Barry Norton, Fellow, Milestone Systems considers control room operator fatigue in the age of AI-driven video security.
Ask any control room operator how long they can watch a wall of monitors before their attention starts to slip, and most will give you an honest answer that’s well short of a full shift.
That is not a character flaw; it’s human physiology working exactly as it should, and the research on sustained visual attention backs it up.
Guidance from the UK’s Center for the Protection of National Infrastructure, now part of the National Protective Security Authority, describes what researchers call the vigilance decrement, which typically sets in after 20 to 30 minutes of continuous monitoring, depending on the level of concentration required.
A peer-reviewed study of 42 working video monitoring operators, published in the journal Applied Ergonomics, put real numbers behind it: the operators detected only about half of the target behaviors in a 90-minute video review task, with a high rate of false alarms alongside the misses.
These were not novice operators; they were trained professionals running into limits that no amount of experience fully overcomes.
That distinction matters because it reframes the problem. Video security has never really had a staffing problem so much as a physics problem.
A single operator can be asked to watch dozens of camera feeds, but the human visual system was not built to process that much simultaneous, low-signal information for hours at a stretch.
Consider further night shifts, the cumulative fatigue of shift rotations, and the psychological weight of knowing that missing something has real consequences, and the gap between what the job demands and what a person can reliably deliver only widens.
No training program closes that gap on its own, because it is not a knowledge problem. It is a capacity problem.
The industry’s early would-be solutions to this problem often made it worse, and these mechanisms still show up in poorly-designed newer deployments: motion-based detection and simple rule triggers generate constant alerts.
A shifting shadow, a blowing tree branch, a delivery truck idling near a fence line. Multiply that across a facility with dozens or hundreds of cameras, and operators are staring at an alarm queue that never stops growing, the overwhelming majority of which turns out to be nothing of note.
That volume quickly cultivates the wrong behavior; when most alerts are noise, operators learn – understandably and rationally – to treat all alerts with less urgency. It is not laziness.
It is a reasonable adaptation to an unreasonable signal-to-noise ratio, and it happens whether the operator is a two-year veteran or someone six weeks into the job.
Response times slow as a result, and genuine threats blend into a background hum of false positives, with the operator who eventually catches something serious often catching it despite the system, not because of it.
Over time, this erodes something harder to rebuild than a response metric: security leadership loses the ability to demonstrate that the technology is worth the investment, and turnover climbs in roles where burnout is already a known risk, compounding the staffing problem the technology was supposed to help solve.
This is not an anecdote, but the default failure mode of any detection system that generates more alerts than a person can meaningfully evaluate, which is why the fix cannot simply be asking operators to concentrate harder or hiring more people to stare at more screens.
Staffing alone cannot outrun a psychological limit; the fix has to change what reaches the operator in the first place, not how hard the operator tries to keep up with what reaches them.
This is where AI can earn its place, not as a replacement for human judgment, but as a filter so operators can spend their limited attention on the alerts that deserve it.
Most of what’s doing that work today is properly described as agentive AI: systems that assist people and keep humans in the loop, distinct from fully agentic AI, which pursues goals and takes action with less human involvement.
Modern analytics platforms, increasingly built on vision language models capable of describing what’s happening in a scene rather than just flagging pixels,can distinguish a person from an animal, a vehicle on an approved route from someone loitering near a restricted area and routine activity from behavior that warrants a look.
The value increases when analytics move beyond simple filtering into anomaly detection. A vehicle circling a building perimeter once might not, in itself, look unusual; the same vehicle appearing near the same fence line on three consecutive evenings, lingering just long enough to observe before moving on, is a pattern that requires connecting dots across staff shifts whose operators individually saw nothing alarming.
The same logic applies to post-incident investigations, where the workload problem is just as real. Reviewing hours of stored video data manually is slow, tedious, and prone to the same attention limits that affect real-time watching.
Technologies designed for condensed video review compress hours of footage into minutes by layering detected objects and events from different times into a single, searchable view, each still linked back to its original timestamp for verification.
One law enforcement agency using this kind of platform analyzed roughly 1,200 hours of video in under an hour during a kidnapping investigation, identifying the vehicle and driver in time to locate the victim safely.
That is the difference between an investigation that stalls and one that closes while the trail is still warm.
Security leaders evaluating whether to invest in analytics do not need to take the operator-fatigue argument on faith, because the return on investment data is now substantial enough to stand on its own.
A survey on the ROI of video analytics, conducted by Omdia, and based on the responses of 140 end users across North America and Europe, found that over 85% of organizations achieved a return on investment from the video analytics deployment within one year, and more than half saw returns within the first six months, with banking and manufacturing leading the pack at 95% and 90% respectively.
What these numbers demonstrate goes beyond budget justification: reducing operator load is not a soft benefit dressed up in return-on-investment language.
It shows up in solvability rates, in response times and in whether a security program can prove its worth to the people funding it, which is often the difference between a program that gets renewed and one that gets quietly cut in the next budget cycle.
Again, none of this is an argument for taking humans out of the loop. The distinction matters: agentive AI assists people and keeps them in the loop, while fully agentic AI pursues goals and takes action with greater autonomy, raising real questions about trust, accountability and operational risk that the industry is still working through.
What’s described here – filtering, pattern flagging, evidence review – sits firmly on the agentive side of that line.AI-driven video analytics exist to protect the value of human judgment by making sure it gets applied where it matters, not to substitute for it.
An operator who reviews five well-qualified alerts an hour, each with real context attached, will consistently outperform one drowning in fifty raw notifications, not because the second operator is less capable, but because attention is a finite resource and no amount of training changes that fact.
The organizations applying new technology correctly aren’t asking whether it’s AI agents or human operators who should run their security operations; they’re asking how to design the handoff between the two so each does what it is good at.
Machines are well suited to tireless, continuous monitoring across more feeds than any person could watch, and to spotting patterns across time that no single shift would ever see.
People remain essential for judgment, context and the kind of decision that carries real consequences.
Vigilance decrement is not going away; it’s a fixed feature of human physiology, as real and well documented as the limits of eyesight or reaction time.
What has changed is that the industry finally has tools that account for that reality rather than working against it.
Video security programs built around that fact, rather than around the assumption that operators can simply try harder, are the ones that will hold up under real operational pressure.
About the Author
Dr. Barry Norton is a Fellow, and former VP for Research, at Milestone Systems. He has 25 years of experience in AI, and has worked on large-scale applications in many industries, including as Head of Digital Platform at Mærsk. He completed his PhD in Computer Science at the University of Sheffield and carried out research and teaching at several universities across Europe.
This article was originally published in the September edition of Security Journal Americas. To read your FREE digital edition, click here.