Begin with a decision and a data contract
Write the question, audience, unit of analysis and time period before opening an assistant. Define each important field, allowed values, missing-data rule and source owner. This prevents a fluent analysis from quietly answering a different question.
Save an immutable source snapshot or query identifier. Record extraction time, filters and row counts. If the source changes during analysis, create a new version instead of replacing the evidence behind an existing conclusion.
Profile before interpreting
Check schema, types, duplicates, missingness, date ranges, extreme values and category frequencies. Ask AI to propose checks, but run them as formulas, queries or code that another analyst can inspect and repeat.
Investigate unexpected results before cleaning them away. A spike may be a data error, a changed definition or the most important event in the period. Preserve both the raw value and the documented transformation.
Make calculations executable
For every KPI, store the formula, numerator, denominator and inclusion rules. Prefer visible worksheet formulas or version-controlled queries to a number produced only inside a chat response.
Reconcile totals against the source system and calculate a few representative records independently. When AI writes code, review joins, grouping keys, null handling and date boundaries—the places where plausible analyses most often drift.
Separate exploration from evidence
Use an assistant to suggest segments, alternative explanations and visual forms. Treat those suggestions as hypotheses until the underlying calculation is validated. Do not publish a correlation as a causal result without an appropriate design.
Create a claim ledger linking each sentence in the final summary to a table, chart or cited source. If a statement cannot be traced, rewrite it as a question, qualify it or remove it.
Review the narrative last
Generate prose only after the validated analysis table is stable. Ask a reviewer to compare every number, direction and comparison with the evidence, including units and time periods. Keep uncertainty and limitations near the conclusion they qualify.
Archive the source version, transformation steps, outputs, model and prompt context needed to reproduce the work. AI can compress the time between question and draft; the audit trail is what makes the result trustworthy enough to use.
Sources & further reading
Social-media activity is treated as a signal of attention, not proof. Product claims are attributed to the linked publisher or announcement.