Your problem is not AI — nobody trusts the data

Every time I walk into a plant asking for "AI", week one turns into the same conversation: nobody knows where the number they already look at comes from.

The three questions

They are simple, and I use them as a filter before accepting any advanced analytics project.

1. Who produces this value? A sensor, an operator typing into a spreadsheet, an intermediate calculation in another system. Those three have wildly different reliability.
2. How often and how precisely? A value every 5 seconds at ±0.1% and an eyeballed value per shift do not support the same decisions.
3. What happens when it is missing? Interpolated, last-value-held, left as a gap? The answer completely changes how any average should be read.

If none of the three has a written answer, any model on top amplifies the mess instead of fixing it.

The good news

Fixing this does not take a year. Minimal instrumentation, a context model (line, shift, order, product) and a trustworthy historian usually make 80% of decisions better without a single neural network.

The order is always the same:

• Signal inventory with the real source, not the one on the drawing.
• A context model agreed with production, not with IT.
• One historian, with defined retention and a tested restore.
• Only then: dashboards, alerts and, if the case justifies it, models.

That fourth step is the only negotiable one. The first three are not.