The traditional video surveillance model is reactive by design. Companies treat it like an insurance policy, installing cameras and paying for monitoring in the hope they’ll never be needed. When problems do arise, security teams spend hours reviewing footage to understand what already occurred.

That model is starting to break, and ironically, not because organizations suddenly have fewer cameras, but because they have far too many. In campuses, hospitals, retail chains, logistics hubs, and city centers, the modern reality is camera-dense, multi-site sprawl. 

But here’s the problem: there’s not enough human attention to actually make sense of it all in real time. The video wall era created three issues that keep getting worse: operators burn out from staring at screens all day, manual triage can’t keep up with the volume, and systems cry wolf so often that security teams learn to ignore alerts.

In a recent interview, Edan Sorski, Head of Cloud at Lumana, an AI video security platform, explained why he believes the next era of video surveillance will be about agentic workflows: systems that can monitor, verify, and trigger actions, pushing humans up the stack from constant watchstanding to oversight and exception handling. 

As AI-powered surveillance becomes an increasingly competitive space, companies are racing to move beyond detection-only systems.

Why Security Operations Cannot Scale the Old Way

Retail offers a clear example of why traditional surveillance is failing. Shrink losses reached $112.1 billion in 2022, putting enormous pressure on loss prevention teams. The problem in 2026 is that detecting theft isn’t enough. Teams need to verify what’s happening and respond before losses occur, not hours or days later when someone finally reviews the footage. Systems built around recording everything for post-incident investigation can’t stop theft in progress. More cameras create more events to monitor, more false alarms to sort through, and exponentially more work for human operators trying to make sense of it all. The old approach of people staring at video walls doesn’t work at this scale.

This problem is fueling market growth. The U.S. video surveillance market is expected to hit $37 billion by 2030 as companies move away from just investigating incidents after they happen. Organizations now expect these platforms to work like any other enterprise software: fewer false alerts, better automation, smoother integration with their existing systems.

In this context, surveillance that mainly records and stores footage is starting to look outdated.

The Solution to Security Teams Ignoring Their Own Cameras

Traditional video analytics work like a basic alarm system. They detect motion, classify what they see, and send an alert. Simple enough. But the problem isn’t the technology itself. It’s the messy real world these systems operate in. Changing light conditions, weather, shadows, animals wandering through frames, reflections off glass… all of these trigger alerts. Organizations end up with false alarm rates that can hit 90%. 

When that happens, operators stop trusting the system. They start ignoring alerts altogether. What you get is massive alert fatigue. Security teams become numb to notifications, and video goes back to being something you only look at after an incident, not a tool that helps prevent one.

Lumana built VIA-1 specifically to solve this. Instead of being trained in a lab under perfect conditions, the system learns from the actual environment where each camera is installed. A loading dock operates differently from a corporate lobby or a school parking lot. VIA-1 adapts to those differences. The goal is getting cameras to understand what normal looks like for their specific location so they can flag actual threats instead of routine activity.

Four Agents Redefining What Security Teams Actually Do

Building on this foundation, Lumana has deployed four AI agents positioned as a digital workforce handling distinct stages of security operations:

• Monitoring Agent: Watches cameras continuously and flags specific activities that require human attention.
• Response Agent: Notifies stakeholders and can trigger interventions like lockdowns, loudspeakers, alarms, or emergency dispatch.
• Investigation Agent: Searches footage using vision-language intelligence based on natural language queries such as “Find a person in a red jacket” or specific visual attributes.
• Insight Agent: Converts video into structured operational data for tracking trends like foot traffic patterns, safety violations, and staffing efficiency across multiple sites.

These agents represent a vision for how security operations might work in the coming years. Machines handle the constant monitoring, filter through alerts, and connect related events. Humans oversee the system, step in when something unusual happens, and make the decisions that matter. The idea isn’t replacing security teams but amplifying what they can do. Smaller teams can manage workloads that used to require full security operations centers.

Lumana’s growing network of 50,000 cameras, serving organizations like Salesforce, Meta, and McDonald’s, provides early validation of this approach. According to the company, customers report a more than 90% reduction in false alerts, with security teams describing a fundamental change in how they interact with surveillance infrastructure.

Where Agentic Surveillance Lands First

Think camera-dense environments across multiple sites. More cameras just meant more noise. In sectors like education, government, retail, hospitality, healthcare, manufacturing, and logistics, organizations often run dozens or hundreds of locations with thousands of cameras between them. The problem isn’t that they don’t have enough video. It’s that they have way too much for any human team to actually watch, sort through, and investigate.

Schools want to know what’s happening right now so they can respond fast. Hospitals and hotels are juggling safety issues, access control, and crowd management. In manufacturing and logistics, teams need eyes on the floor to spot where things are slowing down and stop accidents before they happen.

These workflows are designed for exactly that kind of reality. They separate useful signals from background noise. Flag only what needs attention. Create consistent response protocols across locations, even when teams are understaffed.

Local Intelligence, Central Control

Agentic security has a core technical problem. The system needs to react in seconds, but it also needs to learn and improve across thousands of cameras in hundreds of different places. You can’t just put everything in the cloud because that creates lag. Real-time response stops being real-time. But if you keep everything local, the systems can’t talk to each other. They’re hard to update and they miss patterns that only show up when you’re looking across the whole organization.

Lumana’s answer is delivering a distributed, hybrid-cloud architecture that splits the work. Local processing handles immediate needs. When something happens, the camera detects it, verifies it, and triggers a response right there. No waiting for the cloud. The system keeps working even if internet goes down. Cloud processing tackles everything else. It manages oversight and coordination. Security teams use it to roll out policy changes, refine how the system detects threats, and identify patterns that only become visible when you’re looking at data from multiple locations at once.

This changes what video surveillance actually does. Detection and verification happen locally. Responses fire immediately on-site. But the intelligence behind those responses, the rules governing what happens when, the improvements over time, all of that flows from a central platform that sees across the entire operation. You get speed at the edge and consistency from the center. 

What used to be a passive recording system becomes something that stops problems before they escalate, learning and adapting as it runs.

Why Automated Surveillance Needs Human Guardrails

When surveillance moves from sending alerts to taking action, everything changes. A false alert is annoying. A false action can be a real problem. Locking doors when you shouldn’t, blasting a message over loudspeakers, setting off alarms, calling emergency services. These create consequences. That’s the tension at the heart of this approach. 

Automation brings speed and scale, but people still need to own the outcomes. The system lets organizations decide how much autonomy to give it at each step. A warehouse might allow automatic responses at night when nobody’s there. A school might require a security officer to sign off before anything happens. Hospitals might need extra controls to avoid disrupting patient care. 

Then there’s privacy, which adds a whole other layer. AI surveillance in schools brings up questions about monitoring students and keeping their data. Healthcare has patient privacy laws to deal with. Public spaces get into civil liberties debates that look different depending on where you are.

The real competitive edge in this space won’t just be about accuracy. It’ll be governance. Can you show why the system did what it did? Can you limit what it’s allowed to do? Do you have clear processes for who decides what and when? Security leaders, legal teams, regulators all need answers to those questions. Agentic can’t just mean automated. It has to mean accountable.

Conclusion: Video as a Decision Layer

The old surveillance model measured success in coverage and storage. The agentic model changes the scoreboard to operational outcomes: fewer false alerts, faster response times, quicker investigations, and measurable safety improvements. But this evolution from passive recording to active intervention introduces serious questions about governance that organizations must address through careful controls on system autonomy.

In the next phase of this market, winners will be defined not by who stores the most footage but by who turns footage into outcomes while maintaining accountability for how those outcomes are reached.

“The next evolution in video surveillance is agentic. It is not just detecting an event, it is understanding context, validating what is happening, and driving the right next step automatically.”

Edan Sorski, Head of Cloud at Lumana

Origianl Creator: Ekaterina Pisareva
Original Link: https://justainews.com/applications/face-and-image-recognition/lumana-vision-for-agentic-ai-and-videosurveillance/
Originally Posted: Wed, 28 Jan 2026 04:50:49 +0000