Is matter data ready for AI and BI?

A firm can spend months pursuing abstract data cleanliness without knowing which defects matter. Define the operating question first, then list the minimum fields, events and relationships required to answer it. A stalled-matter review may need a stable matter identifier, status, responsible person, selected activity timestamps and agreed exclusions; it may not need the full client file. For organisations subject to the Privacy Act, APP 3 considerations include whether collection of personal information is reasonably necessary for the entity's functions or activities and, for sensitive information, additional requirements may apply. The exact analysis depends on the organisation and data. Minimisation is therefore both a governance and engineering discipline. It reduces unnecessary exposure and makes field-by-field validation possible. If the decision cannot be supported without broad speculative collection, narrow the use case rather than importing everything for possible future use.

Inspect how the firm actually uses each field. A required field can still contain placeholders; an optional field can be highly reliable within one practice. Custom fields may vary by matter type, and exports may omit context that is present in the application. Document the definition, owner, allowed values, population pattern, update process and known exceptions. Test a representative matter sample with the people who create and use the data. Look for duplicate identifiers, stale statuses, inconsistent dates, unmatched users and ambiguous nulls. Do not turn missing data into a negative conclusion. Mark it as unknown and decide whether the workflow should pause, exclude the matter or ask for correction. Vendor documentation helps explain product capabilities, but only the firm's configuration and records show whether a field is suitable for the proposed decision.

Every surfaced fact should carry enough context to be checked: source system, record or field, observation time and any transformation. Derived indicators should state the rule and which inputs were unavailable. Confidence is not a decorative percentage. It can be expressed as a practical status such as complete source set, partial source set, stale observation or mapping unresolved. The reviewer then knows whether to investigate, refresh or disregard the item. If AI extracts a value from a document, preserve the document link and make the extraction distinguishable from an authored source field. Do not silently write the extracted value back into the PMS. A verified correction belongs in the canonical system through its governed process. On the next refresh, the intelligence view should reflect that source change and retain only the operating history needed to explain the review.

Data problems often reveal workflow problems: unclear matter-opening ownership, unused fields, inconsistent practice definitions or access that was never removed. Treat these findings as Pilot outputs. Assign each recurring issue to the source-system or process owner, decide whether the field should be corrected, redefined, retired or excluded and measure whether the change persists. Avoid mass rewriting historical data unless the purpose, authority and impact are understood. A Pilot can succeed even when it concludes that a source is not ready for the proposed use. The firm may simplify the question, add a controlled field at matter opening or use a manual review until the process matures. Readiness is not a one-time certification. It is an evidence-based statement about a particular dataset, decision, cohort and point in time.