Data analysis prompts for ChatGPT and Claude: 10 reliable templates for SMEs
A good data analysis prompt first describes the file (columns, units, period), asks a precise business question and requires the tool to show its calculation. ChatGPT and Claude can run code on an uploaded spreadsheet; your job is to frame the question and check the result before you present it to your leadership team.
SME owners, management controllers, sales managers, nonprofit administrators: you have Excel or CSV exports sitting idle (sales, invoices, memberships, survey responses) and no analyst on staff. This article gives you a three-step method: prepare the file, ask the right questions with ten ready-to-use prompts, check the numbers. For the general principles of writing prompts, the prompt engineering guide for SMEs is the starting point of this series.
What ChatGPT and Claude actually do with your files
When you upload a spreadsheet, the tool does not "read" the numbers like text: it writes and runs code (usually Python) to calculate, filter and draw charts. That is what makes the analysis usable. An LLM adding up 3,000 rows "in its head" would get it wrong; a script that sums a column does not, provided it sums the right column.
The four assistants most common in SMEs all offer this, with some differences:
| Tool | How it analyzes | Strength | Watch out for |
|---|---|---|---|
| ChatGPT | Runs Python code on uploaded files | Charts and cleaned file exports | Check which column was used in the calculation |
| Claude | Runs code on uploaded files, detailed explanations | Commented reasoning, long documents | State the file's separator and date format |
| Gemini in Google Sheets | Works directly in the sheet | No export needed if your data already lives in Sheets | Answers depend on how clean the sheet is |
| Copilot in Excel | Works in the open workbook | Stays inside the team's Microsoft environment | Works best on a structured Excel table |
Features, size limits and eligible plans change several times a year: check the terms of your own subscription before uploading a large file. To choose the tool itself, see our ChatGPT, Claude and Gemini comparison; for Excel, our Copilot 365 beginner's guide.
What does not change from one tool to another: the model knows nothing about your business. It does not know that the "Revenue" column excludes credit notes, that August is always slow, or that "DUPONT SAS" and "Dupont" are the same customer. If you do not say so, it calculates correctly on a wrong assumption.
Step 1: prepare the file before the first prompt
Ten minutes of preparation prevent most misinterpretations: a single header row, one piece of data per column, and a short column dictionary pasted into the prompt.
Files exported from accounting software or a CRM almost always have the same problems: merged cells, subtotals between rows, title lines above the headers, dates stored as text, amounts with the currency symbol inside the cell. The tool can often work around them, but it does so silently, and not always the way you would have.
Minimum preparation:
- remove title, subtotal and total rows;
- keep one header row, with explicit column names;
- remove or replace columns that identify people (see below);
- note the period covered and where the export came from.
Then write the column dictionary in a few lines:
File: 2025 invoice export, invoicing software, 2,840 rows. Columns: invoice_date (DD/MM/YYYY), customer_id (internal code), segment (Micro, SME, Nonprofit), amount_excl_vat (euros, credit notes as negative values), status (paid, overdue, cancelled). Specifics: cancelled invoices must be excluded from any revenue calculation.
This dictionary plays the same role as the context sheet of a writing prompt. You can reuse it from one analysis to the next and store it in a project's instructions, as explained in our article on system prompts and user prompts.
Step 2: ten prompts to analyze your data
Paste each template after the column dictionary and fill in the brackets. They follow the order of a real analysis: understand the file, clean it, explore it, then answer a management question.
1. Have the file described before any analysis
Before any analysis, describe this file: number of rows, columns detected and their type, period covered, missing values per column, values that look abnormal to you. Do not calculate any indicator yet. Tell me if your reading differs from my column dictionary.
This first exchange is the most useful of all. If the tool reports 2,790 rows instead of 2,840, or reads the dates in US format, you see it before a wrong chart is in front of you.
2. Clean without losing track
Propose the cleaning steps needed (duplicates, date formats, labels to standardize) as a numbered list, with the number of rows affected by each. Apply nothing until I agree. Then produce a cleaned file and a change log.
3. Get the core indicators
For [period], by [segment / month / product], calculate: revenue excluding VAT, number of invoices, average order value, share of each segment in the total. Present the result as a table, then give the grand total so I can reconcile it with my accounts: [known total, if you have it].
The known total is your safety net: if it does not match, a rule (credit notes, cancellations, VAT) was applied incorrectly.
4. Compare two periods
Compare [period A] and [period B] on [indicator]. Give the difference in value and in percentage, then break it down: how much comes from the number of customers, how much from the average amount per customer? Name the customers who weigh most in the change.
5. Spot outliers
In the [column] column, find values that clearly differ from what is usual for their [segment / month]. Explain the method in one sentence, list the 10 most striking cases with their original row, and do not conclude they are errors: I will check.
6. Segment a customer base
Using the purchase history, sort customers into 3 or 4 groups based on frequency, time since last purchase and total spend. For each group, give its exact criteria, its size and its share of revenue. Suggest one sales action per group, in one sentence.
7. Read open-ended survey answers
The [comment] column contains [number] open answers to the question "[question]". Group them into 5 to 8 themes, with the number of answers for each and two exact quotes. Flag the answers you could not classify. Do not make up any quote.
For a nonprofit or an HR team, this prompt often pays off the most: hundreds of comments become a readable summary in a few minutes. Giving two or three classification examples makes the output more consistent, which is the principle behind few-shot prompting.
8. Match two files
Here are two files: [file 1, join key] and [file 2, join key]. Match them on [key]. Tell me how many rows found a match, how many did not, and give 5 examples of unmatched rows so I can understand why.
The unmatched rows often tell you more than the match itself: customer codes entered differently, contacts never invoiced, invoices with no contact in the CRM.
9. Build the right chart
I need to present [message in one sentence] to [audience: leadership, board, sales team]. Suggest the most readable chart type, justify the choice in one sentence, then draw it with a title that states the conclusion. No 3D charts, no pie chart with more than 5 slices.
10. Write the summary note
Based on the previous analyses, write a one-page note for [recipient]: 3 findings with figures, 2 possible explanations clearly presented as hypotheses, 2 possible decisions. Separate what is measured from what is interpreted.
The line between finding and hypothesis is the weak spot of AI-generated summaries: a drop in sales quickly becomes "due to seasonality" with no data to back it up.
Step 3: check the numbers before you present them
A table generated by AI looks right because it is well formatted; that does not mean it is right. Errors rarely come from the arithmetic itself: they come from the wrong column, a row excluded without saying so, a business rule ignored.
Four checks are enough in most cases:
- Ask for the method. "Show the code you used and list the filters applied, with the number of rows removed at each step." You do not need to read Python fluently to spot an unexpected filter.
- Reconcile a known total. Annual revenue, number of members, headcount: one reliable comparison point is enough to catch most discrepancies.
- Recalculate a sample. Take one segment or one month and redo the calculation in your spreadsheet. Five minutes.
- Have the result critiqued. "Review your analysis and list the assumptions you made that I did not give you." Getting the model to reason step by step, as in chain of thought prompting, brings those hidden assumptions to the surface.
Fabrication is also a risk in analysis: with no file attached, or when the question goes beyond the data, a model can produce an industry average or a market trend with no source. Our article on AI hallucinations and the hallucination glossary entry explain the mechanism. Simple rule: any figure that does not come from your file must be sourced or removed.
Which data can you upload?
A file uploaded to an AI assistant leaves your information system: before sending it, remove what identifies people, unless your tool and your internal policy explicitly allow it.
The GDPR (EU Regulation 2016/679) applies as soon as a file contains names, emails, phone numbers or identifiers that can lead back to a person. The French data protection authority (CNIL) draws a useful distinction: pseudonymisation (replacing a name with a code) is reversible, and pseudonymised data remain personal data; only anonymisation that makes re-identification impossible in practice takes data outside the scope of the GDPR.
In practice, for an SME:
- No issue: aggregated data, business figures with no personal names, files that are already public.
- With a properly configured business account: customer files where names have been replaced by internal codes, keeping the lookup table on your side.
- Avoid: employee data (salaries, absences, appraisals), health data, information covered by professional secrecy or a confidentiality agreement.
A nonprofit analyzing its memberships does not need names to measure a renewal rate by age group or town. Removing the column before upload settles the question.
Also make sure that conversation reuse for training is turned off, and write these rules into your AI policy. Our article on AI and data confidentiality details the settings for each tool.
When a prompt is no longer enough
Prompts work well for one-off analyses; as soon as the same table has to be produced every month, a dedicated tool is a better fit. Redoing the same analysis by prompt each month means accepting small changes in method from one month to the next, and indicators that are no longer comparable.
A few signals:
- you ask the same question every month on an updated export;
- several people need to see the same figures;
- the data come from several software tools that have to be matched each time.
In that case, use AI to design the indicator and write the first version of the calculation, then set it up in an automatically fed dashboard. Choosing the indicators themselves is covered in our article on AI KPIs in business. If your question is mostly about documents (reports, minutes, contracts) rather than tables, the document-based approach described in RAG vs prompting is a better match.
For more on each tool, our pages on ChatGPT use cases for SMEs, Claude for SMEs and Gemini for SMEs give further examples.
Frequently asked questions
What is the best prompt for analyzing data with ChatGPT?
There is no single one. The most reliable combines three things: a description of the columns and business rules, a precise question tied to a decision, and an explicit request to show the method and the filters applied. Always start by having the file described before asking for a calculation.
ChatGPT or Claude: which is better for data analysis?
Both run code on your files and give comparable results on typical SME analyses. ChatGPT is often preferred for its charts and exports, Claude for its detailed explanations. If your data live in Google Sheets or Excel, Gemini or Copilot save you an export. The choice matters less than the quality of the prompt and the checks.
Can you trust AI calculations?
A calculation run by code is reliable; the interpretation of your file is not always. Ask for the method, reconcile at least one known total and recalculate a sample before sharing a figure.
Can you upload a customer file to ChatGPT or Claude?
With care. Personal data remain subject to the GDPR, even when replaced by codes. Use a properly configured business account, remove columns that identify people when the analysis does not need them, and follow your organization's AI policy.
Do you need to code to analyze data with AI?
No. Being able to read a spreadsheet and ask a precise question is enough for most analyses. Being able to spot a filter or a column name in the generated code helps you check the result; that takes a few hours of guided practice to learn.
What file size do ChatGPT and Claude accept?
Limits depend on the tool and the plan, and they change regularly. Check the documentation for your subscription. For a large file, filtering the period and the useful columns beforehand speeds up the analysis and reduces errors.
How GrowthPerf can help
Analyzing data with AI depends less on the tool than on the method: describe the file, ask a management question, check before sharing. GrowthPerf, a Qualiopi-certified training provider, trains SME, startup and nonprofit leaders and teams in this method, using their own exports.
The Operational AI course includes analysis workshops on your anonymized files: column dictionary, a library of analysis prompts, a checking routine. For a program focused on prompting techniques, see our page on prompt engineering training. In France, these courses can be funded by your OPCO depending on your entitlements.
To go over the whole method again, return to the complete prompt engineering guide for SMEs.
Have exports nobody uses? Book a free 30-minute audit: we look at one of your files with you and identify the three analyses that would best support your decisions.