Why Your Camera Alerts Get Ignored — and How to Fix Alert Fatigue
A camera system that fires more than a handful of alerts per shift stops being read within two weeks. In the US, 94–98% of burglar alarm activations are false — the same cry-wolf pattern kills factory alerts. Tune for decisions, not detections: three to seven alerts per supervisor per shift, each with one named owner and a two-tap response.
If you run a 200–1000 worker plant in Chakan, Hosur, Sanand or Coimbatore, you have probably already bought cameras — and possibly analytics on top. The cameras work. The problem is the WhatsApp group that now buzzes 300 times a day, which your production head muted in week three. This article is about that failure, because it is the single most common reason a camera-AI project quietly dies.
Key points
- Alert fatigue is a documented failure mode, not a soft complaint. The US Department of Justice's problem-oriented policing guide False Burglar Alarms is the standard reference for what happens when a channel is dominated by noise: responders deprioritise it. Hospital alarm-fatigue literature reaches the same conclusion. Factories are neither police departments nor hospitals, but the human response is identical — after roughly a week of noise, people stop looking.
- Price the noise. At an all-in shift-supervisor cost of ₹45,000/month over 26 working days of 8 hours, one supervisor-minute is about ₹3.6. Three hundred alerts a day at 40 seconds each is 200 minutes — roughly ₹720 a day, about ₹19,000 a month per supervisor, spent reading messages that mostly say nothing. Those are illustrative numbers; run them on your own payroll before your next vendor demo.
- Set an alert budget, then fit the rules to it. Decide first how many decisions a shift each role can absorb, then switch off every rule that does not fit. Process-industry alarm practice codified this decades ago: the ISA-18.2 alarm-management standard and EEMUA 191 treat the alarm list as something designed to a rate and measured, not something accumulated.
- Route by who can act, not by who is senior. A helmet violation belongs to the line supervisor. A gate-weighbridge mismatch belongs to stores plus you. Nothing goes to everybody.
- Indian plants have Indian false-trigger sources. Monsoon rain streaks on outdoor domes, insects swarming IR illuminators, grinding dust, forklift beacons, shift-change crowds, and DVR reboots after a grid dip. All are tunable; none are the AI "being stupid."
- Some alerts you should not send at all. India's DPDP Rules were notified on 13 November 2025, with the substantive notice, security and breach duties commencing 14 May 2027. Behaviour alerts naming individual workers are an industrial-relations risk today and a compliance risk on that clock.
Why the buzzing group chat is the default outcome
Most analytics ship with every rule enabled, because that is how a demo impresses. Loitering, intrusion, crowding, PPE, motion, tamper, line-crossing — on twelve cameras, at 24 hours a day, that is thousands of events.
Then two things happen. First, the true positives get buried: a real pilferage event at the scrap gate arrives as message number 214 of the day. Second, the supervisor learns, correctly, that reading the group has a low hit rate. Muting is a rational decision by a busy person. Blaming him is the wrong diagnosis.
The honest way to say it: a system nobody reads is worth close to zero. Worse than zero, arguably — because you paid for it and you now believe you have coverage you do not have.
Set the alert budget first
Work backwards from human attention. Below is a starting budget for a mid-size Indian plant running two or three shifts; adjust to your span of control. It is a design proposal, not a benchmark drawn from published data.
| Recipient | Alert budget | What they should get | Channel | Response expected |
|---|---|---|---|---|
| Line supervisor | 2–4 per shift | Machine stopped >15 min, guard/interlock bypass, danger-zone entry | WhatsApp, immediate | Act on floor, mark done |
| Security in-charge | 1–3 per shift | Gate vehicle without gate pass, scrap loading outside window, perimeter breach after 22:00 | WhatsApp + control room screen | Physical check, log |
| Stores / dispatch | 0–2 per shift | Weighbridge mismatch, loading bay idle >30 min | Reconcile against challan | |
| Safety officer | Daily digest | PPE compliance %, near-miss clips, night-shift escort and transport checks where your state's conditions require them | Once daily, 09:00 | Weekly review |
| Owner / plant head | 1 per day | One shift summary: output vs. plan, top three losses, anything unresolved | Once daily, 07:30 | Read in 90 seconds |
Notice what the owner gets: one message. Not a feed. If you as the owner are receiving live alerts, the routing is broken — you are being used as the escalation path for a system that has no escalation logic. Our view of what that daily message should contain is in /efficiency/daily-shift-report-factory-owner-india/.
Where Indian false triggers actually come from
Most "the AI is unreliable" complaints trace to a handful of physical causes. They are fixable in days, not quarters.
| False-trigger source | Typical plant | Fix |
|---|---|---|
| Monsoon rain streaks and water on the dome | Any outdoor camera, Jun–Sep | Hood/shroud, re-angle away from prevailing rain, raise motion threshold seasonally |
| Insects swarming IR illuminators at night | Rural and peri-urban units | Separate the illuminator from the lens, or use white-light at gates |
| Grinding, foundry and cement dust on the lens | Rajkot, Coimbatore, Ahmedabad clusters | Weekly lens-cleaning SOP; a dirty-lens health alert instead of 200 motion alerts |
| Forklift beacons and welding flash | Assembly and fabrication bays | Exclusion masks on the light source region |
| Shift-change crowd at gate and canteen | Every two- or three-shift plant | Time-window suppression across your actual changeover windows |
| Grid dip → DVR/NVR reboot storm | Almost universal | Suppress the first 5 minutes after reconnect; alert on the outage itself instead |
| Stray dogs and birds in yards | Yards, scrap areas | Size and dwell filters; classify object type before alerting |
One rule of thumb: if a rule fires more than twice a shift and nobody has ever acted on it, it is not an alert. It is a metric. Move it into the daily digest and take it out of the live channel. Camera placement decides how much of this you avoid up front — see /cameras-and-placement/where-to-place-cameras-on-a-shop-floor/.
The 30-day tuning protocol
Days 1–10: shadow mode. Every rule runs, nobody is notified. All events land in a log. You are measuring the natural rate.
Days 11–20: score the log. For each rule, take a sample of 50 events and mark them acted / ignorable. Anything below roughly 50% "would have acted" gets switched off or narrowed by time window, zone or object size. Keep the count of live rules small enough that you can name them all from memory.
Days 21–30: go live with confirm/dismiss. Every alert must have two taps: "handled" or "not real." The dismiss tap is the record you tune against. If your vendor cannot show you a weekly false-positive rate per rule after this, you cannot manage the system, and you should say so before renewal.
Then review monthly. Alert budgets drift upward the way email does; someone always wants "just one more rule."
What cameras do not fix — and where they can hurt you
Being straight about this is more useful than another feature list.
Cameras do not create accountability. If nobody currently follows up on a stopped machine, an alert saying the machine stopped changes nothing. The alert lands on the same non-decision. Fix the escalation path first; the camera only shortens the time to knowing.
Cameras do not survive bad networks. In plants where the shop-floor link drops daily, cloud-only analytics generate a second class of noise — connectivity alerts — that trains people to ignore the channel even faster. Weigh the architecture honestly: /cost/nvr-vs-cloud-vms-factory-video-storage-india/.
Individual-level behaviour alerts can backfire. Naming a worker in a WhatsApp group for a helmet violation is a fast route to union friction, and it points at your weakest spot under the DPDP framework. The Act's ceiling for failure to take reasonable security safeguards is ₹250 crore, though that duty is enforced from 14 May 2027 rather than today. Alert on the zone and the shift, discipline through your existing HR process, and post a proper worker notice: /compliance/worker-cctv-privacy-notice-dpdp-india/.
Some areas should have no cameras and therefore no alerts. Toilets, changing rooms and crèche areas are off-limits regardless of what the analytics could do there. See /compliance/cctv-banned-areas-factory-india/.
Compliance changed — do not tune against the old law
If your consultant is still writing SOPs against the Factories Act 1948, stop. The four labour codes, including the Occupational Safety, Health and Working Conditions Code, 2020, were brought into force on 21 November 2025; the OSH Code replaces 13 central labour laws, the Factories Act among them. Your safety-alert design — danger zones, night-shift arrangements for women workers, contractor entry — should be written against the Code and the applicable state rules, not the old Act. State rules are still being notified in many states, so check yours rather than assuming.
This matters for alert fatigue specifically, because compliance-driven alerts are the ones most often left on "just in case." They are also the ones most often ignored. Better to run a small, defensible set and retain the footage properly: /compliance/cctv-footage-retention-period-india-factories/.
What good looks like after 90 days
Five to seven alerts a day across the whole plant, each one acted on. A single 07:30 message to you. A weekly false-positive number that a person owns. Rules you can list from memory.
If instead you have a muted group with 4,000 unread messages, the honest read is not that AI does not work in factories. It is that nobody set a budget, and the system was allowed to spend your supervisors' attention without a limit.
FAQ
How many alerts per day is too many? Practically, more than about ten live alerts a day across a plant and people begin skimming; more than thirty and they stop reading. Aim for three to seven per supervisor per shift. Everything else belongs in a once-daily digest, which is read deliberately rather than reflexively dismissed. These are working rules of thumb from deployment experience, not published thresholds.
Should alerts go to WhatsApp or a separate app? WhatsApp, for Indian plants. Supervisors already live there, adoption is free, and nothing needs installing on personal phones. The discipline must come from volume control, not from a separate app people will not open. Use a dedicated group, not the general plant group.
My vendor says false positives will reduce with more data. True? Partly. Model tuning helps, but most false triggers in Indian plants are physical — dust, rain, insects, beacons, shift crowds. Those are solved by masks, time windows, hoods and cleaning SOPs, usually within two weeks. Ask for a per-rule false-positive rate; if the vendor cannot produce one, more data will not save you.
Does turning off rules mean missing real incidents? You will miss some low-value events. You will catch more high-value ones, because they are no longer buried. A rule that fires 200 times and is never acted upon is not providing coverage — it is providing the appearance of coverage, which is worse.
What is a reasonable first set of live rules? Three to five. Typically: danger-zone entry on the two most hazardous machines, machine stopped beyond a threshold, vehicle at gate without a gate pass, and after-hours perimeter breach. Add rules only when one is retired. See /safety/machine-guarding-danger-zone-detection/.
Who should own alert tuning inside the plant? One named person — usually the safety officer or the production planning head — with a monthly 30-minute review. Ownership by "the vendor" or "IT" fails, because neither feels the cost of noise. The person who receives the most alerts should have the authority to switch rules off.