Imperfect data is a starting point

“Our data is too messy” can be an accurate diagnosis and an unhelpful stopping point. Asset identifiers differ across systems. Important details live in scanned reports. Work orders use inconsistent descriptions. None of that makes the information valueless. It makes the effort of assembling it the problem to solve.

AI creates another way to approach that effort. It can propose relationships between records, extract candidate fields from documents, and present inconsistencies in a form a domain expert can review. The starting question becomes: which useful workflow can we improve while making its supporting data better?

Recover the evidence you already have

Imagine an illustrative utility with a maintenance spreadsheet, a GIS asset layer, and years of inspection reports. A planning team repeatedly searches all three to understand a small group of assets.

An initial workflow could gather likely matches and show the original evidence together. It need not settle every ambiguity. Even an unresolved conflict is more useful when the relevant records and reasons are presented to the person who can investigate it.

This is why entity matching matters. A new dashboard adds little if each underlying source still describes a different version of the same asset. Connecting identities, preserving provenance, and making uncertainty visible can support several future workflows.

Improve the record without inventing the asset

There is a large difference between extracting a rating from a readable inspection image and estimating it from nearby equipment. Both may have a place in analysis, but they deserve different labels and different uses.

A record should preserve whether a value was observed, imported, inferred, or verified. An inference should not quietly become an authoritative engineering attribute because it fills an empty cell.

The same applies to duplicates. A plausible match can be wrong, and merging unrelated records can hide the mistake. Keep the evidence, test matching quality, and provide a way to review and undo corrections.

Change what the project is buying

A proposal to “clean all the data” is hard to finish and hard to evaluate. A proposal to “reduce the time needed to assemble inspection evidence for capital planning” gives the team a clearer outcome.

The supporting work may still involve extraction, normalization, matching, validation, and field verification. The difference is that priorities come from the decision being supported. The team can explain which defects matter now and which can wait.

Measure the change in the workflow as well as the dataset: accepted corrections, fewer blocked tasks, less searching, and the cost of handling exceptions. Track recurring defects so new imports do not erase the improvement.

Start with a manageable slice

Pick one asset class, geographic area, document collection, or recurring task. Preserve originals. Define the quality needed for the intended use. Let AI prepare the evidence and proposed corrections; give the responsible team a clear route to validate and approve them.

That creates something more useful than a one-time cleanup exercise: a repeatable way to improve information as work happens. The data quality guide shows how to build that loop, from identifying an issue through verifying the correction.