Before You Trust AI With Your PI Data, Clean It Up

AI can summarize, classify, detect patterns, and answer questions about industrial data. It cannot correct an unknown source problem simply because the model is advanced.

Before an AI workflow uses PI data, teams should validate the data path and the context that gives the values meaning.

Start with source reliability

Check whether important signals are:

• Current

• In valid PI states

• Mapped to the intended source

• Expressed in known engineering units

• Collected at suitable fidelity

A model cannot infer that a pressure tag was remapped to the wrong transmitter unless the system provides that context.

Validate calculated data

AI workflows often consume derived KPIs rather than raw signals.

For these values, know:

• Calculation purpose

• Inputs

• Output

• Schedule

• Bad-data handling

• Change history

A stale or incorrectly configured calculation can produce plausible numeric values that are still wrong.

Provide lineage

Lineage lets teams trace an AI input back to the source and identify important transformations.

It also supports impact analysis when a source or calculation changes.

Without lineage, a data scientist can see the value but not the chain of assumptions behind it.

Provide asset and unit context

A tag name alone is rarely enough for AI or analytics.

Provide asset hierarchy, measurement meaning, engineering units, source, and relevant equipment metadata.

This reduces ambiguity and improves the ability to validate model output.

Separate data errors from model errors

When an AI result is wrong, investigate both the model and the data.

Ask:

• Was the source data valid?

• Did the model use the correct time range?

• Were units consistent?

• Did a configuration change alter the input?

• Did the model make an unsupported inference?

Keep a human verification step for consequential use

AI output used for engineering, safety, maintenance, or regulatory decisions requires appropriate review.

The model can accelerate analysis. It should not remove accountability.

The objective

AI readiness is not the same as perfect data.

It means that important data is sufficiently reliable, contextualized, and traceable that teams can understand why the model received a value and can investigate when the result is questionable.