# Model Maintenance at Scale
In industrial AI, the model doesn't stop needing attention after deployment. Process conditions drift: raw materials change, equipment ages, seasonal patterns shift, operators adjust procedures. A model trained on last quarter's data degrades this quarter.
If model maintenance is manual, each customer adds ~0.2-0.5 FTE of ongoing support. At 5 FTEs, you cap at 10-15 active deployments before hiring. This is the hidden scaling ceiling that doesn't show up in pitch decks.
Autonomous model maintenance requires four automated steps:
1. **Drift detection.** Catching when the model's accuracy degrades. Techniques like CUSUM or KL divergence monitoring.
2. **Diagnosis.** Understanding what changed. Sensor drift vs. process change vs. seasonal pattern.
3. **Retraining.** Automatically retraining with new data while respecting safety constraints.
4. **Validation and safe deployment.** Ensuring the retrained model meets safety thresholds before going live, with instant rollback if it doesn't.
Each step is commodity practice. Automating the full loop for safety-critical processes, with no human intervention, is genuinely hard. The orchestration layer that connects detection > diagnosis > retraining > validation > safe deployment might be the real innovation in industrial AI, even though no individual step is novel.
Related: [[Industrial MLOps]], [[Deployment Velocity]], [[Industrial AI Unit Economics]], [[Industrial AI MOC]]
---
Tags: #deeptech #systems