Why your AI strategy is really a data strategy

Data issues can undermine a useful model: inconsistent formats, missing records, duplicate identifiers or information that cannot be accessed appropriately. Audit those constraints alongside the model and the business process.

The uncomfortable audit

Before any model is trained, ask three questions. Where does the data live, and who can access it tomorrow morning? Is it complete enough to represent the decisions you want to improve? And is it being captured consistently going forward — because changing data and business conditions can affect performance.

• Consolidate: one governed store for the data that matters, fed automatically
• Clean: deduplication, validation rules and standard formats at the point of entry
• Capture: fix the forms and processes that create tomorrow's data today

Small data is still data

There is no universal minimum dataset size. Requirements depend on the task, signal, model complexity and acceptable error. Evaluate on representative held-out data, avoid leakage, and use time-aware splits for forecasting. Learning curves can help assess whether more data is likely to improve performance.

Data volume alone does not establish fitness for use. Coverage, provenance, permissions and representativeness matter alongside consistent capture and clear ownership.

Start with data you understand, can govern, and can evaluate against the task.

A practical sequence

Start with a decision and an inventory of the data it needs. Fix material quality and access issues, then establish a reporting baseline. Choose pipelines or storage changes only where needed; validate a model against that baseline before scaling.