How to verify an AI-generated business analysis
Review an AI-written analysis by separating facts, calculations, interpretations, and recommendations. Use a small worked example to identify unsupported causes, correct a percentage-point error, and produce a checked explanation you can defend.
Before you start
No coding required
A claim-by-claim review and corrected short analysis.
- Ability to compare a response with supplied source notes
- Text-capable AI chat — access: varies
- Calculator — access: varies
A readable answer can combine correct numbers with an unsupported explanation. Review the claims individually before trusting the paragraph as a whole.
You will inspect a fictional onboarding report, check its arithmetic, and rewrite it. Use a calculator and optionally an AI chat tool. No programming or file upload is needed. Hands-on time has not been measured.
1. Keep the evidence visible
These original synthetic records describe checklist completion, not sales or customer satisfaction:
| Record | Invitations | Completed checklists | Period |
|---|---|---|---|
| S1 | 40 | 20 | August 2026 |
| S2 | 50 | 30 | September 2026 |
Definition: completion rate = completed checklists divided by invitations in the same period. There are no supplied interviews, experiment results, or revenue data.
Review this deliberately flawed editorial illustration:
Completion rose from 50% to 60%, an improvement of 10%. Our new reminders caused the increase, proving they will grow sales next month. We should roll them out immediately.
2. Separate the claims
Calculate independently before asking another model to check:
August rate = 20 / 40 = 50%
September rate = 30 / 50 = 60%
Difference = 60% - 50% = 10 percentage points
Relative increase = (60% - 50%) / 50% = 20%
The percentage-point change and relative percentage increase answer different questions. If invitations were zero, the completion rate would be unavailable, not zero.
| Claim | Review |
|---|---|
| Rate changed from 50% to 60% | Supported by S1 and S2 |
| Improvement of 10% | Ambiguous/wrong unit for the stated subtraction; use 10 percentage points |
| Reminders caused the increase | Unsupported; reminders are not documented in the inputs |
| Sales will grow next month | Unsupported forecast; no sales evidence supplied |
| Immediate rollout is justified | Recommendation exceeds the available evidence |
3. Use AI to assist the review
Paste the table and flawed paragraph into this prompt where indicated:
Review this fictional onboarding analysis against the supplied inputs.
[Paste records S1–S2, the rate definition, and the flawed paragraph.]
For each claim, classify it as supplied fact, calculation, interpretation,
or recommendation. Show the supporting record ID or say what evidence is missing.
Recompute rates using each period's invitations. Distinguish percentage points
from relative percentage change. Do not invent a campaign or a causal explanation.
Return a review table, then a corrected paragraph under 80 words.
Check the review itself against S1 and S2. A second AI response can repeat the same error. Arithmetic and source inspection remain your responsibility.
4. Produce a checked version
This is the corrected editorial example:
Checklist completion increased from 20 of 40 invitations in August (50%) to 30 of 50 in September (60%). The rate rose by 10 percentage points, equivalent to a 20% relative increase. The inputs do not establish the cause or any effect on sales. Review the differences between cohorts and any documented process changes before proposing a rollout.
It preserves the observed change while leaving causal and commercial questions open.
Reusable review checklist
Confirm the scope, dates, units, denominators, and source labels. Recalculate consequential figures. Identify each causal explanation and ask what supports it. Separate proposed actions from approved actions. Check that omitted information would not materially change the conclusion.
If a source is missing, request it or remove the claim. If totals disagree, stop drafting and resolve the data. If a recommendation sounds stronger than its evidence, narrow it to an investigation or clearly labelled proposal.
For another exercise, change September completions to 25. The rate becomes 50% and the difference becomes zero percentage points; any claim of improvement should disappear. Next, apply this review to the budget commentary, procedure, or feedback guide.
Source: original synthetic records S1–S2. The review method and examples are educational editorial material; no external model execution is claimed.