Better models cannot repair ambiguous operations
An AI system can process a large amount of data and still produce a poor result when the underlying records describe the same asset, customer, or job in incompatible ways.
Five foundations to improve
- Identity: stable IDs for customers, vehicles, equipment, work orders, and locations.
- Ownership: a named system and team responsible for each important field.
- Context: timestamps, units, status definitions, and reason codes.
- Quality checks: detection for missing, stale, duplicated, or impossible values.
- Feedback: a way to record whether a recommendation or workflow helped.
Start with one decision
Do not begin with “clean all data.” Begin with a decision the business wants to improve, such as maintenance prioritization, ETA communication, or parts replenishment. Trace the data required for that decision and fix the smallest broken link.
A reliable data foundation is not glamorous, but it compounds. It makes reporting clearer today and makes future automation safer to evaluate.
Trace one decision end to end
Choose a decision such as maintenance priority, service ETA, or replenishment. List the records required, their identifiers, units, timestamps, owners, allowed values, and acceptable freshness. Then sample real records, including missing and exceptional cases. Data quality is not a single percentage; it depends on whether the data is fit for the decision being made.
Create a small quality contract with checks for duplicates, impossible values, stale records, missing relationships, and unexpected volume changes. Send failures to an owner with a reason and a deadline. Keep corrections traceable so a later model or report can explain which version of a record it used.
Make feedback part of the product
Capture whether a recommendation was accepted, changed, ignored, or disproved, and why. Protect personal and sensitive information through access controls and minimisation. Read the NIST AI Risk Management Framework alongside your own legal and operational requirements; it is a risk-management resource, not a certification or compliance promise.
Build a data dictionary that people use
For each important field, record its definition, unit, source, owner, permitted values, freshness expectation, and example. Include the business meaning of null, unknown, not applicable, and not yet measured. These distinctions prevent dashboards and models from treating missing information as a real zero.
Use representative samples
Review current, historical, corrected, duplicated, and exceptional records. Ask the people who make the decision whether the data supports their reasoning. A data engineer can measure completeness, but the operational owner must judge whether the record is meaningful.
Govern access and change
Keep a record of who can read, change, export, or delete sensitive data. Version important definitions and transformations. When a model or report changes, preserve enough lineage to explain which data and logic produced the result.
