Computer vision is one of the few AI domains where a plant manager can watch the model work. That visibility is a blessing and a curse: a 92% detection rate looks impressive until the 8% misses become customer escapes. ROI appears when vision systems change a physical process—divert, stop, rework—not when they add another TV in the war room.
Use cases with honest payback math
- Final visual inspection on high-volume, stable SKUs with expensive escapes.
- Label and packaging verification before goods issue in the ERP.
- PPE and restricted-zone safety with privacy-preserving poses rather than facial ID.
- WIP location to reduce “lost in plant” expedites.
Simple ROI model inputs
| Input | Where to get it | Notes |
|---|---|---|
| Escape cost | Quality + sales credits | Include freight and brand risk |
| Inspector labor | Std hours × fully loaded rate | Do not assume 100% removal |
| False reject cost | Scrap/rework + line stoppages | Often the silent killer |
| Integration + lighting | OT / MES vendor quotes | Lighting beats a fancier model |
| Model drift labor | Hours/month to relabel | Budget it or the system dies |
Integration is the project
A bounding box that never writes a nonconformance in the QMS or blocks a shipment in the ERP is a science fair. Define the handshake: vision event → MES reason code → quality hold → planner notification. Test it on the graveyard shift when the integrator is not standing there.
People and privacy
Unions and operators will ask who is watched and why. Publish a one-page notice: purpose, retention, who can view footage, and that performance reviews will not be run from raw video without a policy. Trust is a throughput variable.
Conclusion
Computer vision pays off when it is treated as an operations project with MES/ERP write-back, false-reject economics, and a drift budget. Start with one stable line, prove escape reduction, then clone the pattern. That is how factories get AI ROI that survives the next kaizen event.