They asked to watch the workers. The process was what nobody could see.
The situation
A multinational apparel manufacturer running at 55% line efficiency against a 60% goal, with each point worth roughly a million US dollars a year. Production data was written by hand, hourly, onto paper sheets and a whiteboard that was erased at the end of every shift. The data went nowhere. Finding a bottleneck took an hour and a walk of the floor. Downtime durations were unknowable because repairs were logged from memory. Managers steered the operation off a weekly email.
The diagnosis
The ask was facial recognition: identify every worker on camera, track individuals, manage by watching. The diagnosis said the problem was pointed the wrong way. Surveillance would have measured people while the process stayed invisible, and poisoned trust on the floor for the privilege. The real problem was that the factory's operating data died nightly on a whiteboard. Nobody could see a stall forming, compare shifts, or know whether a slow station was a machine problem, a materials problem, or a training problem, because the answer was erased before anyone could ask.
The prescription
Cameras as process sensors, not surveillance. Identity handled by badge and QR code only, no faces, by design. And the data treated as two problems at once: a micro problem, what is happening at this station in this minute, and a macro problem, what this plant can promise a customer next quarter.
The build
One camera per two workstations, with edge processing that alerts a supervisor in under 60 seconds when a line stalls. A unified data layer fusing the video signal with the production, HR, and maintenance systems, so for the first time the numbers agree. Defect detection with agreed precision and recall targets and a monthly retraining loop. Zone-based activity sensing for line balance. Three tiers of dashboards in plain language: supervisors who act in seconds, managers who adjust daily, and a managing director who plans capacity on live numbers instead of last week's email.
What the team kept
At the micro level: a stall is visible in under two minutes instead of an hour, mechanics are dispatched against timestamps instead of recollection, and the line can be rebalanced inside the same shift. The station-level data quietly became a people tool too: skill gaps are visible, training is targeted at the stations that need it, and promotion cases are backed by evidence instead of impressions. At the macro level: the managing director plans capacity and accepts orders against a live factory rather than a remembered one, and can compare plants on the same numbers. At three points of recovered efficiency, the program is worth roughly $3M a year, and the whiteboard is gone.