Every year, more than 395 million workers worldwide are injured on the job, and almost 330,000 lose their lives in occupational accidents, according to the latest global estimates from the International Labour Organisation (ILO). When deaths caused by work-related illnesses are included, that number rises to an estimated 2.93 million.
There has been real progress in certain places. For instance, in 2024, employers in the American private industry noted their lowest level of injuries and illnesses for the last twenty years according to the Bureau of Labor Statistics (BLS). Nevertheless, the risks associated with the worldwide statistics are the same on every factory floor: a near-accident involving a forklift, a worker’s hand near the press, or lifting an object over and over again.
I have seen these risks firsthand. I began my professional journey as a building engineer and have spent the past several years on industrial sites across Asia, the Middle East, and Africa. Across the industrial sites, cameras are common. The harder question is what happens after a risk appears on screen: who assesses it, who acts, and how does the team know the problem has been addressed? So, the challenge isn’t detection anymore, it is making that data actually useful on the ground.
Why Traditional CCTV Falls Short on a Moving Floor
Most camera systems in manufacturing were installed for the purpose of security and incident review. The footage is reviewed after the event has happened. Even with a control room, one person may be responsible for dozens of feeds across a shift. Important details can be missed, not necessarily because someone is careless, but because continuous monitoring places practical limitations on human attention. The next question is whether the site has a reliable process for escalating and resolving what is noticed.
Manufacturing floors are dynamic. Layouts change, temporary work zones appear, and forklift routes get re-routed around new lines. Early rule-based video systems struggled with this. For instance, a motion alarm in a busy aisle fires all day, and teams learn to mute it. At that point, the system is technically running but practically ignored.
What Changes When AI Can Reason About a Scene
The first generation of computer vision safety was mainly classification: Is that worker wearing a hard hat? Is someone in this marked zone? Useful but narrow.
Newer approaches, including those using Vision Language Models (VLMs), combine context with object interpretation. A worker without gloves walking into the break room is totally different than the same worker leaning into a machine guard. Similarly, a worker behind a barrier next to a forklift does not have the same risk as one walking into its path. Recognising these differences may help safety teams focus on situations that warrant closer attention.
When video analytics identifies a potential risk, the next step naturally is deciding who should review it and what action is required. Recording incidents by area, and shift can also help safety teams identify recurring problems.
Posture and repetitive movement risks, which can be difficult to notice simply by looking around, can also be detected using video analytics on the video footages captured by the cameras already installed in the jobsite.
None of this replaces a supervisor’s judgement. These systems can still make mistakes, and they are more useful as an additional pair of always-watching eyes, than as a decisive solution.
Fitting AI Into the Safety Program Manufacturing Sites Already Run
Manufacturing plants can get more value from AI, when they see it purely as an operational improvement, rather than a software purchase. A practical starting point is to use existing camera footage and incident logs to identify one or two hazards that appear most frequently.
Before switching on the system, teams should decide who would be responsible for reviewing particular type of alert and what action is expected and how that action will be followed up. Without that clarity, alert can quickly turn into background noise.
Workers should also be involved in these discussions. They may have concerns about how camera footage is used and what behaviour is monitored and who can access the recordings. Explaining them the purpose of monitoring and the privacy safeguards in place, such as face blurring and end-to-end encryption, can help build trust.
The criteria of measuring success also play an important role. While the number of alerts is important, it cannot be viewed in isolation from how quickly the hazard is rectified as well as whether similar conduct is repeated or not.
What This Looks Like in Practice
One such example comes from a power generator manufacturer in Dubai with over 8,000 employees. As the company consolidated several operations into a new facility, the usual safety assumptions stopped holding. Forklift drivers worked unfamiliar routes, and assembly teams had to adapt to redesigned zones. Periodic audits and manual checks could not keep pace.
The team used video analytics to monitor helmet and safety shoe compliance on the assembly lines, along with forklifts and pedestrians sharing common space. Alerts were triggered, and the operators and supervisors were informed of the events, and every violation was logged in one place. The site witnessed a decline in overall safety violations by 54% and reduction in response times by 71%. Spearheaded within that, PPE violations along assembly line operations reduced by 88%, with forklift operations becoming 65% safer via reduced pedestrian collision risks.
The lesson learned here is about timing. Technology mattered the most when the facility was least predictable, not after operations had settled into a routine.
The Real Measure of Progress
The next gains in manufacturing safety are unlikely to come from one camera or more rules. They will come from closing the gap between seeing a hazard and acting on it, ideally within minutes. Record-low injury counts are worth acknowledging. For the people who still get hurt on the floor, the more useful question is what their supervisors could have known, and when.
