When a KPI drops, AI can generate twenty possible explanations in seconds. That is useful. It is not root-cause analysis.
Root-cause work becomes valuable when you move from possible explanations to tested explanations.
1. Quantify the change
State exactly what moved, by how much, when the movement began and which comparison you are using. A 12% drop week-over-week is a different problem from a 12% drop versus the same week last year.
2. Rule out data problems
Before blaming the business, check refresh timing, source changes, missing records, duplicate records, broken joins, filter changes and metric-definition changes.
This matters because a dashboard can report a real numerical change that is not a real business change.
3. Use AI to create competing hypotheses
Ask for operational, customer, product, pricing, marketing and data-quality explanations. Ask for the evidence that would support or weaken each one.
4. Segment the change
Break the metric by the dimensions that could plausibly explain it: geography, product, customer type, channel, device, cohort or operational stage. Look for where the movement is concentrated.
5. Test, don’t narrate
For each hypothesis, define a test. If you think a pricing change caused the decline, compare affected and unaffected groups and inspect timing. If you think a pipeline issue caused it, reconcile source and reporting totals.
6. Use AI to challenge the conclusion
Once you have evidence, ask AI: “What alternative explanation still fits these findings?” That is much more useful than asking, “Why did revenue fall?”
A simple RCA worksheet
| Observation | What changed? |
| Data check | Could the measurement be wrong? |
| Hypothesis | What could explain the change? |
| Test | What evidence would distinguish the explanations? |
| Conclusion | What is supported—and what remains uncertain? |
For the broader metric-investigation process, see the root-cause analysis guide.