A trustworthy AI analysis Report contains at least a decision question, definitions, principal findings, evidence, limitations, and uncertainty or next steps. A polished summary speeds reading but cannot replace them. Readers must understand what a number means, where it came from, and which conclusions remain unsupported.
A six-part Report template
1. Decision question: what action should this support?
Replace "analyze last month's performance" with "decide whether the new onboarding should expand to all new users, primarily using completion and seven-day activation rates." Point toward a decision without embedding the desired conclusion. If the question changes, update the analytical contract rather than quietly changing the Report.
2. Definition: how were the numbers constructed?
State the population and grain; numerator, denominator, and exclusions; date range, time zone, and cohort method; and important join keys and source version/extraction time.
"Completion was 62%" is incomplete. "Among accounts first entering onboarding in July, 62% completed every required step within seven days, excluding test and deleted accounts" can be reviewed.
3. Principal findings: answer first, then qualify the evidence
Use "conclusion + number + comparison + condition":
New-version completion was 62% versus 55% for the old version in the same period, a seven-percentage-point difference. This observational comparison has not controlled for differences in traffic-source mix.
This separates what happened from whether causation is supported.
4. Evidence: make every number traceable to a query
Evidence does not mean pasting all SQL. Give key numbers a locatable SQL artifact, execution step, result field, and validation checkpoint:
| Finding | Evidence | Validation |
|---|---|---|
| Completion 62% | 02_completion_rate |
Distinct account denominator |
| Mobile has highest drop-off | 03_dropoff_by_device |
NULL rate for device field |
A conclusion that cannot identify its retrieval logic is not yet a formal result.
5. Limitations: what can the data not answer?
Limitations define the decision boundary, not merely legal cover. Common examples include missing channels, mid-month event-definition changes, small samples, missing confounders, or an observational design that cannot support causality. NIST treats AI risk management as lifecycle work spanning design, use, and evaluation; explicitly stating known limitations is actionable risk communication. See the NIST AI RMF.
6. Uncertainty and next steps: what should happen now?
Distinguish low-regret actions available now, hypotheses requiring more data, and tradeoffs for an owner. For example: "fix copy in mobile step two now, while using a randomized experiment to test whether the new version causes higher completion."
Copyable Report skeleton
## Decision question
Decision, population, and time range.
## Definition
Grain, numerator/denominator, exclusions, and data timestamp.
## Principal findings
Conclusion + number + comparison + condition.
## Evidence
SQL/execution step and validation for every key number.
## Limitations
Coverage gaps, assumptions, biases, and unsupported inferences.
## Next steps
Immediate actions, hypotheses to validate, owner, and review time.
How Lantide Data keeps the Report connected to analysis
Lantide Data Project Analysis aligns definitions in a Plan. After user Approve & Execute, it retains formal SQL and execution evidence, then creates a Report with numbers and limitations. Reports support comments and HTML delivery. The roles of Plan and Report are respectively the pre-execution contract and post-execution result, not just chat summaries.
This artifact-first model lets readers trace conclusions to definitions and evidence, while separating presentation changes from calculation changes. HTML wording and layout can be refined; if a denominator or number is wrong, return to a new Plan and execute again rather than editing the displayed number. See the HTML Report guide.
Lantide cannot automatically decide whether a conclusion reflects business reality or turn observational data into causal evidence. Domain judgment, risk ownership, and the final decision remain human responsibilities.
Conclusion
Review with six questions: why was it done, how was it calculated, what was found, where is the evidence, what cannot be claimed, and what happens next? Treat any missing item as unfinished analytical work, not something a prettier summary should conceal.