€1.2M of sensors, two points of OEE

The upgrade was the biggest capital project the site had run in five years. Sensors on every line. A real-time OEE dashboard on screens mounted above each cell. A new MES module promising paperless batch records and live downtime tracking. The vendor's business case put a fifteen-point OEE lift in year one.

Eight months in, OEE had moved from 61% to 63%. Two points. The screens were on, the data was streaming, and on the floor nothing had changed — supervisors still ran the morning meeting off a whiteboard and a feel for which line had hurt them yesterday.

Finance wanted to know why €1.2M of sensors had bought two points of OEE. The honest answer wasn't on any of the new dashboards.

Following the data backward

The reflex in the room was the predictable one: the platform was wrong, or the install was under-scoped. The vendor had already been asked to quote a phase two — more tags, more integration, a bigger historian. The working assumption was that they hadn't bought enough Industry 4.0.

Walking the data backward told a different story. The OEE dashboard pulled stop events straight from PLC signals, accurate to the second. But stop reasons were still keyed in by operators at the end of each shift, from memory, into a spreadsheet that lived nowhere near the dashboard. The system could tell you a filler had stopped for 47 minutes across a shift and not one reliable thing about why. Quality results sat in the LIMS. Maintenance history sat in a standalone CMMS. Neither was connected to the OEE layer — and that layer was the only thing anyone had been shown how to use.

Then the tags. Close to 40% of the new sensor points were never mapped into the historian; they streamed into a buffer nobody queried. The predictive-maintenance module sold as the headline feature had no clean failure history to learn from, so it threw false alerts almost daily. Within a month maintenance muted it and went back to the run-to-failure they'd always run.

The plant hadn't bought too little. It had bought the wrong rung. Data was now recorded in more places, to finer precision than ever — and almost none of it was connected. There was no single trustworthy picture of line state, which meant nothing built on top of it could work. They had installed a predictive capability and were still running at the level of a paper logbook with better handwriting.

A ladder, not a menu

An Industry 4.0 maturity model gets read like a catalogue: pick the capabilities you want, buy them in any order. It isn't a catalogue. It's a ladder, and the rungs are load-bearing in sequence. The version I use for food plants runs five levels.

The order is not a suggestion. You cannot see what you haven't connected. You cannot predict what you can't see. You cannot let a line self-correct on a signal you don't trust. The plant bolted Level 4 onto Level 1, and Level 4 did what it always does on disconnected data: it produced confident noise, fast.

This gives you a sharper question to run at the next capital review. Stop asking "what's the most advanced system we run?" — that question rewards whoever bought the shiniest platform, and it's how you end up installing Level 4 on a Level 1 floor. Ask instead: "where does our data still break?" The answer is rarely at the top of the stack. It's a stop reason typed from memory, a quality result that never leaves the LIMS, a sensor tag streaming into a buffer nobody opens. Find the lowest layer that isn't trustworthy yet — that's the rung you're actually standing on, whatever the logos on your dashboards say. Build there. Every level above it inherits what you fix, and inherits what you don't.

Audit the rung below the spend

Before you sign off the next "predictive" or "AI" line item, audit the rung beneath it. If your stop reasons are still keyed from memory at end of shift, you are not ready for predictive maintenance — you are ready to connect your stop-reason capture, and that cheaper fix is what makes everything above it possible.