A polished analysis can contain an incorrect calculation, an unsupported explanation or a recommendation based on a fact that was never supplied. Verification should happen before the output influences a consequential decision.
Break the analysis into claims you can inspect. Numbers, factual statements, assumptions and recommendations need different checks.
Recalculate the important numbers
Return to the original data. Check units, dates, denominators and whether the compared periods are equivalent. Use a trusted calculation method instead of asking the same model to reassure you that its answer is correct.
For illustration, growth from 80 to 100 is 25% relative to 80. The increase of 20 units is a different quantity. A small denominator mistake can materially change a recommendation.
Trace claims to evidence
Open cited sources and verify that they support the specific claim. Separate supplied facts from assumptions added by the model. A plausible causal story is not evidence that one event caused another.
NIST identifies confabulation as a generative AI risk: systems can confidently present false material. That is why verification needs an independent reference, not just confident wording.
Test the recommendation against constraints
Ask whether the proposed action fits your resources, commitments and decision authority. Identify what evidence would reverse the recommendation and what downside a small test could reveal.
For financial, legal or other high-consequence decisions, involve the appropriate qualified professional. The analysis can organize questions without replacing that person's judgment.
Reference: NIST — Generative AI Profile.