A packaging plant in northern Chile installed a predictive maintenance model on their conveyor belt system. The model was solid. The data pipeline was not.
The sensors had been logging temperature and vibration data for 14 months, but nobody had audited the logs. Three sensors had been physically replaced during that period, with no recalibration records attached. The model trained on that data started flagging healthy equipment and missing real faults within six weeks of deployment.
The baseline audit problem
Experienced teams often skip the baseline data audit because they assume operations staff would have flagged sensor issues. They rarely do. Maintenance logs and sensor metadata live in different systems, managed by different people, with no automatic reconciliation.
Before any model touches production data, someone needs to cross-reference physical maintenance records with sensor output timestamps. This is not glamorous work, but skipping it produces models that erode trust faster than no model at all.
What the audit should cover
- Sensor replacement and recalibration dates matched against data timestamps
- Gaps longer than 4 hours flagged and explained, not imputed automatically
- Drift patterns in raw readings over rolling 30-day windows
The plant eventually fixed the pipeline, but spent four months rebuilding confidence with the maintenance team who had dismissed the system entirely after the early false positives.
The lesson is not about data quality in the abstract. It is about who owns the physical-to-digital translation layer, and whether that person is in the room when the AI project kicks off. Usually, they are not.