THE SHORT ANSWER
AI can help explore data, flag unusual patterns, generate hypotheses and communicate findings. Start with the decision the analysis must support, validate the data and calculations, and distinguish observed relationships from evidence of cause. Human context is essential to choosing the next action.
Replace ‘analyse this’ with a decision question
‘What happened last month?’ invites a narrative of whatever changed most visibly. ‘Should we expand this campaign given lead quality and sales capacity?’ specifies a decision, a unit of value and a constraint. It also reveals which information the dashboard may not contain.
In a hypothetical business-to-business (B2B) campaign, a lower cost per lead could reflect better targeting, a less demanding form or an influx of irrelevant submissions. Ask AI to propose competing explanations and the evidence that would distinguish them. Do not ask it to select the most persuasive story from incomplete data.
Check the dataset before explaining its patterns
- Define each metric, its denominator, date range and reporting time zone.
- Inspect missing records, duplicates, refunds and delayed outcomes.
- Check whether tracking, consent, attribution or campaign settings changed.
- Reproduce key totals with a deterministic calculation outside the generated narrative.
- Compare meaningful cohorts and retain the limits of small or incomplete samples.
A generated query can be syntactically valid while joining tables in a way that counts a sale twice. Inspect the logic and reconcile totals before treating a fluent explanation as analysis. Do not send customer-level data to an unapproved tool just because it accepts a spreadsheet.
Evidence & context: NIST
Use anomaly detection to begin an investigation
An unusual drop can indicate a market change or a broken tag. Compare the timing with operational events, campaign changes and independent records. Ask what should also have changed if the proposed explanation were true.
Predictive features are similarly conditional. GA4, for example, documents data and model-quality prerequisites for its predictive metrics. An eligible prediction estimates an outcome; it does not explain why that outcome will occur or prove that contacting a predicted buyer will change their behaviour.
| Question | Evidence that helps | Common overreach |
|---|---|---|
| What happened? | Reconciled descriptive records | Treating an incomplete total as the whole business |
| What might happen? | A model evaluated on appropriate unseen data | Treating a prediction as certainty |
| What caused the change? | A credible experiment or causal design | Treating correlation or attributed credit as cause |
| What should we do? | Evidence plus costs, constraints and judgment | Letting a recommendation hide its assumptions |
Evidence & context: Google Analytics Help
Finish with a decision and its uncertainty
Randomised advertising experiments are one way to estimate causal effects, as illustrated in Google's brand-lift research. They still need suitable design and interpretation. When experimentation is not feasible, explain the assumptions and uncertainty behind an observational conclusion instead of relabelling it proof.
For the B2B example, a useful recommendation might be to hold budget steady while checking qualified outcomes and a recent form change. That is more actionable than announcing that AI has ‘discovered’ an efficiency gain.
Evidence & context: Google Research
Sources & further reading
- GA4 predictive metrics
Google Analytics Help. Documents data and model-quality prerequisites. Checked 11 September 2026; prediction is not evidence of causal impact.
- Methods for Measuring Brand Lift of Online Ads
Google Research. Original research using randomised experiments to estimate advertising effects; no universal lift or ROI benchmark is inferred.
- Generative Artificial Intelligence Profile (NIST AI 600-1)
NIST. Risk-management guidance, including confabulation. It does not establish a universal error rate.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗