A responsible analysis workflow
Copy and scope the data
Work from a copy. Specify included files, relevant date ranges, important fields, and excluded sensitive information.
Profile before calculating
Ask the agent to report types, missing values, duplicates, and suspicious records before drawing conclusions.
Define the question
Replace “find insights” with measurable questions such as revenue change, retention by cohort, or anomalies by category.
Preserve the calculation path
Require formulas, grouped tables, or code used for important metrics so a person can reproduce them.
Separate evidence from explanation
Measured patterns are facts about the dataset. Reasons for those patterns are hypotheses unless supported by additional evidence.
Review and export
Verify key rows and totals, then save a report and any transformed data separately from the source.
Example task specification
Analyze the attached order export from January through June.
Questions:
1. How did monthly revenue and order count change?
2. Which product categories contributed most to the change?
3. Which records look anomalous and why?
Requirements:
- do not alter the source file
- show calculations for every headline metric
- label explanations as hypotheses unless supported by data
- return a report plus a CSV of the summarized tables
- ask before contacting an external serviceKeep the data boundary explicit
Choose execution based on the dataset:
- Local AI: appropriate when files should remain on your computer.
- Private Network: useful when an approved machine provides a needed model or more compute.
- Public Network: reserve for general or sanitized tasks that are suitable for public providers.
Remove unnecessary personal fields, credentials, secret keys, and unrelated columns before analysis. Data minimization matters even when execution is local.
What to verify in the result
| Check | Why it matters |
|---|---|
| Row counts before and after cleaning | Reveals accidental exclusions or duplicate handling. |
| Units and date boundaries | Prevents mixing currency, time zones, percentages, or incomplete periods. |
| Headline calculations | Confirms that the narrative matches the underlying numbers. |
| Outlier treatment | Shows whether unusual records were removed, capped, or retained. |
| Facts vs hypotheses | Stops plausible explanations from becoming unsupported conclusions. |
| Saved artifact paths | Makes it clear which files were created or changed. |
Where human expertise remains essential
An agent can accelerate inspection, computation, and reporting. It cannot determine whether the source data represents reality, whether the sampling process is biased, or whether a decision is lawful and appropriate without relevant context. Financial, medical, employment, legal, and safety-critical conclusions require qualified review.
Learn how the agent loop works, or compare local and cloud execution before choosing where a dataset should be processed.
Turn one dataset into an inspectable report
Start with a copy, define the question, and require the agent to show its work.