Alert fatigue is a design problem, which means it’s ours to fix. And a quieter queue is only half the job.
At 02:40am on a winter Tuesday morning, an operator on the north side of Johannesburg clears an alert in about two seconds. Not verified, just cleared. A glance at the thumbnail, a click, next. It was a scheduled delivery truck pulling into a side bay. So were the six before it. So, likely, are the next forty. Somewhere around alert sixty in that queue is an unverified vehicle that has actually stopped at an unauthorised gate. Unless the interface intervenes, it gets the same two seconds as the delivery trucks.
We’ve covered this ground before. We’ve written about how alert fatigue shows up through camera drift, and we’ve laid out the case for AI-driven alert management that filters out the noise. But even with well-tuned models, every time we revisit the problem, whether through camera drift, model accuracy, or false-positive thresholds, we land on the same uncomfortable conclusion. The industry talks about alert fatigue as something that happens to operators. Look closely, and most of it is “inflicted on them” by the interfaces they work in.
What Fatigue Looks Like From The Chair
Spend enough nights behind operators and you watch the same pattern play out, in roughly the same order, though nobody in a control room would describe it in clinical terms.
First, the pop-ups stop reading as possible intrusions and start reading as admin. When most of the queue is wind, wildlife and headlight glare, that’s simply when the brain quietly stops sorting signal from noise. Then the slowdown creeps in. Every unnecessary click, window swap and menu hunt is paid for out of the same attention budget the operator needs, six hours into a night on the desk, and by then, the budget is overdrawn. And finally, the habit nobody wants to admit out loud: clearing the screen becomes the job. The queue number drops. The risk doesn’t.
On the sites we monitor, from residential estates and logistics yards to mine-to-port corridors, the gap between a response that arrives in time and one that arrives to take notes is usually measured in tens of seconds. We don’t have a tidy academic study to wave at you here. We have incident logs and control room footage, and they say exactly the same thing.
The Old Interfaces Made It Worse
Legacy system design ran on a simple, brutal philosophy: show everything, hide nothing, and let the user sort it out.
So a genuine exception, an unverified vehicle in a loading zone at 03:00am, sets off a little dance. Acknowledge the pop-up, hunt through a camera tree for the adjacent angle, cross-check a timestamp against a gate log, and scroll a dense table to type a remark nobody will read.
None of that is security work. All of it is billed to the same attention budget.
Fewer False Alarms Is The Easy Promise
The industry’s answer to all of this has settled on a single headline. Pick up almost any vendor pitch today and it leads with the same claim: fewer false alarms.
Fewer false alarms matter. We’ve made that argument ourselves. But it’s the easy half of the problem, because a false alarm at least announces itself. The operator sees it, judges it, and clears it. The harder question, and the one almost nobody in this market is asking out loud, is what happens to the events that never generate an alert at all.
The person who walks through the gate in a contractor’s bib at 14:00pm, because nobody thought to cover the daytime. The vehicle that idles just short of the line it was supposed to cross. The camera that was knocked ten degrees off its view on Sunday and has watched a wall ever since. The perimeter that stays perfectly calm because its rules were tuned too tightly. None of these show up in a false-alarm statistic. They show up in an incident report, after the fact.
Any system can drive false alarms towards zero by simply detecting less. A silent queue looks like success on a dashboard, and it may be the most dangerous screen in the control room. So the question we’d rather the industry argued about is this: what did the system miss, and how would anyone know? However quiet the queue gets, the measure we hold ourselves to is whether the right alerts are still in it.
What We Built Instead
When we started on the Refraime Alert Manager, we gave ourselves one constraint, and it sounds easier than it is. The operator should never have to search for what they need to see.
In practice, that means one queue, ranked by risk. Not a camera tree, not four tiled windows. A single prioritised list where high-risk alerts go straight to the top instead of waiting their turn behind the delivery trucks. Ambient noise is filtered upstream so it never earns a pop-up in the first place, and that filtering is tuned site by site, so cutting noise never means cutting coverage.
It means the frame is the workspace. The event arrives with its detection box drawn on the image, the triggered zone outlined on the scene, and, on the same screen, the context rail: which site, which camera, which zone, what time, and a defined SOP, right next to it. One click switches between the single frame and the clip, so the operator sees the movement, not just the moment. Nobody should be hunting for a phone number, or for footage, while a car idles at a gate.
Our notes panel rides alongside the frame: check for uniforms and tools, count the people in view, treat the 20:00pm – 06:00am window as high risk, phone the client on the office line during the day and the after-hours number at night, and send armed response if nobody answers. Having the procedure sit next to the picture, instead of in a binder under the desk, is one of the cheapest safety upgrades we’ve ever shipped.
And it means verdicts cost one click. Incorrect. Flag. Escalate. Flag matters most when the answer isn’t the operator’s to give. The alert stays open and moves into an awaiting-feedback queue while the client or site confirms, and the operator carries on working the queue in the meantime. Nothing gets closed on a guess, and nothing behind it waits.
Every action writes its own remark into the audit history, so accountability stops being a typing exercise.
And when the scene needs more eyes, it means the rest of the site is one click away. Live view takes the operator straight from the alert into the camera’s live feed, to see what’s happening at that gate now. The video wall goes further. It can open every camera on the site, or only the relevant and adjacent ones, so the operator can follow which way someone is moving and see what lies ahead of them on the next camera. Situational awareness stops depending on an operator’s memory of the camera layout. All the information sits right where they need it.
Then there’s the widget we added to keep ourselves honest: unactioned alerts per site, sitting in plain sight on the same screen. When a site’s number climbs, that’s a tuning conversation we need to have, not an operator failing a stamina test. And when a busy site suddenly goes quiet, that deserves the same conversation. Fatigue and silence should be visible to the people building the system, not only to the people living with the consequences.
Attention Is The Product
This industry has spent a decade bragging about detection scores, frame rates and bounding boxes, and lately about how few false alarms it sends. A model that finds everything while exhausting the only person who can act on anything has achieved precisely nothing. So has a model that stays quiet by missing what matters. We still get it wrong sometimes: a zone tuned too hot, a note that reads like legalese. When we do, the unactioned alert counter is usually the first to tell us.
But at 03:00am, the delivery truck should get two seconds and a click, while the vehicle at the gate gets an operator’s full attention, a call to the client, and a response team on the road. That difference, not the number on a model card or a false-alarm percentage, is what real security should look like.
Come and sit behind an operator with us. Book a demo, and bring the one question every vendor should be able to answer: what does your system miss?
Author: Hercules Le Roux
Lead Engineer @ Refraime
