Weapons Screening vs. AI Video Weapon Detection: The 2026 Market Analysis of the Modality Buying Decision

Two categories of weapons detection systems now compete for the same line item on a security director's budget, and they are routinely confused with one another. The first is concealed-weapons screening: walk-through pedestals that a person passes through at a controlled entrance, including the AI-assisted free-flow systems that have largely displaced the airport-style metal detector. The second is AI video weapon detection: software that runs on a facility's existing camera network and flags a brandished firearm in the open. Both promise to find guns. They do fundamentally different things, at different points in an attack, at costs that differ by more than an order of magnitude. Choosing between them, or sequencing them, is one of the highest-consequence procurement decisions a risk leader makes, and the market does not make the distinction easy.

This Market Analysis separates the two modalities on the variables that actually move a buying decision: where in the timeline each one detects a threat, what it costs to acquire and to staff, how often it cries wolf, and what the public record now says about the claims vendors have made. The evidence base is deliberately drawn from primary sources. A 2024 Federal Trade Commission enforcement action established what one screening vendor could no longer claim. A 2025 Maryland Center for School Safety study, conducted under legislative mandate, put government-collected cost and staffing numbers on the table for the first time. Federal performance standards define what "detection" even means. Read together, they reframe the screening-versus-detection question from a feature comparison into a question of operating model and total cost.

The two modalities detect threats at opposite ends of the attack timeline

The single most important difference between weapons screening and AI video detection is not accuracy or cost, it is when each one fires. A walk-through screening pedestal is a checkpoint. It interdicts at a defined chokepoint before a person enters a protected space, which is its entire value proposition: stop the weapon at the door. That model assumes three things hold true. There is a controlled entrance everyone funnels through, there is staff stationed at it to resolve every alarm, and the threat actually walks through the front door rather than entering through a loading dock, a propped fire exit, or a parking structure.

AI video weapon detection operates on a different premise. It watches the camera feeds a facility already runs and raises an alert the moment a firearm becomes visible anywhere in frame, whether that is a lobby, a corridor, a parking lot, or a perimeter approach. It does not interdict; it shortens the time between a weapon appearing and a human knowing about it. The tradeoff is symmetrical to screening's. Video detection sees a gun only once it is brandished or otherwise visible to a camera, so it cannot find a weapon concealed in a waistband or a bag. What it can do is cover the ninety percent of a property that no pedestal will ever stand at, and compress the response window during the pre-attack and early-attack phases that perimeter and interior cameras observe. This is the architectural shift from passive recording to real-time detection on existing video infrastructure.

The Maryland Center for School Safety, in the interim report it was required to produce under House Bill 782, drew exactly this line in its own taxonomy. It classified walk-through detectors, with and without AI, under "metal detection," and classified camera-based AI as "visual detection," noting plainly that "objects (e.g., firearms) that are not visible to the camera cannot be identified by video detection technologies." The same report noted that walk-through systems detect concealed metallic weapons but depend heavily on sensitivity settings, "particularly smaller knives." That dependence on configuration is one of several documented failure modes that any detection program has to plan around. Neither modality is a superset of the other. They detect different threats, in different places, at different moments. Treating them as substitutes is the first analytical error a buyer can make.

Live CAM 04

The federal record reset what a screening vendor can claim

For most of the past decade, the marketing of AI-assisted weapons screening ran well ahead of its measured performance. In November 2024 the Federal Trade Commission ended that gap for the category's most prominent vendor. The FTC took action against Evolv Technologies over allegations that the company made false claims about the extent to which its AI-powered screening system could detect weapons and ignore harmless items, "including in school settings." According to the FTC's complaint, Evolv's Express scanners were located in over 800 schools across 40 states, with school systems making up half the company's business.

The specifics in the FTC's filing are the most useful part for a buyer, because they describe failure modes rather than slogans. The Commission alleged that Evolv misrepresented that its system would detect all weapons, would ignore harmless personal items without requiring people to empty their pockets or bags, would detect weapons more accurately and faster than metal detectors, would reduce false alarm rates, and would cut labor costs by 70 percent compared to metal detectors. The complaint described scanners that "failed in several instances to detect weapons in schools while flagging harmless personal items," such as laptops, binders, and water bottles, and cited a seven-inch knife that was not detected and was used to stab a student in October 2022. After that incident, the complaint states, school officials increased the system's sensitivity, "prompting a 50% false alarm rate."

The proposed settlement order, approved by a 5-0 Commission vote, prohibits Evolv from making misrepresentations about its products' detection ability, accuracy and false-alarm rates "including in comparison to the use of traditional metal detectors," screening speed, labor costs, and the results of any testing. It also required the company to notify certain K-12 customers that they could cancel contracts signed between April 1, 2022 and June 30, 2023. The order does not establish that the technology does not work. It establishes that specific performance and cost-savings claims could no longer be made without support. For a procurement officer, that is the more durable lesson: the burden of proof on a screening claim now sits with the vendor, and "tested" is a word that requires a citation.

The cost gap is an order of magnitude, and the hidden cost is staffing

The Maryland HB 782 study is the most useful public cost reference available, because the numbers were collected by a state agency directly from the 24 local education agencies that operate the systems, not from vendor price sheets. The reported acquisition costs span a remarkable range. A handheld wand runs about $210 per unit. A standard walk-through weapons detector averaged $17,500 per device. A walk-through detector with integrated AI averaged $100,000 per device. AI video analytics averaged roughly $3,000 per school in initial software and installation, with ongoing licensing of $100 to $300 per camera per year. A gunshot detector ran about $1,100 per unit. Multiple devices are typically needed per entrance to manage throughput, so a multi-lane screening deployment scales into the hundreds of thousands of dollars before anyone is hired to operate it.

And operating it is where the real cost lives. The Maryland report is unambiguous: "the requirement for dedicated staff is consistently reported as the most significant and often prohibitive operational expense, serving as the primary barrier to WDS adoption for the majority of LEAs." A walk-through lane requires a minimum of two staff to operate and conduct secondary inspections, with three to four per entrance described as ideal because most arrival entrances need at least two lanes. The study cited one agency's figure of roughly $82,000 for a full-time employee. That is the line item that compounds: a pedestal is a one-time purchase, but the people standing next to it are a recurring annual cost that does not end when the warranty does.

This is also where the Evolv labor claim matters. The 70-percent labor-reduction comparison the FTC barred from Evolv's marketing was, in effect, a claim that the AI pedestal needed far fewer people than a metal detector. The Maryland staffing data, collected independently, shows walk-through systems still requiring two to four staff per entrance regardless of the AI label. AI video analytics, by contrast, generates an alert that "human verification" resolves, work the report describes as typically performed by existing safety-and-security staff who already monitor the system, rather than by new personnel posted at a portal. The verification burden is real for both modalities, because no system is free of false alerts, but the staffing geometry is different: screening demands people at the point of entry during every arrival window, while video verification distributes across staff already watching the network.

Three detection modalities, side by side

Government-collected cost and operating data from the Maryland HB 782 study, mapped against where each modality detects a threat. Figures are LEA-reported 2025 acquisition costs.

False alarms are the operating reality, not an edge case

Every detection technology in this market generates alerts for things that are not weapons, and how a system handles that fact determines whether it is workable in practice. The Maryland study states the principle directly: "None of the systems discussed in this report are able to identify a weapon with 100% accuracy. Consequently, every alert generated by a device or system requires human verification." That sentence applies equally to a pedestal and to a camera. The difference is in the cost and friction of each verification.

On a screening pedestal, a nuisance alarm halts a person at the entrance and triggers a secondary search, hand-wanding or bag inspection, conducted by staff while a queue forms behind them. The FTC's Evolv complaint captured how this dynamic degrades: raising sensitivity to catch more knives drove a 50-percent false-alarm rate at one school, and the agency noted that mitigating false positives pushed the system "more like traditional lower-cost metal detectors," with conveyor belts and hand-diversion of harmless items. Federal performance standards put numbers to the tolerance: NIJ Standard-0601.02, the U.S. reference standard for walk-through metal detectors in concealed-weapon detection, specifies detection-performance and nuisance-alarm testing, and European performance requirements that reference it target a nuisance-alarm rate below 5 percent. The gap between a 5-percent design target and a 50-percent in-the-field rate is the gap between a published standard and an operational reality shaped by sensitivity settings and the threat you are trying to catch.

On a camera, a false alert is a notification to a verifier rather than a person halted in a doorway. The cost is an operator's attention and the risk of alert fatigue, a genuine failure mode, but it does not stop the flow of people through an entrance or require a physical search. This is why the verification layer, and whether it is automated, staffed in-house, or sold as a monitoring subscription, belongs at the center of any modality comparison. IntelliSee's platform routes a visual detection through a verification step and an alert pipeline that completes within seconds, on existing camera infrastructure, with no facial recognition, no stored video, and no personal data collected. The relevant procurement question is not "does it false-alarm," because everything does, but "what does each false alarm cost in staff time and friction, and who absorbs it."

National research finds deterrence perceptions, not proven incident reduction

A market analysis owes buyers the uncomfortable finding alongside the favorable one. The Maryland report, summarizing national research, states that studies "found no consistent reduction in threats, fights, or victimization" from weapons screening, "despite some reports of lower rates of students bringing weapons," and that visible security measures "can sometimes negatively affect students' sense of safety and may actually increase fear." It cites peer-reviewed and institutional work to that effect, including a 2011 review in the Journal of School Health and analyses from WestEd and the National Association of School Psychologists. At the same time, seven of the surveyed Maryland agencies reported increased perceptions of safety and a "huge deterrent effect," and several reported "reduced fear" when systems were in use.

Both findings can be true at once, and the reconciliation is instructive. Screening produces a visible deterrent and a measurable interdiction at the door it stands at, but the academic literature has not isolated a consistent reduction in overall incidents, partly because screening is almost always deployed inside a broader security posture that makes its independent effect hard to measure. The policy implication is the one the Maryland agencies themselves reached: weapons detection is "a component within a multilayered system," not a standalone solution, and foundational measures, functioning doors and locks, secured vestibules, working communications, were "often prioritized over and implemented before weapon detection technologies." The buying calculus, in other words, is not screening-or-video. It is where each modality fits in a layered architecture, and what each layer is being asked to do.

A buyer's framework for weapons detection systems: match the modality to the threat geometry

The choice resolves cleanly once a buyer stops asking which technology is better and starts asking which threat geometry they are defending. A facility with a single, genuinely controlled entrance that everyone passes through, a courthouse, a stadium gate, a secured clinic intake, has a geometry that screening was built for, and the question becomes whether the staffing model is sustainable. A facility with many entrances, large open areas, parking structures, and a perimeter, a campus, a hospital complex, a manufacturing site, a multi-building school district, has a geometry that no pedestal can cover, and AI video detection on the existing camera network addresses the space between the doors. Most real properties are the second kind, which is why coverage, not point accuracy, is usually the binding constraint.

Throughput is the other hard limit on screening. AI free-flow pedestals are marketed at high throughput, with one prominent vendor citing roughly 3,600 people per hour for a single lane and screening, by its 2025 account, more people per day than the TSA. That throughput is real and is why screening dominates high-volume event ingress. But throughput is a function of how many lanes are open and staffed, and each lane carries the two-to-four-person staffing cost the Maryland study documented. The economics that make screening compelling at a stadium gate for three hours on game day are the same economics that make it prohibitive at twelve school entrances every weekday morning. This is precisely the bind that drives the K-12 detection conversation, where dozens of doors and tight budgets collide. Video detection has no throughput limit because it does not gate movement; it observes it.

The honest synthesis is that these modalities are complements more often than substitutes. A stadium can run free-flow screening at the gate and AI video across the concourse and parking structure. A hospital can screen at the emergency department intake and run video on the dozens of other doors no one screens. The error is to buy one modality, believe it has closed the gap, and discover during an incident that the threat used the geometry the chosen modality could not see. The procurement task is to map the threat geometry first, then assign each layer to the modality whose detection model fits it, and to count the staffing cost honestly before signing. A disciplined proof-of-concept methodology turns that mapping into a defensible buying decision, and understanding how detection-to-response actually works across a facility keeps the comparison grounded in operations rather than slogans.