Analysing a data table with AI, without being an analyst

A 4,000-row table, a question from the boss like “what caused sales to drop in June?”, and no data analyst on hand. This is exactly the use case where AI changes things: it reads a CSV or an Excel and answers in natural language, with no formula to write.

What AI can really do with a table

Concretely, it turns a plain-language question into a calculation run on your real data, not an approximate estimate. You upload the file, you ask “what is the average by region” or “show me the monthly trend as a chart,” and the tool writes code behind the scenes (Python or JavaScript depending on the platform) to produce a verifiable answer, not one invented from the model’s memory.

This is an important difference from a simple prompt with no attached file: here, the model calculates on your data, it does not guess it.

Which tool for which need

Your situation	Recommended tool	Why
A single CSV or Excel file, need for quick charts	ChatGPT (Plus or above)	Built-in data analysis, generates tables and charts directly in the conversation
Large table, cross-referenced with other documents (reports, contracts)	Claude	Large context window, comfortable cross-referencing several files and producing structured analysis
You already work in Google Sheets	Gemini	Acts directly in the sheet (summaries, trends, formulas) without export, via the AI features of Sheets

For a one-off use, any of the three is largely enough. The difference mainly comes down to your work habit: staying in Sheets or switching to a conversation.

The 4-step method

1. Clean the bare minimum before uploading. One clear header row per column, a single data type per column, no merged cells or multiple tables in the same tab. AI copes with an imperfect file, but it makes fewer mistakes on a clean one.

2. Ask a precise question, not a vague one. “Analyse this file” produces a generic answer. “Calculate revenue by month for 2026 and identify the month with the sharpest drop” produces a usable answer. The same principle as for writing a good prompt applies to data: the more framed the request, the better the answer.

3. Ask to see the reasoning, not just the result. A “show me how you calculated this figure” or “which rows did you exclude” lets you spot a badly-set filter before presenting the result in a meeting.

4. Check against a sample you control. Take three rows or a subtotal you already know and compare with what the AI calculated. If it matches, you can extend the trust to the rest of the table — a reflex close to the checklist to verify an AI answer before relying on it.

The pitfalls to avoid

• Confusing calculation and interpretation. AI calculates correctly on what it understands of your columns, but it can misinterpret an ambiguous heading — for example “Q1” can mean the first quarter or the first quartile (the lowest 25% of values in a series). State the meaning right in the question.
• Uploading sensitive data without thinking. Client names, salaries, health data: on a consumer account, this information may be kept depending on the tool’s privacy settings. Anonymise or use a professional plan with contractual guarantees.
• Trusting a chart without checking the axes. An automatically generated chart can truncate an axis or choose a misleading scale. A glance at the displayed values is enough to spot the problem.
• Forgetting the context window on a very large file. The context window is the amount of text an AI can process at once: beyond a certain volume, it may process only an excerpt of the table without clearly saying so. On a file of several tens of thousands of rows, explicitly ask how many rows were taken into account — see our guide on tokens and the context window.
• Ignoring missing or outlier values. An empty cell or an obviously out-of-range figure can skew an average without AI flagging it on its own. Systematically ask “did you exclude empty or outlier rows, and which ones” before validating a numeric result.

A concrete example, end to end

Let’s take the sales table mentioned in the introduction. You upload the file and ask: “Calculate monthly revenue for 2026 and identify the month with the sharpest drop, excluding rows with no amount.” The AI answers with a summary table and points to June. You follow up: “Show me how you calculated the June total, and how many rows you excluded.” It details its calculation and flags, say, twelve rows with no amount removed. You check three of those rows in the source file: they indeed correspond to cancelled orders, so legitimately excluded. The figure is reliable, you can present it — all without opening a spreadsheet or writing a single formula.

Key takeaway

You do not need to know how to code to draw reliable conclusions from a table: upload, ask a precise question, request the reasoning, check against a known sample. It is the same logic of controlled delegation as for any other task given to AI at the office — you save time on production, you keep control of the checking.