How can AI be used to analyse data?
AI can be used at almost every step of a simple analysis: clarifying the question, describing the dataset, suggesting cleaning steps, computing summaries, writing formulas or Python, building charts and drafting a written interpretation. The time saving is mostly in the mechanical steps. The risk is mostly in the last one, where a fluent paragraph can state a conclusion the data does not support.
This guide is about everyday business analysis by non-specialists: a manager looking at sales by month, a support lead looking at ticket volumes, an operations coordinator comparing supplier lead times. The companion guide to AI for Excel covers formulas and spreadsheet clean-up in detail.
What kinds of data analysis can AI help with?
Analysis is often described in four types, and AI helps with each to a different degree. The more the question moves from what happened to why and what to do, the more human judgement and context the answer needs.
| Type of analysis | Question it answers | How much AI can help |
|---|---|---|
| Descriptive | What happened? | A lot: totals, trends, breakdowns and charts, all checkable |
| Diagnostic | Why did it happen? | Some: it can suggest hypotheses and splits to test, but cannot confirm cause |
| Predictive | What is likely next? | Limited for non-specialists: forecasts need method and validation |
| Prescriptive | What should we do? | A starting list of options; the decision and its trade-offs stay with people |
Does ChatGPT do data analysis?
General assistants such as ChatGPT, Claude and Gemini can help with data analysis in two ways: by reasoning about data you paste or describe, and, on some plans, by running code on a file you upload. When a tool runs code, its arithmetic is usually more reliable than when it estimates figures in prose, but the code can still answer a slightly different question from the one you meant. Check which capabilities your approved plan includes in the vendor's current documentation.
Spreadsheet assistants are another route. Microsoft describes a chat only mode in Copilot in Excel that analyses workbook data and gives insights without changing the workbook, and Google says Gemini in Sheets can generate data analysis and insights and build charts. Google's page also warns that Gemini may suggest inaccurate information. Whichever route you use, the checks in this guide apply.
What is a safe method for analysing data with AI?
Write the question down before you open the tool, describe the data accurately, ask for the method before the answer, and verify every number you will repeat. The steps below work for most small business analyses.
State the question
Write one sentence: what decision will this inform, and what would change your mind? Without it, the tool will find patterns, and some will be noise.
Describe the data
Columns, units, time period, how each row was recorded and anything known to be missing. Ask the tool to restate it so you can catch misunderstandings.
Ask for the plan first
Ask how it would answer the question and which calculations it would run. Fix the plan before it produces numbers.
Recompute the key figures
Check the two or three numbers that will go in your summary with a formula, a pivot table or a calculator.
Separate finding from explanation
Write what the data shows, then separately what might explain it and how you would test that.
Record what you did
Keep the question, the data version, the method and the checks, so someone else can follow it.
Want to see how this connects to a role?
Explore the planned team trial ↗What does a misleading AI analysis look like?
Here is a worked example with fictional figures. Lumen Box, a small subscription company, raised the price of its monthly plan on 1 September. The founder uploads a cancellations sheet and asks: Did the price increase cause more cancellations? The data shows 1,950 active subscribers at the start of August with 100 cancellations that month, and 2,000 at the start of September with 140 cancellations.
First AI answer: Cancellations increased by 40 percent in September, which shows the price increase drove customers away. You should consider reversing it. Three problems. The 40 percent is a rise in the count of cancellations, from 100 to 140; the cancellation rate went from about 5.1 percent to 7.0 percent, a rise of roughly two percentage points. Mixing up percent and percentage points is one of the most common errors in AI-written analysis. Second, two months of data cannot show that the price change caused the rise; September might differ for other reasons. Third, the recommendation goes well beyond what was asked or shown.
Better follow-up prompt: Using only the uploaded sheet, give the cancellation rate for August and September with the calculation shown. Then split September cancellations by plan type, since only the monthly plan changed price. Do not state a cause. List what other data would help test whether the price change was responsible.
Reviewed finding: Monthly cancellation rate rose from about 5.1 percent in August to 7.0 percent in September. The rise coincides with the 1 September price change on the monthly plan. To test whether it was the cause, compare cancellations on the monthly plan with the unchanged annual plan, read the cancellation reasons customers gave, and look at the same months last year. The founder checked both rates by hand before sharing.
What data should you not give an AI tool?
Do not give an AI tool personal or confidential data unless your company has approved that tool and plan for it. Customer lists, employee records, health information and unreleased financial results usually need explicit approval. Replacing names with codes helps, but it does not make the data anonymous if someone can link the codes back. The UK ICO says pseudonymised data is personal data in the hands of someone who holds the additional information needed to identify people.
Often you can avoid the question. Aggregate first, so you share monthly totals rather than individual rows, or describe the columns and use made-up sample rows to get the formula or code, then run it on the real data inside your own systems. This is general information, not legal advice; check your policy and, where needed, your data protection lead.
Is Copilot good for data analysis?
Copilot in Excel is a reasonable choice for everyday analysis if your organisation already has it, because it works on the workbook where the data lives and, according to Microsoft, offers a mode that analyses without editing. Whether it is good for a particular analysis depends on the same things as any tool: how clean the data is, how precisely you ask and whether you check the result. It does not replace knowing what a sensible answer should look like.
If your team is choosing between assistants for analysis, test each on one real, non-sensitive question you already know the answer to. That tells you more than a feature comparison.
When should you call an analyst instead?
Call an analyst, or someone with statistical training, when the decision is expensive, the data is large or messy, or the question is about cause, forecasting or whether a difference is real rather than noise. AI makes it easy to produce a confident chart and a tidy paragraph. It does not make the underlying statistics any simpler, and a non-specialist may not know which question to ask the tool.
AI is also not a substitute for missing data. If the information needed to answer the question was never recorded, the right output is a note saying so and a plan to collect it.
Sources and scope
Microsoft Support, Get started with Copilot in Excel, accessed 16 October 2026 ↗
Google Docs Editors Help, Collaborate with Gemini in Google Sheets, accessed 16 October 2026 ↗