AI-Powered Video Analytics: Separating Real Capability From Marketing Hype
By Gina Stuelke, CEO of Kenton Brothers
Every security camera vendor now claims some form of “AI-powered” analytics. Scroll through enough product pages and you’ll see promises of predictive crime prevention, instant threat detection, and systems that seem to understand intent, not just motion. Some of this is real. A lot of it is marketing language stretched well past what the underlying technology can actually do.
For buyers trying to evaluate systems, the gap between what’s demoed and what performs reliably in a real facility, at scale, in varied lighting and weather, is often significant. Here’s a practical breakdown of what’s genuinely working today, what’s overstated, and what questions separate a real capability from a slide-deck feature.
What Actually Works Well Today
- Object and motion classification. Distinguishing a person from a vehicle from an animal, and filtering out irrelevant motion (blowing leaves, shadows, weather), is mature technology at this point. This alone meaningfully reduces false alarms compared to older motion-only systems.
- License plate recognition (LPR). Reading and logging plates for access control or investigation purposes is a well-established, reliable capability, especially in controlled conditions like a gated entrance with consistent lighting and vehicle speed.
- Loitering and dwell-time detection. Flagging when a person or vehicle remains in a defined area longer than expected is a solid, practical use of analytics — useful for perimeter monitoring or after-hours alerts.
- Line-crossing and zone-intrusion alerts. Defining a virtual boundary and triggering an alert when it’s crossed is dependable and widely deployed, particularly for perimeter security.
- Facial recognition for access control (in constrained settings). For matching against a known, enrolled set of faces at a controlled access point — not for open-ended crowd identification — accuracy is generally strong under good lighting and camera angle.
Where Hype Outpaces Reality
- “Predicting” crime or intent. No current system can reliably infer that someone is about to commit a crime based on behavior alone. What’s being sold as “predictive” is usually pattern-based anomaly detection (something looks statistically unusual) — which is useful, but it is not prediction, and vendors that imply otherwise are overselling.
- Crowd-scale facial recognition accuracy. Performance drops meaningfully in uncontrolled environments — poor lighting, partial occlusion, extreme angles, or large crowds. Claims of near-perfect accuracy in these conditions should be treated skeptically and tested against your actual environment, not a vendor’s demo reel.
- “Understanding” behavior or emotion. Analytics marketed as detecting aggression, distress, or emotional state from video are built on much shakier ground than object detection. These systems have high false-positive rates and raise real questions about reliability, especially if alerts are used to justify intervention.
- Universal, out-of-the-box accuracy. Many demos are run on curated footage. Real-world performance depends heavily on camera placement, resolution, lighting consistency, and how well the system has been tuned to the specific site. A capability that performs well in a demo can perform far worse in your parking lot at dusk.
Questions to Ask Vendors
When evaluating a system, a few direct questions tend to cut through the marketing quickly:
- What’s the false positive rate in real deployments, not lab conditions, and can you provide reference customers with similar environments to mine?
- Does this run on-device (edge) or in the cloud, and what does that mean for latency, bandwidth, and data privacy?
- How much site-specific tuning is required before the system performs at the accuracy being advertised?
- What happens in degraded conditions, night, rain, fog, glare, and is performance data available for those conditions specifically?
- How is the training data sourced, and has the system been tested for bias across different demographics, lighting conditions on skin tones, and clothing types?
- What’s the actual role of a human in the loop? A system that flags anomalies for a human to review is a very different (and generally more trustworthy) product than one that claims to make autonomous determinations.
The Practical Takeaway
AI video analytics is a genuinely useful category. It reduces false alarms, speeds up investigation, and extends what a limited security staff can monitor effectively. But the honest framing is “smart filtering and pattern flagging,” not “AI that understands and predicts human behavior.” Buyers who ask pointed questions about real-world performance, rather than relying on demo footage, are far more likely to end up with a system that performs the way they expect once it’s actually running in their facility.
Kenton Brothers Systems for Security is a WBE-certified commercial security integrator headquartered in Kansas City, Missouri. Our services include commercial security solutions for video surveillance, access control, locking hardware, intrusion detection, mass notification, remote services, and specialty products including robotics, video analytics, and metal detection systems.
To learn more, give us a call.
We have been protecting Kansas City businesses since 1897. We’ve watched security technology evolve from simple deadbolts and night watchmen to AI-driven analytics and remote monitoring platforms. And in all that time, one truth has never changed: a security system that only records what happened is not doing its most important job.
