What data-driven actually means on a factory floor

Descriptive: what happened. Diagnostic: why. Predictive: what will happen. Prescriptive: what to do about it. Vendors sell level four. Most plants are somewhere between one and two, and inconsistently.

The honest sequence is to make level one trustworthy and fast before buying level three. A prediction built on data nobody trusts changes no behaviour.

Start from a decision someone makes on a schedule: which Flows to expedite, which asset to take down, which supplier to call, which line to staff. Then ask what would change the answer, and what evidence would be needed to change it.

If no realistic data would change the decision, do not build anything. That test alone eliminates a large share of dashboard backlogs.

Veteran judgement is usually right and rarely documented. The highest-return exercise in most plants is to sit with the person who 'just knows' and turn their rules into explicit logic — thresholds, exceptions, sequences.

Encoded, that knowledge runs every shift, survives retirement and can be improved by argument rather than repeated from memory.

A metric with no owner and no action attached is decoration. Track how many decisions are now made from the shared picture, how long the daily meeting takes, and how many spreadsheets have been retired. Those move margin; dashboard counts do not.