Data & Analytics Practical insights
Audit marketing data before connecting it to an LLM
Check row meaning, joins, missing values, freshness and access permissions before connecting marketing datasets to an LLM for reporting or decisions.
Connecting an LLM to marketing data does not make the underlying tables easier to interpret correctly. A fluent answer can hide duplicated revenue, missing markets or an inappropriate join. Audit a representative dataset before widening access, focusing on the errors that could change a real business decision.
Identify what one row represents
Write down the grain of every table: one event, order line, customer or daily campaign record. Inspect keys and test their expected uniqueness. Joining campaign totals to several creative rows can multiply spend if the query sums the repeated values. Prepare a small, manually checked example showing the correct aggregation before allowing the assistant to build similar queries.
Examine absence and freshness
Distinguish zero from missing, unavailable and not yet processed. Profile missingness by source, period and market so an apparently complete average does not conceal an excluded group. Check late-arriving records, currency units, timezone handling and changes in field meaning. Keep a clear boundary between an observed value and an estimate.
Limit the accessible surface
Expose approved views containing only data needed for the task. Enforce permissions in the data layer; a prompt that says not to reveal restricted records is insufficient. Clear tool descriptions can help an agent understand the outputs it receives. [1] Return freshness and scope information with the data so the system can identify an incomplete answer.
Reconcile and release gradually
Choose several questions with known answers and compare results with the authoritative source. Include awkward cases: refunded orders, duplicated identifiers, a missing day and a user without permission for another business unit. Record discrepancies and who must fix them. Start with read-only access to a limited dataset, and broaden it only after the observed failures are resolved. The audit’s output should be approved data contracts and reproducible checks, not merely a declaration that the database is ready for AI.
Sources and evidence
Sources checked on 4 October 2026. Proposed workflows and hypothetical examples are editorial analysis.
From insight to practice