Section 1
The gap between the signal and the wrench
A predictive maintenance model that emails a dashboard link to a maintenance supervisor has not automated anything. It has added a notification to a person who already had too many. Value appears only when the prediction connects to the systems that control the response: the CMMS that holds the work order, the inventory system that holds the spare, and the scheduler that holds the line time. That is why plant automation projects should be judged on how far the workflow reaches, not on model accuracy. A mediocre prediction wired into a work order beats an excellent one that ends in a chart. The same logic runs through connected equipment generally, covered in [AI and IoT: Smart Devices, Smarter Businesses](/blog/ai-and-iot-smart-devices-smarter-businesses).
Section 2
Where it pays on a plant floor
Three areas repay the effort. Unplanned downtime, where the cost per hour is already known to finance and the case writes itself. Quality inspection, where vision systems catch defects earlier than a downstream check and stop you shipping scrap. Changeover and scheduling, where a plan that reacts to actual cycle times rather than a standard rate recovers hours nobody had budgeted. A fourth area is quieter but often larger: the paperwork around production. Shift handover notes, non-conformance reports, supplier quality correspondence, and compliance documentation are drafted by people who should be on the floor. Automating that documentation does not sound like a smart factory. It frees the people who make one work.
Section 3
Safety systems are out of scope
Draw this boundary explicitly and early. Interlocks, emergency stops, guarding, and anything governed by a functional safety standard stay deterministic and stay out of any model-driven path. Advisory outputs may inform a human operator. They may not actuate. Plants that blur this line do not get a warning first.
Section 4
Piloting on one line, not the plant
Pick one asset with a known failure mode and a known cost of stoppage. Establish the baseline from maintenance history: how often it fails, how long it takes to recover, what the recovery costs in lost output. This is usually the hardest step, because the history lives in a mix of a CMMS, a spreadsheet, and a supervisor's memory. Run the model in advisory mode across at least a full maintenance cycle. Log every alert and what actually happened next. If it never predicted a real failure and generated forty false alarms, you have learned something cheap. Only after the alert quality is credible should the workflow start creating work orders on its own, and even then a planner approves the schedule.
Section 5
OT, IT, and who owns the failure
Governance on a plant floor collapses into one word, access, which is where the NIST emphasis on design and use actually bites. What can the system read from the control network. What can it write. What sits behind a one-way boundary. Manufacturing has spent two decades keeping operational technology separated from corporate IT for good reasons, and an automation project is a common way that separation quietly erodes. Name one owner per automated workflow, someone in operations rather than in a data team, and give them the authority to disable it. Log every automated action against the asset so a stoppage can be reconstructed. Assume the vendor will change the model at some point and ask in advance how you will find out.
Section 6
Metrics the plant already trusts
Resist inventing new measures. The plant already runs on overall equipment effectiveness, unplanned downtime hours, mean time between failures, scrap and rework rate, on-time delivery, and inventory turns. Attribute automation to those numbers or admit it did not move them. Add two operational checks. Alert precision, meaning what share of flagged issues turned out to be real, and time from alert to action, which tells you whether the wiring works. Review monthly with maintenance and production in the same room. If the automation improved a metric on a dashboard but the supervisors do not believe it, the honest reading is that it did not work.