Workstation ergonomics with AI — anticipating the musculoskeletal disorder before the sick leave.

The problem

The postural assessment arrives when the operator is already off work.

Industrial workstation ergonomics is still assessed the way it was twenty years ago: the prevention technician visits the workstation, observes for an hour, fills in the RULA or REBA sheet, proposes an adjustment, files it. The method is sound — but the classical procedure has three gaps that AI can close:

1. The observation is a snapshot. One hour, one week a year, on the morning shift. The workstation is worked 8 hours, 3 shifts, 220 days — and the postural peaks never make it into the photo.
2. The operator adapts their posture when they are being watched. What the technician observes is not the real posture of an ordinary shift.
3. Rotation changes the risk. A workstation that is fine for a tall person can be hell for a short one. And the shift change reassigns people without the matrix ever being updated.

The result: the MSD (musculoskeletal disorder) remains the number-one cause of sick leave on the plant floor — and it is diagnosed once the sick leave has already started, not before. The baseline is broken, and without a baseline there is no way to prove any improvement.

How it fits the IRIS system

iLEAN gives continuous observation to what prevention used to assess with a snapshot.

The problem is not a missing method (RULA and REBA have worked for decades), it is missing continuity of observation. iLEAN provides that continuity without turning the plant into a panopticon: the Edge piece analyses posture as a pose-estimation skeleton, storing no faces, tying nothing to individuals — only to the workstation. It is the filler that closes the crack between the annual assessment and the reality of the day.

Edge sees the skeleton, not the face. Connect hears the operator say "my lower back is hurting today". The agent cross-checks with sick leave and proposes. The person signs the change.

The three iLEAN pieces applied to ergonomics:

• Edge — a terminal with vision and a CNN over the workstation. It runs pose estimation locally: it extracts the joint silhouette, computes angles on every frame, and aggregates them into continuous RULA/REBA across the shift. It records no video of people: what persists are the angles and the timestamp of the peak. It works with no network — if the plant loses WiFi, it keeps measuring.
• Connect — the operator reports discomfort by voice from the headset or the phone ("my back is heavy today"). The medical service gets the alert at second zero, with no form. And it also captures what comes from outside: the insurer's report, the physiotherapist's recommendations, the return of a lift assist that arrived new.
• Agent — it cross-checks the workstation's postural peaks with the sick-leave history, rotation, takt and the average height and weight of the assigned team. It proposes adjustments with a reasoned priority — add a lift assist, redistribute micro-tasks, schedule an active break. The prevention technician validates and signs; the line never reorganizes itself.

See the full IRIS architecture →

Before and after

Classical ergonomics vs. continuous ergonomics with iLEAN

Aspect	Classical RULA/REBA assessment	With iLEAN Edge + Connect + Agent
Frequency	Annual, or after an incident	Continuous across every shift
Workstation coverage	One isolated hour	Every real hour of the shift
Operator privacy	Video of the workstation, face included	Pose-estimation skeleton, no face, no identity
Cross-check with sick leave / insurer	Manual, after the fact	Automatic, in the live workstation dossier
Detection of extreme peaks	Only if it happens while the technician is looking	Every peak across the year is on record
Time to improvement proposal	Weeks, depending on the calendar	Prioritized proposal within hours

Impact estimate

Impact estimate for your plant — to be validated against your numbers.

The block below is an estimate to be validated with the specific data of your plant. We set it out so that management and prevention have an order of magnitude; we refine it during the diagnostic.

• A plant of 200–500 operators, 30–80 workstations with repetitive manual tasks, a history of 2–5 MSD sick leaves a year across the family of workstations to be addressed.
• An Edge pilot over 3-5 critical workstations plus the cross-checking agent. First value expected within a few weeks: a real postural risk ranking per workstation and a proposal of priority adjustments.
• Indicative payback between 6 and 9 months, on the conservative side. Levers: an expected reduction of MSD sick-leave days of ≥30% (estimate), the insurance premium over the medium term, and the productivity of the redesigned workstations.
• The hard lever: a single long sick leave avoided pays for the pilot.

And the works council's fair objection

"Isn't this about watching the operator?" — quite the opposite. What Edge keeps is the skeleton of the workstation, not the face of the person. No identity is ever attached to a joint angle. The system measures the workstation and hands prevention the data so people can work better, not so someone can be written up. This is process verification, not people: fight the problem, not the person. Continuous ergonomics is the best defence tool an operator has — the first time the system says "this workstation has a lumbar-twist peak at the fixture changeover", the insurer has it as evidence. And hallucination is a problem of free-form generation, not of anchored tasks such as scoring a joint angle against a RULA/REBA scale: on anchored tasks the best models brought the error rate below 1.5% ^[1].

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.