In January, a safety manager at a mid-sized logistics firm told us he was tired of demos. Every vendor promised a miracle, every pilot fizzled, and every board meeting ended with the same question: where is the actual reduction in incidents. By December, that same manager had a wall-mounted screen showing live near-miss counts by aisle, and a monthly report his insurer had started asking to see. Nothing about the technology had changed in twelve months. What changed was that it finally started keeping its promises.
That, in a sentence, is the story of 2025. Safety AI stopped being a science project. It grew up.
The year the pilot ended
For most of the last decade, computer vision safety lived in a permanent state of "pilot". A camera in one corner of one warehouse. A proof of concept that ran for six weeks and quietly died when the champion changed jobs. The technology worked in the demo and struggled in the rain.
2025 was the year that stopped being acceptable. The conversations we had this year were not about whether the detection worked. They were about integration, roll-out schedules across multiple sites, and how the alerts fit into an existing permit-to-work process. The centre of gravity moved from the innovation team to operations. That is what maturity looks like: when the interesting question is no longer "does it see the forklift" but "who acts on the alert, and how fast".
The buyers grew up too. A year ago, the first question was often "what can your AI do". This year it was "prove it did that on a site like mine". Reference customers, error rates, false-alarm figures, and on-premise deployment became the qualifying questions, not the nice-to-haves.
Three milestones that mattered
If you want to mark the year by its genuine turning points rather than its press releases, three stand out.
From detection to leading indicators
The first was a quiet shift in what the technology is for. Early systems counted incidents. Mature systems in 2025 counted the behaviours that precede incidents — the near-misses, the pedestrian who crossed a vehicle lane, the worker who entered a zone without a helmet and left before anyone noticed. For the first time, a large number of EHS teams had a continuous stream of leading indicators rather than a lagging tally of things that had already gone wrong. That is the difference between a smoke alarm and a weather forecast.
Privacy stopped being an afterthought
The second milestone was privacy moving to the front of the conversation. As adoption widened, so did scrutiny — from works councils, from data protection officers, from workers themselves. The market responded. On-premise processing, where footage never leaves the site and no video is streamed to a third-party cloud, went from a differentiator to an expectation in serious procurement. The systems that treated privacy as a design principle rather than a compliance chore were the ones that got signed off.
The proof bar rose
The third was the arrival of real, boring, verifiable proof. Case studies stopped being glossy and started being specific: measured error rates, documented reductions in unsafe behaviour, deployments you could telephone and check. The industry has spent years selling potential. In 2025 the credible players started selling evidence.
It is worth being honest about our own vantage point on that last one. Our detection was not born in a lab. It was forged offshore, on drill floors — heavy moving equipment, zero tolerance for error, lives at stake every shift. From there it went into the operations of a national oil major, a major international port and an international airport, running at a sub-0.05% error rate and delivering field-measured reductions of around 90% in unsafe behaviour. We mention it not to boast but because it explains why the "prove it" year felt less like a threat and more like a relief. The industry finally started asking the questions we had already been answering.
What still needs to grow
Maturity is not the same as finished, and it would be dishonest to pretend 2025 solved everything.
Alert fatigue remains the quiet killer of good systems. A camera that cries wolf forty times a shift gets ignored, and an ignored system is worse than none because it breeds false confidence. The best deployments this year were ruthless about tuning — fewer, better alerts that people actually trust.
Integration is still harder than it should be. Detecting a hazard is only useful if the alert reaches the right person through the channel they already use, whether that is a control room, a supervisor's radio, or a daily report. The vendors who understood that the AI is the easy part did well. The ones who shipped a clever model and a clumsy workflow did not.
And the human question is not going away. Safety AI does not replace a safety professional any more than radar replaced the pilot. It extends their reach — turning a handful of cameras nobody watches into a tireless second pair of eyes that never blinks, never gets bored, and never looks away at the wrong moment. The organisations that framed it that way, as augmentation rather than surveillance, got the buy-in. The ones that framed it as a way to catch people out did not.
Where 2026 points
If 2025 was the year safety AI grew up, 2026 is the year it goes to work. Expect the conversation to move further from "can it see" to "what does it change" — fewer demos, more before-and-after numbers. Expect insurers and regulators to lean in as continuous monitoring produces the kind of leading-indicator data that risk models have always wanted and never had. And expect the quiet specialisms — dropped objects, confined spaces, moonpool monitoring in the offshore world — to matter more, because the easy detections are now table stakes and the hard environments are where the real value sits.
The technology stopped being the story this year. What it prevents is the story now. That is exactly as it should be.
2025 in numbers: what the deployment data showed
Near-miss reduction at production-scale deployments
The most significant data point from 2025 was not from a controlled pilot — it was from the ongoing operational data at production-scale deployments that have been running long enough to show multi-year trends. Sites that deployed AI safety monitoring in 2023 are now able to show a three-year near-miss trend: a period of baseline data collection, a period of active detection and alert response, and a period of sustained reduction following engineering changes informed by the monitoring data. This is the proof-of-concept that early adopters needed and that the wider market has been waiting for.
The categories that reduced fastest
Across the deployment base, vehicle-pedestrian near-miss rates and PPE compliance failure rates showed the fastest and most consistent reduction. Both of these are categories where the deterrence mechanism is immediate and the monitoring is continuous — a driver who receives an alert for a proximity violation changes their approach within days. Zone breach rates and speed violation rates showed more gradual reduction, consistent with the longer behavioural change cycle for ingrained habits.
The false positive improvement curve
Alert quality improved significantly across the deployment base in 2025 as the calibration process matured and site-specific model fine-tuning became more systematic. The average false positive rate across production deployments fell over the year as the accumulated site-specific training data improved model accuracy for each environment. This improvement is not automatic — it requires active calibration reviews and feedback from the operations teams — but it is measurable and consistent with the expected learning curve.
Implications for 2026 safety programme planning
The data from 2025 production deployments provides the evidence base that safety teams need to build 2026 programme budgets around real outcomes rather than vendor projections. For organisations that deployed in 2023-2024, the 2026 planning cycle should include a quantified near-miss reduction ROI calculation from the actual deployment data. For organisations that are still evaluating, the 2025 production data from comparable sites is the most reliable proxy for what deployment would deliver on their own sites.
See what your cameras have been missing — book a demo.
