percent of total U.S. retail sales. The headline finding: retailers reported an 18 percent increase in the average number of shoplifting incidents per year in 2024 versus 2023, and threats or acts of violence during shoplifting or theft events increased 17 percent over the same period. More than half of respondents reported increases in phone scams, digital and e-commerce fraud, shoplifting and merchandise theft, and cargo or supply-chain theft conducted by organized retail crime groups over the prior twelve months, and 67 percent reported the involvement of a transnational organized retail crime group in thefts against their company.
Those numbers describe what retailers are experiencing and choosing to report to their own trade association. They do not describe what shows up in a police blotter, and the survey itself explains why. Sixty-four percent of retailers told NRF and LPRC that they report less than half of their store-related theft incidents to law enforcement, citing a lack of law enforcement response as the primary reason. That single admission does more to explain the confusion around retail crime statistics than any methodology footnote. If two-thirds of retailers are not reporting the majority of what happens on their sales floor, then any statistic built on police-reported data, which is most of what makes national news, is measuring a minority subset of a much larger problem, and the size of that minority can shift from year to year based on nothing more than a retailer's changing confidence in local prosecution, not a change in actual theft volume.
"Retailers are contending with rising levels of theft, fraud and violence, while continuing to refine security measures, utilize technologies and partner with law enforcement in efforts to curtail loss across the retail landscape."
David Johnston, NRF Vice President for Asset Protection and Retail Operations, October 2025 press release
The dollar-shrink side of the ledger is even harder to pin down. The National Retail Federation's last full National Retail Security Survey, covering fiscal year 2022, put the average shrink rate at 1.6 percent of sales, translating to roughly $112.1 billion in annual losses, up from 1.4 percent and $93.9 billion in fiscal year 2021. NRF did not publish a comparable shrink-dollar survey for 2023 or 2024, stating that the prior methodology no longer captured the evolving nature of the problem, and pivoted instead to the incident-frequency format behind the 2025 Impact of Theft & Violence report. That is a defensible research decision. It also means the retail industry's single most-cited historical shrink figure, the number still quoted in press coverage and vendor sales decks in 2026, is now several years old and was never designed to be extrapolated forward without a comparable successor.
None of this means retail theft and violence are not increasing. The direction of the NRF survey data, the FBI data discussed below, and independent city-level reporting all point the same way. What it means is that the specific magnitude, the number a CFO wants on a slide, cannot be pulled cleanly from any single national source in 2026, because the industry's own trade association has told you that most of the underlying incidents never get reported, and the government's own reporting infrastructure produces two different answers for the incidents that do. Even the federal government's general workplace injury data cuts against a simple narrative: the U.S. Bureau of Labor Statistics' January 2026 Survey of Occupational Injuries and Illnesses release showed the retail trade sector's total recordable injury rate declining in 2024, alongside the lowest employer-reported injury and illness count in the series' history back to 2003, a very different signal than the rising-violence trend retailers describe in their own survey responses. Three federal and industry data series, three different directions, all describing the same year.
Two Federal Systems, One Crime, Two Different Answers
The Council on Criminal Justice, a nonpartisan research and policy organization, has published the most detailed independent analysis of this problem. Its November 2024 report, updated through October 2025, examined shoplifting data from the FBI's two Uniform Crime Reporting Program data sources: the Summary Reporting System, the older monthly-summary format, and the National Incident-Based Reporting System, the newer incident-level format that is gradually replacing it. The two systems, drawing on the same underlying pool of law enforcement agencies and the same underlying crime, produce dramatically different trend lines for the identical period.
By the Summary Reporting System's count, the number of shoplifting incidents reported in 2023 was roughly comparable to 2019, essentially flat over four years. By the National Incident-Based Reporting System's count, the shoplifting rate was 93 percent higher in 2023 than in 2019, rising from 159.3 to 308.8 incidents per 100,000 population. That is not a rounding difference. It is two federal measurement systems, tracking the same crime over the same period, arriving at conclusions that differ by more than ninety percentage points.
Part of the explanation is a composition effect, not a crime-rate effect. In 2019, only 46 percent of law enforcement agencies serving jurisdictions of 250,000 or more residents submitted data through the National Incident-Based Reporting System. By 2023, 86 percent did. The number of NIBRS-reporting agencies nationwide nearly doubled over that period, from 8,497 to 16,334. When large urban departments join a reporting system for the first time, their historical volume of incidents appears in the data as new growth, even if the underlying crime rate in that city did not change. The Council on Criminal Justice's own analysis, using only agencies that reported consistently to NIBRS from 2019 through 2021, found a trend pattern much closer to the flat Summary Reporting System numbers than to the 93 percent NIBRS headline. The FBI's own Crime Data Explorer acknowledges that figures may not be comparable across the SRS-to-NIBRS transition, but as the Council's researchers note, general caution language is not the same as clear guidance on which figures a policymaker, or a retail security director, should actually use.
"It is unclear why such a substantial difference in reported shoplifting exists between these two sources... If NIBRS is not only the future but also the present of crime data reporting in the U.S., providing general statements is insufficient."
Council on Criminal Justice, Between the Aisles: A Closer Look at Shoplifting Trends
City-level data compounds the confusion rather than resolving it. Chicago's reported shoplifting rate for the first ten months of 2024 ran 46 percent higher than the same period in 2023, and higher than any full prior year in six years. Los Angeles reported shoplifting 87 percent higher by the end of 2023 than in 2019, though Los Angeles Police Department's March 2024 transition to a new NIBRS-based records system caused the reported incident count to drop from over 17,000 in February 2024 to roughly 8,000 by June, an artifact of the reporting-system change, not a sudden crime decline. New York's reported shoplifting rate was 55 percent higher in 2023 than in 2019, following a 48 percent jump from 2021 to 2022 and a slight pullback in 2023. Three of the country's largest cities, three different trend shapes, and at least one of the three data series is known to be distorted by a records-management transition rather than a change in actual theft.
A data-quality problem this large is not an academic footnote when it sits underneath a capital request. If a security director builds a business case on "shoplifting is up 93 percent," a finance committee member who has read a single news article about the SRS-NIBRS discrepancy can dismantle the entire proposal in the first five minutes of the meeting. The credibility of the ROI model depends on where the baseline number comes from, and a national statistic that even its own publisher cannot fully explain is a fragile foundation for a multi-year technology investment.
What This Means for the AI Video Analytics Business Case
The reporting gap described above is not an argument against acting on retail theft and violence. It is an argument for where the baseline data should come from. A national shoplifting statistic, however it is measured, describes an aggregate trend across thousands of dissimilar stores, jurisdictions, and reporting practices. It was never designed to tell a single facility what is actually happening on its own sales floor, in its own parking lot, or at its own loading dock, and the discrepancies documented above show that even the aggregate trend is contested at the federal level.
AI-verified detection data solves a narrower, more tractable version of the same problem. Instead of asking "what is the national shoplifting rate," a facility asks "what did our own cameras classify, with what confidence, and when." Modern computer vision platforms process existing camera feeds and return a bounding box with a confidence score around a classified object or behavior, a person entering a restricted zone after hours, a vehicle lingering at a loading dock, a person loitering near an entrance outside business hours, in real time. That detection event carries a timestamp, a camera location, and a confidence value, and it exists independent of whether anyone later decides to call police, independent of which federal reporting system that police department uses, and independent of the retailer's own appetite for filing a report.
This is where the retail loss prevention conversation connects to IntelliSee's broader detection platform. The same computer vision models that support loitering detection in a parking lot and perimeter control at a loading dock apply directly to the pre-incident behaviors that precede organized retail crime and shoplifting: a vehicle staged near an exit, a group loitering near an entrance before closing, a person testing an access point after hours. None of it requires new camera hardware in most stores; the detection layer runs on the existing IP or analog camera systems a retail facility already has, which matters in a sector where capital budgets for physical security compete directly against merchandising, labor, and store-remodel spend.
Retail is a setting where biometric identification carries real legal exposure, from state biometric privacy statutes to customer-facing reputational risk. IntelliSee performs no facial recognition, stores no video, and does not build a customer identity profile; the platform classifies an event, such as a person in a restricted zone or a vehicle at a loading dock, and returns a bounding box with a confidence score without identifying who that person is or retaining the underlying footage. That architecture is what allows verified detection data to function as an internal measurement system, rather than a surveillance liability layered on top of an already-scrutinized retail environment.
Building the Four-Input ROI Model for Retail Loss Prevention
A defensible ROI model does not need a single perfect national number. It needs a small set of inputs that a finance committee can trace back to a named source, and a structure that shows how those inputs combine into a dollar figure. The framework below decomposes the retail AI video analytics business case into four inputs, each grounded in a specific, citable data source, and each auditable independent of the others.
Compare what AI-verified detection surfaces at a facility, every classified event with a timestamp and confidence score, against what currently reaches an internal incident log or a police report. With 64 percent of retailers reporting fewer than half their theft incidents to law enforcement, the delta between verified detection volume and reported-incident volume is itself the first quantifiable input: the size of the previously invisible loss stream.
NRF/LPRC underreporting rate, facility-level detection log
Apply a dollar value to the verified-incident count. Absent facility-specific average-loss data, use NRF's last full figure, a 1.6 percent shrink rate against sales, clearly dated to fiscal year 2022, rather than an unsourced or rounded figure from a vendor calculator. Update this input the moment a facility has its own average-loss-per-incident data from actual case resolutions.
NRF National Retail Security Survey, FY2022, dated explicitly
Unverified alerts, motion triggers, generic alarm activations, consume loss prevention labor and, when escalated, law enforcement dispatch time on events that turn out to be nothing. Nationally, the International Association of Chiefs of Police has estimated that up to 98 percent of security alarm activations are false, at a combined national cost near $1.8 billion in wasted response resources. A facility-level estimate of hours spent investigating unverified alerts, multiplied by loaded labor cost, is the third input, and it shrinks directly as verified-only alerting replaces raw motion triggers.
IACP false-alarm data, facility alert-investigation logs
Threats or acts of violence during theft events rose 17 percent year over year according to the 2025 NRF/LPRC survey. Earlier detection of a pre-incident behavior, staging, casing, or an escalating confrontation, compresses the time between the behavior and a staff or law-enforcement response, which is the same mechanism that reduces both injury risk to employees and a facility's exposure in a foreseeability-based premises liability claim.
NRF/LPRC violence trend, incident response-time logs
The model is deliberately conservative about Input 2. Rather than inventing a current shrink-dollar figure the industry itself has stopped publishing, it anchors to the last full, explicitly dated NRF figure and treats anything more recent as a placeholder to be replaced with a facility's own data as soon as it exists. That is the entire point of the framework: every input traces to either a named, dated, external source or the organization's own detection log, and nothing in the model depends on a number a board member cannot verify by opening the cited report.
Comparing the Three Loss-Signal Sources Retailers Actually Have
Retail security leaders are not choosing between good data and bad data. They are choosing among three imperfect sources, each with a different latency, coverage, and audit trail, and understanding the tradeoffs is what makes the four-input model above defensible rather than arbitrary.
| Loss Signal Source | What It Actually Captures | Latency | Audit Defensibility | Primary Limitation |
|---|---|---|---|---|
| Law enforcement reported data (FBI UCR: SRS and NIBRS) | Incidents formally reported to police and classified into a federal reporting system. | Months to over a year; published on annual or quarterly federal release cycles. | High per reported incident, but the two federal systems produce a 93-point gap on the same national trend. | Excludes the majority of retail theft; year-over-year change is distorted by which agencies newly join a reporting system. |
| Retailer self-reported survey data (NRF/LPRC) | Retailer-estimated incident and violence trends aggregated across a limited respondent panel. | Annual survey cycle; published roughly two months after the summer field period closes. | A credible directional industry figure, but not facility-specific and not independently audited incident by incident. | Reflects 70 respondent companies representing about a quarter of U.S. retail sales, not a facility's own experience. |
| AI-verified detection data (facility-level) | Every classified detection event at a specific facility, with timestamp, camera location, and confidence score. | Real time, continuous, independent of any decision to file a report. | Facility-specific, timestamped, and reviewable by an internal or external auditor at any time. | Reflects only what deployed cameras cover; a facility must extend coverage to the zones where pre-incident behavior actually occurs. |
The comparison is not an argument that AI-verified data replaces law enforcement statistics or industry surveys. National and industry data remain the right tool for benchmarking a facility against its peers and for understanding sector-wide trend direction. What facility-level verified data adds is the one thing the other two sources structurally cannot provide: a continuous, facility-specific, independently reviewable record that does not depend on a retailer's decision to call police or a federal agency's decision to change its reporting methodology mid-decade.
What Loss Prevention Leaders Should Ask Before Building Next Year's Business Case
Security, asset protection, and finance leaders preparing a 2026 or 2027 retail security budget can convert this report into a working diligence checklist:
- Which specific source is behind every statistic in the proposal, the FBI's SRS, the FBI's NIBRS, an NRF survey year, or a vendor's own estimate, and is that source explicitly dated in the document?
- Does the proposal rely on a national or industry-average shrink rate, or does it build toward a facility-specific baseline using the organization's own detection and incident-resolution data?
- How much loss prevention and store-management time currently goes toward investigating unverified alerts, and what would a confidence-scored, verified-only alert stream save in labor hours?
- Can the vendor confirm in writing that its detection architecture performs no facial recognition and does not build customer identity profiles, given the legal exposure retail faces under state biometric privacy statutes?
- Does the proposed detection coverage extend to the pre-incident zones, parking lots, loading docks, and after-hours entry points, where organized retail crime staging typically begins, not just point-of-sale areas?
Retailers evaluating how the Combating Organized Retail Crime Act will change federal coordination should treat that legislative track as separate from, not a substitute for, a facility-level detection strategy: legislation changes how aggregated cases get prosecuted; it does not generate the timestamped, facility-specific evidence a prosecutor or an insurer needs. For the general economic modeling approach behind this report, see IntelliSee's Four-Variable ROI Framework and the companion methodology paper on avoided-incident attribution, which addresses the counterfactual question this report assumes: how to credit a prevented incident that, by definition, generates no police report at all.
Frequently Asked Questions
Why do retail shrink and theft statistics disagree so much from source to source?
Three different measurement systems are answering three different questions. The FBI's Summary Reporting System and National Incident-Based Reporting System only capture incidents formally reported to police, and the two systems disagree by 93 percentage points on the same 2019-2023 shoplifting trend because of changes in which agencies report through which system. The National Retail Federation's Impact of Retail Theft & Violence survey captures what a panel of 70 retail companies estimate and choose to disclose. Neither source is measuring what actually happens inside a specific store.
How many retail theft incidents actually get reported to law enforcement?
According to the National Retail Federation and Loss Prevention Research Council's 2025 survey, 64 percent of retailers report less than half of their store-related theft incidents to police, most commonly because they do not expect a law enforcement response. That means any national statistic built on police-reported data is describing a minority subset of total retail theft, and the size of that subset can shift year to year based on retailer confidence in local prosecution, not the true incident volume.
What is the difference between the FBI's SRS and NIBRS shoplifting data, and why does it matter for a security budget?
The Summary Reporting System is the FBI's older monthly-count format; the National Incident-Based Reporting System is the newer incident-level format replacing it. For the same 2019-2023 period, SRS data shows shoplifting incidents roughly flat, while NIBRS data shows the shoplifting rate 93 percent higher, a gap the Council on Criminal Justice attributes largely to more large-city police departments joining NIBRS reporting during that window, not necessarily a matching rise in actual crime. A security budget request that cites the larger figure without noting this composition effect is vulnerable to being challenged in a finance committee review.
Does AI video detection replace the need for national or industry shrink data?
No. National and industry data remain useful for benchmarking a facility against sector-wide trends and for understanding directional change. AI-verified detection data adds what those sources cannot provide on their own: a continuous, facility-specific, timestamped record that does not depend on a retailer's decision to file a police report or a federal agency's reporting methodology.
What does AI-verified detection data actually capture at a retail facility?
A computer vision platform processing existing camera feeds classifies objects and behaviors, a person in a restricted zone, a vehicle at a loading dock after hours, a person loitering near an entrance, and returns a bounding box with a confidence score and a timestamp in real time. That detection log exists independently of whether the incident is later reported to police, which is what makes it useful as an internal, audit-ready baseline.
Does IntelliSee's AI platform use facial recognition or store video in retail deployments?
No. IntelliSee performs no facial recognition, stores no video, and does not build a customer or employee identity profile. The platform classifies an event and returns a bounding box with a confidence score without identifying who the person is or retaining the underlying footage, an architecture chosen specifically to avoid the legal exposure retail environments face under state biometric privacy statutes.
How should a loss prevention leader build a defensible ROI case when the national statistics disagree?
Build the model on inputs that trace to a named, dated source or to the organization's own detection log rather than a single blended national figure. The four-input framework in this report, verified capture rate delta, a dated shrink-to-dollar translation, response and labor cost avoidance, and violence and liability deterrence value, lets each input be checked independently, which is what allows the total to survive a finance committee's scrutiny. Retailers can request a risk assessment to begin building a facility-specific baseline.
Conclusion: The Baseline Is the Business Case
Retail loss prevention does not have a persuasion problem. Boards already know theft and violence are rising; the NRF/LPRC survey, the FBI's own contested data, and daily news coverage all point the same direction. What retail loss prevention has is a measurement problem, and that problem sits directly underneath every ROI model built to justify a security technology purchase. A finance committee does not reject a security investment because the threat is not real. It rejects the investment when the number behind the threat cannot survive being checked against its own source.
The fix is not a better national statistic. National statistics will keep disagreeing with each other as long as reporting behavior, agency participation, and survey methodology keep changing, which is to say indefinitely. The fix is building the ROI model on a baseline the organization actually controls: a facility-specific, timestamped, confidence-scored detection record that exists whether or not a report ever reaches a police department, and that a facility's own finance team can audit without waiting for a federal agency to reconcile two datasets that, as of this report, still do not agree.
IntelliSee provides no-cost risk assessments that map detection coverage across sales floor, perimeter, parking, and loading-dock zones for retail facilities, building the facility-specific baseline this report's ROI model depends on. Contact our team to scope an assessment for your locations.
Continue the research: AI Retail Security: The 2026 Sector Playbook | The Combating Organized Retail Crime Act: 2026 Standards-Compliance Briefing | The Four-Variable ROI Framework for AI Physical Security | Avoided-Incident Attribution: A 2026 ROI Framework | Weapons Screening vs. AI Video Weapon Detection: The 2026 Market Analysis
Related reading from the IntelliSee blog: 98% of Security Camera Alarms Are False, and What That's Actually Costing You, a deeper look at the false-alarm cost input referenced in this report's response and labor cost avoidance model.
Primary Sources and Citations
- National Retail Federation and Loss Prevention Research Council, The Impact of Retail Theft & Violence 2025, sponsored by Sensormatic Solutions (October 28, 2025).
- National Retail Federation, press release, "New Study Finds Retailers Continue to Contend with Rising Levels of Theft & Violence" (October 28, 2025).
- National Retail Federation, National Retail Security Survey, fiscal year 2022 shrink data (1.6 percent of sales, approximately $112.1 billion), published 2023.
- Council on Criminal Justice, Between the Aisles: A Closer Look at Shoplifting Trends, Ernesto Lopez, lead author (November 2024, updated October 2025), analyzing FBI Uniform Crime Reporting Program Summary Reporting System and National Incident-Based Reporting System data.
- Federal Bureau of Investigation, Uniform Crime Reporting Program, Summary Reporting System and National Incident-Based Reporting System, via FBI Crime Data Explorer.
- U.S. Bureau of Labor Statistics, Survey of Occupational Injuries and Illnesses, Employer-Reported Workplace Injuries and Illnesses, 2023-2024 news release (January 22, 2026), retail trade sector incidence-rate trend context.
- International Association of Chiefs of Police, national security alarm false-activation rate and associated law enforcement response cost estimates.
More intelligence like this
New IntelliSee research drops monthly at most. Subscribe and get the next sector playbook, technology briefing, or threat intelligence report in your inbox the day it ships.
Request a Risk Assessment
Talk to an IntelliSee security specialist. No sales pitch — a structured conversation about your environment, your threat profile, and whether computer vision is the right fit.
Request a Risk Assessment