Analyze Business Data with ChatGPT: A Verification-First Workflow

ChatGPT can inspect supported spreadsheets and documents, write and run analysis code, create tables or charts, and explain results in plain language. A reliable workflow starts by checking the file and defining the decision, then validates calculations against the source before anyone acts on the output.

File types, upload limits, tool names, and plan availability change. Check the current product interface and official documentation instead of relying on a static size table.

Before uploading a business file

• Confirm that the organization allows the selected workspace and use case.
• Classify the data and remove names, account numbers, credentials, payment data, health data, and other fields unnecessary for the analysis.
• Keep the original file unchanged and record its source, export time, timezone, filters, and reporting period.
• Decide who will review the calculations and who owns the resulting decision.
• Do not upload confidential, regulated, privileged, or contract-restricted material without the required approval and controls.

Step 1: Profile the dataset

Inspect the uploaded file without changing it.

Return:
- File and sheet names
- Row and column counts
- Column names and inferred data types
- Date range and timezone if detectable
- Missing, duplicate, and invalid values
- Units and currencies
- Active filters or excluded rows if detectable
- Any mismatch between the file and this description: [description]

Do not begin business analysis yet.
List assumptions and ask for clarification when a field is ambiguous.

Compare this inventory with the source system. A wrong delimiter, truncated export, hidden filter, mixed currency, or date-parsing error can invalidate every later result.

Step 2: Turn the business need into a testable question

“Analyze this data” invites arbitrary observations. Tie the request to a decision:

• Which product categories contributed most to the quarter-over-quarter revenue change, using constant currency?
• Which support categories increased each month, and are the changes large enough to affect staffing?
• Where does the sales pipeline lose the highest proportion of qualified opportunities?
• Which expense categories exceeded the approved budget, and were one-time items responsible?

Define the metric, population, period, comparison, segment, and decision threshold. Separate descriptive analysis from claims about cause.

Step 3: Approve an analysis plan

Propose an analysis plan for this business question:
[question]

For each analysis, state:
- Metric and exact formula
- Required columns
- Filters and exclusions
- Grouping and time period
- Missing-value treatment
- Expected table or chart
- Decision the result informs
- Validation check

Do not calculate yet.
Remove analyses that do not contribute to the stated decision.

Step 4: Run and verify calculations

Ask for code or formulas when they help review reproducibility. Then check:

• Row counts before and after every filter
• Totals against a known figure from the source system
• Currency, units, tax treatment, and sign conventions
• Percentage denominators and divide-by-zero handling
• Duplicate keys and one-to-many joins
• Missing-value assumptions
• Whether outliers are data errors or genuine events

Spot-check several rows manually and recalculate at least one headline figure independently. Code execution reduces arithmetic mistakes but does not fix incorrect definitions or dirty data.

Step 5: Create charts that answer the question

• Use a sorted bar chart for ranked categories.
• Use a line chart for a time trend with consistent intervals.
• Use a stacked chart only when both total and composition matter.
• Use a heat map for a meaningful two-dimensional pattern.
• Use a funnel only when stages are ordered and populations are comparable.

Require a descriptive title, labeled axes, units, source period, readable legend, and accessible color choices. Start a quantitative axis at zero when truncation would exaggerate differences; if a nonzero baseline is necessary, make it obvious.

Finance example: budget variance

Using the approved budget and actual columns:
1. Reconcile total budget and actual spend to the supplied control totals.
2. For each department, calculate actual minus budget and variance percentage.
3. Preserve the sign convention: positive means over budget.
4. Flag rows with missing department, duplicate transaction ID, or unknown currency.
5. Return a table sorted by absolute variance.
6. Draft a five-sentence summary that separates verified results from possible explanations.

Do not infer a cause from the amounts alone.

Marketing example: campaign performance

Compare the five campaigns using delivered messages, opens, clicks, conversions, attributable revenue, and approved campaign cost.

Define each rate and denominator before calculating.
Identify missing attribution and tracking changes.
Return the metrics with confidence limitations.
Do not select a "winner" unless the campaigns target comparable audiences and periods.

Sales example: pipeline analysis

For the selected reporting date:
- Deduplicate opportunities by ID.
- Calculate count and value by stage.
- Calculate conversion only between defined consecutive stages.
- Report median and distribution of days in stage, not only the mean.
- Identify records with impossible dates or missing owners.
- Produce a stage table and a clearly labeled chart.

Do not treat current pipeline value as forecast revenue.

HR example: workforce data

Return headcount by department and tenure band using the approved snapshot date.
Suppress or aggregate small groups according to our privacy policy.
Do not rank individuals or recommend employment actions.
Flag missing dates, duplicate employee IDs, and changes in department definitions.
Require HR and legal review before interpreting protected-group differences.

Working with PDFs

PDF extraction can fail on scanned pages, complex tables, footnotes, or multi-column layouts. Ask for page references and open every important page. For contracts or vendor proposals, extract exact terms first, then compare the reviewed table; do not ask the model to make a legal or procurement decision.

Decision-ready final output

Create the final report from the validated analysis.

Include:
1. Five-sentence executive summary
2. Three verified findings with metric, segment, period, and table reference
3. Recommended next action tied to the stated decision
4. Two data-quality limitations that could change the recommendation
5. Assumptions
6. Reconciliation and spot-check results
7. Follow-up data needed

Do not add new calculations or causal claims at the formatting stage.

Reusable end-to-end prompt

Guide me through a verification-first analysis.

File type and source system: [ ]
Export date, period, and timezone: [ ]
Business question: [ ]
Decision supported: [ ]
Known control totals: [ ]
Required segments and metrics: [ ]
Data classification and approved workspace: [ ]
Reviewer: [ ]

Run these gates in order:
1. Data-sensitivity and scope check
2. File inventory and quality report
3. Analysis plan for approval
4. Calculation with visible formulas or code
5. Reconciliation and spot checks
6. Charts with accessible labels
7. Decision-ready report with limitations

Stop when the file does not match the description, a required field is missing, totals do not reconcile, or the question cannot be answered from the data.
Never invent rows, totals, percentages, causes, or missing values.

Checklist

• The file is approved and minimized.
• The data inventory matches the source export.
• The business question defines the decision.
• Formulas, filters, joins, units, and assumptions are documented.
• Headline numbers reconcile and have been spot-checked.
• Charts are labeled and not visually misleading.
• Correlation is not presented as causation.
• A qualified person reviews high-impact conclusions.