AI agents are seeing increasing use cases in security. Successful implementation requires an understanding of the architecture with which AI agents are deployed. While the technology is new, its potential is not to be ignored.
AI agents are seeing increasing use cases in security. Successful implementation requires an understanding of the architecture with which AI agents are deployed. While the technology is new, its potential is not to be ignored.
AI agents differ from conventional AI in various ways. Compared to traditional AI, AI agents are more autonomous (requiring less human supervision/oversight), more context-aware (knowing not just the what, but also the who, when, where, why and how), and decision-driven (recommending what course of action to take in the event of a security breach). Various use cases are emerging; agents detecting loiterers in a store and sending an alert, or agents seeing a motor scooter rider in a restricted area and alerting the nearest police officer – these are some examples of AI agents at work.
Trending towards the edge
Architecture-wise, AI agents can be deployed at the edge or in the backend server. Each has its own merits and disadvantages.
“From a deployment perspective, on-premise and hybrid approaches are increasingly preferred in physical security. Video data is sensitive, bandwidth-heavy, and often subject to regulatory or customer constraints. Running agents on site ensures data sovereignty, low latency, and predictable performance,” said Koen de Jong, Founder and CEO of Visionplatform AI.
Yet there are also disadvantages of running agents at the edge, the biggest of which is limited processing inside the camera.
“We now have AI models inside the cameras, which allow us to do some basic functionality. So today, for example, we do facial recognition inside the camera. We do license plate recognition inside the camera. But there's a limitation to how much data can be processed,” said Xavier Miota, VP of Sales at Incoresoft.
But Miota notes that AI agents are trending towards the edge, especially with the help of AI boxes.
“The AI box is in between a traditional server and the camera-based AI. So here, you will put an AI box at a site where you have 20 cameras, and then you can do the intelligence both on the camera and on the AI box,” he said. “Now there's a lot of companies and competitors of ours that do half of the tasks inside the camera and the other half on the server, when we used to do everything on the server before. So it's moving. And hopefully in the next five years, we will have powerful enough camera chips in order to run full AI agents.”
Not all smooth sailing
Indeed, AI agents can benefit security in numerous ways. But despite their promises, several challenges still remain. One of them is cost, as some AI agents need expensive servers to run.
“Today, everybody wants to have the best possible system at the lowest possible price. My customers, when they have to buy five servers or six servers with GPUs, they start to complain,” Miota said.
Leo Levit, Chairman of ONVIF Steering Committee, cites input and output language unification as a main challenge. “One of the challenges that ONVIF hopes to overcome is a unified input and output ‘language.’ That is to prevent solutions being dependent on one component or vendor,” Levit said.
Explainability and data quality also present some challenges. “Agents must be able to explain why they recommended or took a certain action. This requires clear logging, traceability, and human-override mechanisms,” de Jong said. “Data quality remains critical. AI agents are only as effective as the signals they receive. This is addressed by combining traditional detection with higher-level semantic understanding and continuous feedback from operators.”
Finally, de Jong cites the challenge of over-expectation. “AI agents are not a magic replacement for security teams. They must be carefully instructed, constrained, and aligned with operational policies,” he said.
Future prospects great
Despite the aforementioned challenges, and the fact AI agents are at a relatively early stage, prospects in security are great.
“We are still in the early days of adoption of this technology. I think we are in the transition between proof of concept types of installation and real-life deployments,” Levit said. “Several companies in the industry are working on this type of solution and showcasing great results. I think the potential is there and it could be a great addition to the industry’s mission to make our work safer.”
De Jong echoes those remarks, saying AI agents will become standard in physical security over the next few years.
“The industry is facing a structural problem: more cameras, more sensors, and more data, but not more operators. AI agents directly address this imbalance. They scale decision-making capacity without scaling headcount. Growth prospects are strong because this shift is driven by necessity, not hype,” he said.
Product Adopted:Other