A model can only work with the inputs it receives. Before evaluating a prediction, make sure the records behind it represent the operation the team thinks it is measuring.
This does not require perfect data everywhere. It requires explicit standards for the data used in a particular decision, and a clear response when those standards are not met.
Agree on what the fields mean
Simple labels can hide important differences. Does “order date” mean creation, confirmation, or payment? Does “stock” include units reserved for another customer? Are cancelled transactions still present in a sales extract?
Write down the definition used for the workflow. Check it with the person responsible for the source system and the person making the operational decision. If their interpretations differ, resolve that before treating the number as a reliable signal.
Check freshness and completeness
A successful data connection does not necessarily mean the information is current. The pipeline may have run while the upstream source was incomplete.
Track both the time a dataset was received and the period it covers. Check whether expected locations, products, or transaction batches are missing. A quiet store and a broken data feed can look similar if only the total is inspected.
Keep the source traceable
When a recommendation surprises someone, they should be able to inspect the records behind it. Preserve useful identifiers and update times so a reviewer can distinguish a source error from a model error.
Traceability also makes correction easier. If an incorrect delivery estimate caused an alert, the team needs to know where that estimate originated and who can fix it.
Define how the workflow fails
An operational tool needs a response to unavailable or unreliable inputs. Continuing with stale data may be acceptable for one decision and unacceptable for another.
Agree on a small set of rules:
Which missing inputs prevent a recommendation?
Which issues can be shown as a warning?
Who is notified when a check fails?
What process can the team use until the issue is resolved?
These rules belong to the workflow design. They should be understandable to the people who rely on it.
Improve one decision at a time
Choose a decision with an identifiable owner and a manageable set of inputs. Establish definitions, checks, and review steps for those inputs. Then evaluate whether the resulting signal helps that owner.
Broader platform improvements can follow from what this work reveals. Starting from a concrete decision makes it easier to distinguish essential data work from work that can wait.
