Guidelines for Developing AI in Physical Security

Recent breakthroughs in artificial intelligence (AI) don’t automatically translate into breakthrough physical security products. Yet some companies are using AI in innovative ways that far exceed the performance of previous generations. The key is clarity—vendors must be able to explain how their products achieve these results, so customers understand the real value behind the claims.

Q:  I’ve been seeing lots of AI video analytics products. Are they all basically the same now?

A: It can look that way from the marketing. In reality, there are significant differences, though they’re not always explained well.

Transformative Improvement

Recent breakthroughs in artificial intelligence (AI) are driving transformative changes in security operations capabilities, shifting the posture from reactive to proactive, preemptive, and even preventive. While nearly all AI-enabled improvements offer value, only a few are truly transformative. Simply adding AI models to an application doesn’t magically make it exceptional.

Creating breakthrough capabilities requires deep know-how across three domains: AI engineering, security technology, and security operations. That’s why it is so important for companies to clearly explain how they apply AI and other engineering expertise to achieve the results they claim.

AI Models Are Not Magic

When people hear “AI,” they often picture a finished, general-purpose capability—a plug-and-play module that works out of the box and “thinks” for you. AI models are just that: models—mathematical structures trained for narrow tasks. A single model rarely delivers the functionality required in a physical security application. Instead, several models are combined into application pipelines built for specific use cases.

It is crucial to understand how a company utilizes AI in the engineering of its product offerings. For many of the breakthrough AI-enabled physical security devices and applications, the innovative aspects of their engineering are significant enough to achieve patent registration. So that is one aspect of a product to ask about. Another important focus is how the company has worked to differentiate its functionality and performance from that of other products in the same category. 

Functionality needs can differ significantly from one business category and facility type to another. 

An Example: Actuate’s “Slicing” and My Mistaken AI Model Assumption

Actuate, for example, provides a cloud-based AI video alarm monitoring platform for central stations and remote guarding. It integrates with existing camera systems to deliver real-time threat detection and actionable alerts, significantly reducing false positives. Some sites use on-camera object-based motion detection to filter out false alarms. For those sites, Actuate’s highly trained AI models can eliminate 50–80% of the alarms that remain. Where only pixel-change motion detection is used, the AI can filter out 85–90% of the alarm traffic—sometimes as much as 95%—while still catching virtually all true alarms.

In a recent project, I assumed Actuate’s “slicing” feature was powered by a Vision Transformer (ViT) model. The way it intelligently divides high-resolution images into smaller segments for parallel processing looked, at first glance, like ViT mechanics to me.

When I asked Actuate, their response corrected my assumption. They explained that although they had experimented with ViTs, the models were too data-hungry and computationally heavy for the sub-second, frame-by-frame analysis required in security video monitoring scenarios. Instead, Actuate relies on its custom-trained models trained using over 1.5 million images of real security scenarios and layers of machine learning processes.

They compared their detection pipeline to overlapping slices of Swiss cheese: if a detection slips through one layer, the next one catches it. Some of these layers use Long Short-Term Memory (LSTM) networks to filter out repeated false positives. Actuate is also exploring Vision-Language Models (VLMs) and Large Language Models (LLMs) for complementary use cases, though their core work remains focused on video.

This level of explanation—about what models are used, what is not used, and why—demonstrates a kind of product transparency that’s rare but increasingly important. And in the current era of rapid technological advancements, a company’s technology roadmap and strategic direction are also essential factors.

How AI Has Brought Speed and Accuracy to Acoustic Threat Detection

I was recently introduced to one clear example of how AI can turn previously unusable technical knowledge into a real-time security response cutting incident response time from minutes to seconds. Acoem’s Acoustic Threat Detection (ATD) builds on decades of acoustics expertise in identifying the sound signatures of gunfire—both the muzzle blast and the ballistic shockwave. In the past, this knowledge couldn’t be applied to real-time security operations without complex, multi-sensor arrays and infrastructure, which limited its use to specialized deployments. Now, with AI-enabled parallel processing, the same knowledge can be harnessed instantly to deliver both speed and accuracy that were previously out of reach.

Acoem’s ATD-300 sensor embeds a machine learning engine directly within the hardware, using AI-driven edge processing to detect and localize gunfire in real time while minimizing latency and reducing reliance on centralized infrastructure. With 360-degree acoustic coverage, a single sensor can provide broad monitoring from one installation point, often reducing the number of units required. The system delivers operator alerts in under three seconds and can integrate with video management platforms to automatically slew PTZ cameras toward the incident. I’m looking forward to learning more about the acoustic threat detection technology at GSX 2025.

Why These Explanations Matter

AI for physical security is not general-purpose AI—it’s an assembly of narrow, specialized models working together. A company’s ability to explain how its AI delivers performance is a key differentiator. Vague “AI-powered” claims are no longer enough. Buyers should expect vendors to show what’s under the hood, and those that can do so demonstrate both credibility and maturity.

Explaining AI Is Part of the Innovation

Both Actuate’s “slicing” and Acoem’s acoustic threat detection show why clarity matters: the breakthrough isn’t “AI” by itself, but how AI is applied to make specific technical knowledge practical in real time. When companies can explain that process—what models they use, what they don’t, and why—it separates genuine engineering achievement from marketing hype. That specificity not only builds trust, it also helps buyers see the true value of a solution in the context of their own security operations.