Agentic AI in physical security: detect, report, action

Most video systems stop at the alert. Agentic AI closes the loop, from spotting an incident to reporting it and putting a response in motion.

Walk into any security operations room and you will see the same picture: a wall of camera feeds and one or two people expected to watch all of them. The research on this is old. A human operator misses most of what happens on screen after just twenty minutes of monitoring. Cameras have never been the problem. Attention is.

The first wave of video analytics helped by generating alerts. But an alert on its own is just noise with a timestamp. Someone still has to notice it, work out whether it matters, log it and decide what to do. When the alerts arrive by the hundred, that chain breaks at the first link.

Agentic AI is the next step, and it changes the shape of the problem. Instead of a tool that flags events and waits, you get a system of AI agents that detect an incident, report it in plain language and then act on it. Detect, report, react. Here is how that works inside icetana AI, using one of the most common incidents in any building: someone slipping on spilled liquid.

Detect: the anomaly surfaces itself

icetana AI learns what normal looks like for every camera it watches. Not through rules someone has to write and maintain, but by continuously modelling the routine movement in each scene. When something breaks that pattern, a person going down on an escalator, a crowd suddenly running or someone on the floor who should be walking, the event surfaces on its own.

That matters here, because nobody schedules a fall. It happens in aisle seven at 2:47pm on a Tuesday, in the blind spot of whoever is nearest. The CCTV camera saw it, the way it sees everything. With self-learning detection running on that feed, the system noticed.

Report: the Triage Agent decides what matters

Detection alone would just move the bottleneck. This is where the agentic part earns its name. The Triage Agent analyses each detected event, filters out what does not matter and explains what does. A kid lying on the ground for fun is not the same as an elderly person who has gone down in a restricted area. The Triage Agent treats them differently.

The operator does not get raw footage and a shrug. They get a prioritised icetana event with a description of the incident: what happened, where and how urgent it is. The event is tagged and logged for incident reporting, which is exactly the evidence trail you want when the question later becomes "when did you know, and what did you do?" For a slip-and-fall claim, the difference between a documented three-second detection and a hazard that sat unnoticed for forty minutes is the difference between a closed file and a settlement.

React: robots and people close the loop

This is the step most video systems never reach, and it is where our partnership with SoftBank Robotics comes in. Together we have launched SmartBX Security, an AI video analytics layer that turns a building's existing cameras into a self-learning intelligence network, and it connects detection directly to response.

Back to our fall. The moment it is detected and triaged, the system can issue instructions, and the two responses run in parallel. Nearby staff get a device alert through Relay Agent, telling them where the person is and what to expect when they arrive. A cleaning robot is dispatched to the same aisle to deal with the liquid that caused it, so the next person through does not go down as well. The incident that used to rely on a customer complaint or a passing employee is now being handled in seconds, not hours.

SoftBank Robotics' integration roadmap uses icetana AI detections as triggers. A detection can deploy a robot for onsite investigation. It can also flag a maintenance issue like a wet floor or a burnt-out light and route it straight to a robot or facility staff. The camera network stops being a passive record of what went wrong and becomes the starting point of the fix.

Why "agentic" is more than a buzzword

Plenty of products now wear the agentic label. The test is simple: does the system carry an incident through to an outcome, or does it hand you a notification and leave? An agent perceives, decides and acts. In icetana AI's case that means self-learning AI that learn each scene, a Triage Agent that filters and explains events in real time and integrations that turn a verified incident into a physical response.

None of this removes people from the loop. It puts them where they are most useful: making judgement calls on incidents that have already been found, verified and described. And because the platform runs on existing cameras, on-premises where required, the path from a monitored building to a responsive one does not start with ripping anything out.

One person slipping on a wet floor is not a small incident. But it is the whole story in miniature: detected in seconds, reported with context, resolved by the right responder before it becomes a second incident. That loop, repeated across every camera and every hour of the day, is what agentic AI actually means for physical security.

See what your cameras are missing. Book a free demo.