For years, “AI for safety” meant a slide about the future. That has changed. In Verdantix’s 2024 global corporate survey, close to half of organisations said they were already using AI widely, partially deployed and looking to expand, or actively piloting it — roughly ten percent at full rollout, another quarter partially deployed, and a tenth in pilots, against about thirty percent with no plans at all. Adoption is still pilot‑led, but the centre of gravity has clearly shifted from “whether” to “how.” For a safety leader, the interesting question is no longer whether these tools work. It is which ones earn their place, and what makes a deployment succeed or fail once the demo is over.
What computer vision is genuinely good at
The most mature category is camera‑based hazard detection. The credible vendors here — Intenseye and Protex AI are the names that come up most — do something deliberately unglamorous: they connect to the CCTV cameras a site already has, rather than asking you to rip in new hardware. From that existing video, the models watch for a defined set of conditions: missing PPE, people straying into vehicle or exclusion zones, pedestrian‑forklift proximity, blocked exits, ergonomic risk in repetitive tasks. When a rule is breached, the system logs it and alerts — turning a camera that used to be reviewed only after an incident into something that flags the precursor before one happens.
The adopter lists are blue‑chip and span heavy industry: Siemens, Unilever, General Motors, Cummins, Coca‑Cola bottlers, Heineken, JBS, Magna, and the miner Anglo American all appear among named users of one platform or another, and Verdantix has recognised this class of tool as a genuine leader in PPE and ergonomics detection. A reasonable note of caution: many of those names are logo‑wall references, and the depth of deployment behind them is rarely independently verified.
On outcomes, treat the numbers as vendor claims until someone audits them. The cleanest example: Intenseye reports that the bottler Swire Coca‑Cola saw a 27 percent drop in its Lost Day Rate within a year, against a stated goal of halving its Total Injury Rate by 2030 — a target that also shows up in Swire’s own sustainability materials. That is a credible, specific figure, but it is self‑reported and not independently assured. (A flashier “80 percent reduction in incidents in ten weeks” claim that circulates in this space did not survive scrutiny and is best left out of any board deck.) The honest framing for executives is that computer vision reliably increases the visibility of hazards and near misses; the injury‑reduction headline is plausible but should be qualified, not quoted as fact.
Wearables and exoskeletons: real physics, narrow evidence
Wearables split into two families. The first is sensing — devices that detect unsafe motion, proximity, heat or fatigue and warn the worker. The second is assistive — exoskeletons that physically share a load. The assistive case has the strongest underlying physics, and the most overstated marketing.
The biomechanical benefit is real and sometimes dramatic. A 2024 systematic review of passive shoulder‑support exoskeletons found reductions in shoulder and torso joint effort of up to 86 percent, and biceps fatigue reductions up to roughly 68 percent during overhead work. For a worker doing sustained above‑the‑head tasks, that is not marginal.
But two caveats matter enormously for anyone signing a purchase order. First, the evidence base is thin where it counts: of 49 studies in that review, 43 were laboratory‑based and only a handful ran in a real workplace — often on healthy young men doing controlled tasks, which is not your workforce. Second, the devices have documented downsides. The same literature records restricted range of motion, interference with the actual task, and discomfort at the points where the device meets the body, particularly hips and thighs. And there is a metabolic‑cost paradox: by offloading the shoulders, some designs push effort into the legs and can raise whole‑body energy expenditure by around a third in dynamic work. An exoskeleton can genuinely protect a shoulder and still leave a worker more tired at the end of the shift. The lesson is to match the device tightly to a specific repetitive task and trial it on your own people before scaling — not to roll out a category.
The cautionary tale everyone should read: Amazon
No company illustrates the failure mode better than Amazon, and the detail is worth sitting with. In December 2024, a US Senate committee published an 18‑month investigation drawing on more than 1,400 internal documents. Its central finding: Amazon’s warehouses recorded injury rates over 30 percent above the warehousing industry average in 2023, with workers nearly twice as likely to be injured as at comparable employers.
The damaging part was not the rate — it was what the committee alleged Amazon knew. Two internal research efforts, code‑named Project Soteria and Project Elderwand, reportedly found that injuries could be reduced: Soteria concluded that more time off and easing speed‑based discipline lowered injuries, and Elderwand identified an upper limit of around 1,940 repetitive movements over a ten‑hour shift and recommended software‑enforced microbreaks. According to the committee, leadership declined to act on both because the changes would slow delivery pace. This is the archetype of productivity monitoring dressed as safety: the same data that could protect workers used, or shelved, in service of throughput.
Two honesty notes. This was an adversarial, politically‑led investigation, and Amazon disputes it — calling the findings wrong on the facts and the methodology flawed — so it should be cited as the committee’s allegations, with Amazon’s rebuttal attached. But the structural lesson stands regardless of who is right on the numbers: when a monitoring system’s incentives point at speed, “safety” becomes the cover story rather than the goal, and workers can tell.
When safety tech curdles into surveillance
Workers’ instincts about being watched are not paranoia, and ignoring them is the most reliable way to kill a programme. The clearest small‑scale example: at a General Electric aerospace plant in Lynn, Massachusetts, a 2018 pilot of StrongArm’s posture‑sensing wearables collapsed when workers simply stopped wearing the devices — a union steward there feared the movement data would be used to push for faster work or to challenge compensation claims, and called it “creepy as hell.” The company denied any productivity use; the devices came off anyway. The data the system needed to function depended on trust the deployment never earned.
Camera‑based monitoring carries a sharper version of the same risk, because it can capture far more than a sensor — body language, apparent emotional state, the texture of how someone works. A 2025 peer‑reviewed analysis makes two points safety leaders should internalise. One: this kind of surveillance is more intrusive than traditional electronic monitoring, and in lower‑risk settings it tends to erode worker autonomy and wellbeing even when it is technically accurate. Two: the systems cannot read intent. A model can correctly detect that eyes are closed and wrongly conclude a focused worker is asleep. The value is real in genuinely high‑risk environments — mines, tunnels, around heavy mobile plant — but it does not generalise to watching everyone, everywhere, all the time.
The legal perimeter is tightening
Regulation is now a live design constraint, not a footnote. Three lines matter.
In Europe, the EU AI Act’s prohibition on inferring workers’ emotions from biometric data took effect in February 2025. Crucially for safety teams, there is a carve‑out: physical states such as pain and fatigue are explicitly not treated as emotions, and the Commission’s own guidance gives driver and pilot fatigue alerting as a permitted example. So a genuine fatigue‑detection system is generally fine; an “engagement” or “mood” analytic is not. The line is thin, and the system has to actually be detecting a physical state rather than dressing up emotion inference as one.
In the United States, biometric privacy law is the immediate exposure. Amazon has been hit with a 2024 Illinois Biometric Information Privacy Act suit (Johnson v. Amazon) alleging it collected employees’ facial‑geometry scans for timekeeping without written consent, a published retention policy, or disclosure of third‑party sharing — allegations that are unproven and pending, but illustrative of where the litigation is heading. Separately, labour law bites: widespread AI monitoring can chill protected concerted activity under the NLRA, the labour board has limited blanket no‑recording rules, and introducing something like AI smart glasses into a unionised site can itself trigger a duty to bargain. None of this makes monitoring unlawful. It makes consent, transparency, and consultation legal prerequisites rather than nice‑to‑haves.
What separates the wins from the wreckage
Across every case, the dividing line is not model accuracy or sensor quality. It is whether the workforce experiences the system as protection or as a leash.
The single most repeated lesson in the practitioner literature is almost embarrassingly simple: be explicit that the tool exists to protect workers, not to monitor them, and involve workers from the very beginning rather than presenting a finished system. But there is a vital qualification — that framing is necessary and not sufficient. “Protect, not monitor” printed on a poster while the same feed quietly scores productivity is worse than saying nothing; it reads as control disguised as care, and it destroys the trust the next safety initiative will need. The framing has to be backed by real function: data that genuinely serves the worker, alerts that are explainable, a human in the loop on anything consequential, manual overrides, and a hard wall between safety data and performance management.
So the practical test for any deployment is a short one. Does it surface hazards the worker would want surfaced? Is the data walled off from discipline and pace‑setting? Were the people being watched in the room when it was designed? Can a person override the machine? If the answers are yes, computer vision and wearables can be a real early‑warning layer — the kind of leading‑indicator signal most programmes lack. If the answers are no, the technology will work exactly as advertised and the programme will still fail, because the people it depends on will switch it off, sue over it, or simply stop trusting it — and they will be right to.
A note on sourcing: vendor outcome figures in this piece (for example the 27 percent Lost Day Rate reduction) are self‑reported and not independently audited; the Amazon findings are the allegations of a US Senate committee that Amazon disputes; and the Illinois BIPA claim against Amazon is pending and unproven. Exoskeleton efficacy figures are largely from laboratory studies. Regulatory points (EU AI Act, BIPA, NLRA) are time‑sensitive and jurisdiction‑specific. Treat the directional lessons as durable and the specific numbers as claims to verify against your own context.