A steel processing facility in the Atacama region completed a 9-month AI project to reduce surface defects in rolled coil. The system was integrated into their MES, alerts were configured, and the project was formally closed.
Defect rates did not change for the first quarter post-deployment.
A process audit revealed that the AI alerts were appearing in a monitoring dashboard that shift supervisors checked once per shift, at the start. By the time a supervisor reviewed the alert, the relevant coil had already moved three stations down the line.
Where the integration spec missed the point
The integration team had delivered exactly what was scoped: alerts in the MES dashboard. Nobody had mapped the actual information flow during a shift. The supervisors were not ignoring the system. They were operating within a workflow that made acting on the alerts physically impossible given the timing.
This is a handoff problem between the AI project team and the process engineering team. Both groups assume the other has validated the operational fit. Neither does, because it falls outside both mandates.
What a workflow validation step looks like in practice
- Shadow two full shifts before go-live, tracking where decisions actually happen and when
- Identify the latest point in the process where an alert can still change an outcome
- Confirm that the alert delivery mechanism matches that timing window
The facility eventually routed critical alerts to floor terminals at the rolling mill entry point. Response time dropped from 6 hours to under 4 minutes. The model had not changed at all.
Technical delivery and operational fit are two separate sign-offs. Treating them as one is where most post-deployment failures originate.