Why 80% of workplace injuries are preventable (and what AI can do about it)

There's a statistic that comes up again and again in workplace safety research: the vast majority of workplace injuries are preventable.

Not theoretically preventable. Not "if we lived in a perfect world" preventable. Preventable with better visibility, better training, and better systems for catching risk before it turns into harm.

The numbers support this consistently. The US National Safety Council classifies workplace deaths as "preventable" when they result from identifiable and controllable hazards, and in 2024, 4,337 preventable work deaths were recorded in the United States alone. OSHA's own data shows that powered industrial truck incidents (one of the leading causes of warehouse fatalities) are overwhelmingly linked to inadequate training, poor traffic management, and lack of real-time oversight, all of which are addressable.

So if we know these injuries are preventable, why do they keep happening?

That's the question worth sitting with. Because the answer isn't that safety teams don't care. It's that the tools most organisations rely on weren't designed for the job they're being asked to do.

The scale of preventable workplace injuries

Before we talk about solutions, it helps to understand just how widespread workplace injuries still are.

In the United States, private industry employers reported 2.5 million nonfatal workplace injuries and illnesses in 2024, according to the Bureau of Labor Statistics. That's a slight improvement on 2023, but still a staggering number. The NSC estimates that 3.95 million work-related injuries required medical consultation in the same year. The AFL-CIO's 2025 "Death on the Job" report suggests the true toll may be even higher, estimating 5.2 million to 7.8 million workplace injuries and illnesses annually when accounting for widespread underreporting.

In New Zealand, Stats NZ recorded 209,400 work-related injury claims in 2024. Serious claims have increased 34.5% over the past decade. If workplace injuries and illnesses were eliminated entirely, Safe Work Australia estimates the Australian economy would grow by $28.6 billion annually.

Behind every one of these numbers is a real person. Someone who went to work that morning expecting to come home safely. A family that got a phone call nobody wants to receive. A team that lost a colleague. The economic figures are confronting, but the human cost is what makes this worth solving properly.

Why 'preventable' doesn't mean 'prevented'

If most injuries are preventable, there's clearly a gap between what we know and what we do about it.

Most workplace safety programmes are built around a combination of training, procedures, audits and incident investigation. These are all important. A strong induction programme, regular toolbox talks, well-written standard operating procedures, and thorough incident reviews are the foundations of any good safety culture.

Bu they share a common limitation: they're largely retrospective or periodic. Training happens at set intervals. Audits happen on a schedule. Incident investigations happen after someone has already been hurt. Safety walks capture a snapshot of one route, one shift, one observer's perspective.

The gap isn't in the quality of these approaches. It's in their coverage. What's happening between the safety walks? On the night shift when there's no supervisor on the floor? At the site nobody visited last week? In the aisle where someone took a shortcut they've taken a hundred times before.

This is where the concept of near misses becomes critical.

The near-miss problem

Safety research has consistently shown that serious injuries don't just happen in isolation. They're preceded by patterns of near misses, unsafe behaviour, and minor incidents that build up over time. Herbert William Heinrich's research (later expanded by Frank Bird and others) established a ratio that's become foundational in safety thinking: for every serious injury, there are hundreds of near misses that went unrecorded.

The challenge is that most organisations only capture a fraction of their near misses. Manual reporting systems rely on people choosing to report something that didn't actually result in harm, which is a big ask in a busy environment. Research consistently suggests that manual near-miss reporting captures only a small fraction of actual events.

This means the data most safety work with is incomplete. They're making decisions based on the incidents that were severe enough to be reported, while the patterns that could predict and prevent future incidents remain invisible.

What computer vision AI changes

This is where technology starts to shift the equation.

Computer vision AI uses existing CCTV camera infrastructure to monitor work environments continuously. Instead of relying on someone being in the right place at the right time, or on workers self-reporting near misses after the fact, the AI watches camera feeds in real time, or on workers self-reporting near misses after the fact, the AI watches camera feeds in real time and identifies specific unsafe behaviours and events as they happen.

The types of events it can detect include pedestrian-forklift proximity breaches, exclusion zone breaches, gaps in PPE usage, vehicle speeding, and loading zone hazards. Each detected event is captured with video evidence, timestamped, and made available to safety teams for review and follow-up.

What makes this genuinely useful (rather than just another alert system) is what happens with the information. The most effective implementations use detected events as coaching opportunities rather than disciplinary triggers. A safety manager gets a short video clip of an actual event from their own site, their own shift, and can have a specific, constructive conversation with their team. "Hey, check out what happened at the loading dock yesterday. How could this have gone wrong?" is a fundamentally different conversation to "remember to follow the rules."

This coaching-first approach matters because safety culture only improves when the people doing the work are engaged in the process. Surveillance-style monitoring tends to drive underreporting and resentment. Coaching-led monitoring tends to surface more honest conversations about why shortcuts and complacency happen, and what can be done about them.

Privacy matters

Any conversation about AI and workplace cameras has to address privacy head-on. It's the first concern most people raise, and rightly so.

The best implementations of computer vision AI are designed with privacy as a foundational principle, not an afterthought. This means processing camera feeds on-site (so footage doesn't get sent to the cloud), automatically blurring faces and identities before any footage is reviewed, and maintaining robust data security certifications like SOC 2 Type II, ISO 27001, and GDPR compliance.

The goal is to identify unsafe events and behaviours, not to monitor or identify individual workers. Safety teams see what happened, not who did it. The coaching conversation happens face-to-face, person-to-person, not through a surveillance screen.

When privacy is built in properly, teams tend to engage with the technology rather than resist it. That engagement is what drives the cultural shift that makes lasting safety improvement possible.

What the results look like

Organisations using computer vision AI for workplace safety are reporting significant, measurable improvements.

At inviol, our customers see an average 67% reduction in site risk after implementing the platform. Actual incidents drop by up to 42% over three years, and machine-on-plant incidents (one fo the most common and dangerous event types in warehousing and logistics) reduce by up to 61%.

These results come from teams at organisations including Coca-Cola, Woolworths, NZ Post, Americold, The Warehouse Group, Linfox, Placemakers, and Whittaker's, across industries ranging from cold storage and logistics to retail, manufacturing, and horticulture.

The improvements happen because safety teams are working with better information. Instead of reacting to incidents after they occur, they're identifying patterns in near misses and coaching on them before they escalate. The data gets richer over time, revealing which zones are highest risk, which shifts see the most events, and whether interventions are actually working.

The 80% that's within reach

The statistic that 80 % of workplace injuries are preventable isn't aspirational. It's a reflection of what the data consistently shows: most serious injuries result from known, identifiable, and addressable hazards.

The gap has never been about awareness. Safety managers know where their risks are. The gap is about visibility, coverage, and the ability to act on information in time.

Computer vision AI doesn't replace the judgment, experience, and relationships that good safety managers bring to their work. What it does is extend their reach. It gives them eyes across every monitored zone, every shift, every day, with the same standard applied consistently. It captures the events that used to fall through the cracks and turns them into coaching moments that build a stronger safety culture over time.

The technology exists. The results are proven. And the 80% of preventable injuries? They're not a statistic to accept. They're an opportunity to act on.

If you're responsible for workplace safety and you're curious what computer vision AI could look like for your sites, we'd welcome the chance to show you. Book a demo to talk through your specific situation.

---

Author

Tane van der Boon

Founder & CEO

---

Frequently Asked Questions

What percentage of workplace injuries are preventable?

Research consistently shows that the majority of workplace injuries are preventable. The US National Safety Council classifies most workplace fatality causes as preventable, and OSHA data indicates that hazards like inadequate training, poor traffic management, and lack of real-time oversight are addressable with the right systems in place.

How does computer vision AI prevent workplace injuries?

Computer vision AI uses existing CCTV cameras to monitor work environments in real time, detecting unsafe behaviours and near misses as they happen. Each event is captured with video evidence and turned into a coaching opportunity, allowing safety teams to address risks before they result in injuries.

Does AI safety monitoring compromise worker privacy?

The best implementations process all camera feeds on-site, automatically blur faces and identities, and maintain certifications such as SOC 2 Type II, ISO 27001, and GDPR compliance. The focus is on identifying events, not monitoring individual workers.

What results can organisations expect from AI safety monitoring?

inviol customers see an average 67% reduction in site risk, up to 42% fewer actual incidents over three years, and up to 61% reduction in machine-on-plant incidents across industries including logistics, warehousing, manufacturing, cold storage, and retail.

Does computer vision AI require new cameras or hardware?

No. Computer vision AI platforms like inviol work with the CCTV cameras a site already has. A dedicated AI processing unit is provided as part of the customer agreement, and most sites go live within two weeks.