Dijam Panigrahi makes a counterintuitive but data-backed argument that routing every sensor alert through human approval does not make cement plants safer.

India’s cement industry has spent the last two years wiring kilns, mills and coolers with sensors and automated control systems, and the safety case for doing so is strong on paper. Contract workers still make up the majority of the industry’s workforce, and fatal accidents remain a recurring problem. The Indian National Cement Workers Federation has noted that around 83 per cent of workers in the sector hold precarious positions, a fact that resurfaced after an oxygen cylinder explosion killed three contract workers at a plant in Chhattisgarh.
Industry tallies compiled by IndustriALL found cement plants recorded at least seventeen accidents in one year with 21 workers killed, and ten accidents the following year with nine killed, most of them contract staff. Automated monitoring, in theory, closes that gap. A sensor never gets complacent and never skips a check because a shift is short staffed.
However, plants that respond by routing every anomaly reading to a person for approval are quietly building a system that fails the same way understaffing does. When operators receive dozens of flagged deviations a shift, most of them minor, they learn a simple lesson: the fastest way through the queue is to approve without reading closely. The safety benefit disappears, not because the technology failed, but because the humans supervising it adapted to the volume.

Why alerts get ignored
A study cited by manufacturing technology publisher Applied SmartFactory found more than 95 per cent of alarms in a semiconductor fab were low priority, and only about 4 per cent ever triggered an action, with just 100 out of 5,000 alarms accounting for 70 per cent of all alarm activity. The mechanism is the same whether the trigger is a vibration sensor or an AI model flagging a kiln temperature swing. Once the ratio of noise to signal crosses a threshold, workers stop treating the system as a decision aid and start treating it as a formality to clear.
The scale of AI deployment underway makes this more than a theoretical risk. Stanford’s 2026 AI Index Report found organisational adoption of AI has reached 88 per cent, even as documented AI incidents rose to 362 in 2025, up sharply from 233 the year before, according to analysis of the report. The Index also found only about a third of organisations have adopted a formal governance framework, with NIST’s AI Risk Management Framework cited by 33 per cent and ISO/IEC 42001 cited by 36 per cent.
Most manufacturers are deploying monitoring systems faster than they are building the judgment for when a flagged event actually needs a person’s attention. In India, plants run by JK Cement have begun pairing CCTV feeds with AI to define safe zones around heavy machinery, a promising direction that still depends on operators trusting and reading the alerts the system generates.

A three-tier model for cement plant
The fix is not less monitoring or more monitoring. It is classifying decisions by risk and by novelty, rather than treating human oversight as a single switch that is either on or off. A workable model sorts factory floor events into three tiers.
The first tier, proceed, covers deviations the plant has seen before that fall within known safe bounds, such as a kiln feed rate adjustment within an established range. These should run without a stop for approval, because routing them to a person only trains that person to click through.
The second tier, pause, covers events that are unusual but not yet dangerous, such as a vibration reading trending toward a limit or a fuel blend shifting outside its typical mix. These warrant a brief human check before the system proceeds, giving an operator the chance to apply judgment the model does not yet have.
The third tier, escalate, covers events that are both high risk and unfamiliar, such as a pressure reading combined with a temperature spike that has no close precedent in the plant’s history. These should stop the process entirely and require a decision from someone with the authority to shut down a line.

Who should set the threshold
Where these tiers get drawn matters as much as the framework itself. Threshold setting is frequently handed to the vendor supplying the monitoring software or to a plant’s IT department, both of which understand the technology but not the specific tolerances of a given kiln, mill or line. Operations staff, who know that a particular grinding unit runs hotter under monsoon humidity or that a calciner behaves differently after a refractory reline, are better positioned to calibrate what counts as routine on their own equipment.
Handing threshold ownership to operations does not remove IT or vendors from the process, but it puts the calibration decision closest to the people who live with its consequences on the floor.

Signals that oversight is actually working
A few concrete indicators reveal whether a monitoring setup is functioning as intended or simply providing the appearance of safety. The escalation rate over time is the first: a rate that stays flat or climbs slowly as operations mature is healthy, while one that spikes and then falls sharply often means operators have started overriding the system rather than engaging with it. Time to resolution is the second: escalations that take progressively longer to close suggest fatigue or confusion about ownership, not diligence. The third, and most telling, is how accurate the system’s own uncertainty estimates turn out to be, meaning whether events flagged as high risk actually correlated with real incidents, and whether events waved through stayed incident free. A system whose escalations do not track with actual outcomes trains operators toward the same complacency that unmonitored equipment produces.
None of this argues against automation in Indian cement manufacturing, where a labor structure built on contract work and a track record of serious accidents make better monitoring an urgent need. It argues for treating human oversight as a design problem with three distinct settings, rather than a single
dial turned up whenever a plant wants to look safer on paper.

About the author:
Dijam Panigrahi, Co-founder and COO, GridRaster, is a spatial computing platform for industrial enterprises and manufacturers.