Measure data quality before connecting business systems
Identify the fields that drive decisions, distinguish missing from incorrect data and fix quality problems at their source.
Data quality is about whether information is suitable for the decision a workflow needs to make. A CRM can contain every required field and still route customers incorrectly. Before connecting systems, identify the important fields, define what acceptable data means and decide how the workflow should handle values that do not meet that standard.
Prioritize the fields that affect actions
Trace each automated decision back to its inputs. A customer identifier, delivery address or approval status may require stricter checks than an optional description. Document the owner and authoritative source for each important field. Avoid treating whichever system was updated most recently as correct when the systems have different responsibilities.
Check several dimensions rather than relying on one completeness percentage. The UK Government’s framework distinguishes completeness, uniqueness, consistency, timeliness, validity and accuracy. For your workflow, select the dimensions that change the decision. A valid-looking email address can still belong to the wrong person, while a correct address can become outdated.
Make missing and conflicting data explicit
Define how the system represents unknown, not applicable, unreadable and intentionally withheld values. Do not convert all of them into an empty string or a default category. Downstream rules should be able to distinguish a field that needs clarification from one that is legitimately absent, without asking a model to invent an answer.
Create rules for conflicts between systems and preserve the values being compared. Some conflicts can be resolved through an authoritative source; others need a reviewer. AI may suggest a normalized description or identify suspicious records, but it should not silently rewrite important identifiers, dates or customer preferences to make the dataset appear cleaner.
Improve the source and monitor the result
Use a representative sample to find recurring problems and estimate the effort needed to correct them. Record examples, affected fields and the process that produced the issue. If a form collects ambiguous dates, improve the form rather than building increasingly complex interpretation rules around the same avoidable ambiguity.
After introducing validation, monitor rejected and corrected records as well as successful ones. Overly strict rules can block legitimate customers or force staff to enter fake placeholder values. Review exceptions with the people who collect and use the data, and update definitions when the underlying service or customer information changes.
Practical checklist
- Identify fields that directly affect automated actions.
- Assign an authoritative source and owner to each.
- Represent missing and conflicting values explicitly.
- Track corrections and improve collection at the source.
Illustrative setup: customer location data
A service business routes enquiries by postcode, but one channel stores a full address in the postcode field. The team introduces a dedicated postcode input and a review state for older records. It does not guess locations from customer names or force ambiguous records into the nearest-looking service area.
Common questions
Is a complete database ready for automation?
Not necessarily. Completeness means required information is present, while accuracy concerns whether it matches reality. Check the fields and quality dimensions relevant to the planned action.
Should all historical data be cleaned first?
Focus on the records and fields the pilot actually uses. Define how older or unsupported records are handled, then expand cleanup where it provides a clear operational benefit.
Further reading
Start with your actual workflow.
Turn the useful parts of this guide into a focused project brief.
Shape your project