ProductionVerified Real deployment at a named or precisely described company, with a measured result from a primary source.

AI camera analytics on existing CCTV spot forklift near-misses, missing PPE and speeding so safety teams can coach (Carlex, Epicurean Dairy)

Added 2026-11-09 · Last updated 2026-11-09

Manufacturers in this case: Carlex · Epicurean Dairy · Smurfit Westrock · all manufacturers using AI

Role

Plant EHS / health and safety (HSE) and operations management · Typical users: EHS or HSE manager, plant manager, warehouse and production supervisor, union safety representative

The problem

Forklifts, pedestrians and manual work share the same floor, and a safety team cannot be everywhere on every shift. Cameras usually only get reviewed after an incident, when time has been lost and the root cause is hard to piece together. At Carlex, an automotive glass maker with four plants and over 1,400 employees, new leaders had no objective data on where to focus a behaviour-based safety programme, and finding leading indicators would have required far more floor observation than the team could do. At Epicurean Dairy (The Collective yoghurt, Auckland, New Zealand), several neighbouring sites with forklifts, vehicles and pedestrians made vehicle safety one of the biggest risks, and the old approach was to review footage after an incident or near miss. At Smurfit Westrock's Lübbecke plant (paper packaging, Germany), forklift drivers moving paper rolls could not see pedestrians behind them.

How AI helps

Computer vision models run on the plant's existing cameras and detect patterns that matter for safety: forklift-pedestrian proximity inside defined zones, high-visibility vests or other PPE that are missing, vehicles exceeding a speed limit, vehicles that do not stop at the end of an aisle or at a door. Events go to a dashboard with video clips, so the safety team sees where and when risk is concentrated, then changes the layout, signage or training and coaches people one to one with the footage. Both Carlex and Epicurean Dairy set the system up as non-punitive: Carlex agreed this with the UAW union and uses it for information gathering and training only; Epicurean blurs faces, with the stated aim 'identifying problems, not exposing individuals'. At Smurfit Westrock, Trio Mobil's TRUE AI takes a different approach: cameras detect pedestrians without tags in defined danger zones and trigger visual and acoustic alarms on nearby forklifts, with no alarm when pedestrians are in safe areas. A viso.ai customer, a European industrial manufacturer, deployed six detection modules (speeding, PPE, proximity, blocked exits, stocking-area overflow, restricted areas) on existing cameras with on-edge processing and face anonymisation.

Step by step

  1. Start from your own data: list the last 12 to 24 months of forklift, pedestrian and PPE incidents and near misses, and mark them on a floor plan, so you know which two or three risks to detect first.
  2. Talk to workers and, if you have one, the union or safety committee before any camera analytics: agree in writing that it is for coaching and not discipline, who can see footage, and how long it is kept (check local privacy law).
  3. Choose a pilot area (one warehouse aisle, one dock, one production cell) and check that the existing cameras cover it at a usable angle and resolution.
  4. Configure only the detections that match your top risks, such as exclusion zones around forklifts, PPE in defined zones and a speed limit, and ask the vendor to blur faces.
  5. Review the event dashboard weekly with line and warehouse supervisors, and act on patterns (a blind corner, shift-change PPE lapses, an intersection that needs a stop sign) before individuals.
  6. Measure against your own baseline from step 1: events per week, near misses, PPE compliance and time spent investigating, and report false alerts as well as the wins.
  7. Extend to other areas only after the pilot has worked and the workers have seen the results.

Tools

  • Voxel (computer vision on existing cameras for forklift and PPE safety): used at Carlex
  • inviol (AI on existing CCTV for vehicle-pedestrian proximity, PPE and speed, with a coaching app): used at Epicurean Dairy
  • Trio Mobil TRUE AI (tagless pedestrian detection with alarms on forklifts): used at Smurfit Westrock
  • viso.ai Health and Safety Suite (AI vision modules on existing cameras, on-edge processing, face anonymisation): used at a European industrial manufacturer
  • Other tools to evaluate: Visionify, the video analytics offered by your camera or forklift telematics supplier, tag-based proximity systems for comparison
  • To think the programme through without new software: your incident and near-miss log and floor plan, plus any LLM (ChatGPT, Claude, Copilot, Gemini) with the prompt below

In the AI tools index: inviol · Trio Mobil TRUE AI · viso.ai Health and Safety Suite · Voxel · all AI tools used by manufacturers

Result

Carlex (automotive glass, Vonore, Tennessee plant), Voxel: 12 existing cameras, in under three months safety vest compliance up 86%, no-stop incidents at the end of aisles down 47% and at doors down 37%; the plant had been live for six months at publication and Carlex was preparing to install the platform in its other plants (Voxel customer story, a vendor publication; no incident or injury rate published). Epicurean Dairy (New Zealand), inviol: a 48% drop in overall risk from peak levels after coaching based on 3 m and 1 m exclusion zones, PPE detection and speed monitoring, and weeks without any one-metre breaches, which the general manager said they had not seen before (inviol blog post, 1 November 2025, a vendor publication). European industrial manufacturer (not named), viso.ai, 90-day deployment on existing cameras: near-miss proximity events down from an average of 45 to 15 a month (-66%), speeding-vehicle incidents down 52%, PPE non-compliance down 41%, blocked exits and aisle obstructions down 71%, investigation workload down 63%, a 2.1x return on investment within 12 months and three more plants added in under six weeks each (viso.ai customer story, vendor-reported; the same page's summary box says 68% fewer unsafe events, while the text says 66%). Smurfit Westrock (Lübbecke), Trio Mobil TRUE AI: rolled out to more than eight Smurfit Westrock plants after a test; the vendor's one-page story says it prevented forklift-pedestrian accidents and publishes no measured figures (and gives both '10 facilities' and 'eight plants').

As reported by the source; most figures are vendor or self-reported and not independently verified.

Human review: cautions

  • All results are published by vendors (Voxel, inviol, viso.ai, Trio Mobil); none gives an injury or lost-time rate, and the viso.ai customer is not named. Treat the figures as what the vendor reports.
  • Workers can feel watched: both named plants made the programme non-punitive and blurred faces. Without that trust the system is likely to be resisted, and works councils or unions may have to approve it. Check privacy law where you operate.
  • Counting fewer events can mean people avoid the camera zones, not that behaviour changed: also check audits and incident reports.
  • Camera analytics detects risk; it does not replace guards, speed limiters or physical segregation of forklifts and pedestrians.

Try it yourself

You are an EHS (health and safety) manager at a manufacturing plant preparing a pilot of AI camera analytics for forklift and pedestrian safety.

I will paste: (1) our incident and near-miss log for the last 12 to 24 months [date, area, what happened, severity]; (2) a description of the floor plan and traffic flows [aisles, docks, crossings, blind corners]; (3) our PPE rules by zone; (4) what cameras we have today and where.

Do this:
1. Group the incidents by area and type, and list the top three risks with the evidence from my log for each.
2. For each risk, say which detection would address it (proximity zone, PPE by zone, speed, no-stop at an aisle end or door) and what we should measure before and after.
3. Suggest one pilot area and why, and which checks I should make on camera coverage there.
4. Draft a one-page note for workers and the union or safety committee: purpose (coaching, not discipline), what is recorded, who sees it, how long it is kept, face blurring.
5. Write 10 questions to ask vendors (where the video is processed, accuracy and false alerts in a plant like ours, privacy features, integration with our incident system, price per camera, reference plants).

Do not give legal advice: list the privacy and labour-law points I should check with our legal team. Use only the data I paste in; if something is missing, write 'not found' or list it as a question instead of guessing. Do not send, order, post or change anything in any system: produce drafts for my review and approval.

Want more like this? Join the newsletter

Related cases

All 21 Production use cases →

Sources

Weekly: 1–2 practical AI use cases for your plant

Newsletter in

Free, one email a week. By subscribing you agree to receive the weekly newsletter (and its sponsor message) from AI Agents for Manufacturing. Every email includes an unsubscribe link, and you can unsubscribe anytime. See our Privacy Policy.