Claude for data analysis: what it does well, and where to check it

A marketing operations manager uploads a 40,000-row export of ad spend and conversions and asks Claude for return on ad spend by channel for the last quarter. Claude doesn't read the rows and estimate. It writes a short Python script, runs it in a sandbox, and returns a table, a chart and, if asked, a spreadsheet to download.

That code step matters. Language models doing arithmetic "in their heads" are unreliable: in one well-known study, GPT-4 got only 59% of three-digit by three-digit multiplications right without a tool (Dziri et al., Faith and Fate, 2023). Both Anthropic and OpenAI route calculations to code for that reason. Anthropic's API docs describe code execution as letting Claude "analyze data, create visualizations, perform complex calculations" in a sandboxed environment (Anthropic, Code execution tool).

In the Claude apps, this is the feature Anthropic calls file creation, or code execution. It replaced the earlier JavaScript "analysis tool" from October 2024, which Anthropic said gave "mathematically precise and reproducible" results (Anthropic, October 2024). Today, Claude can:

• analyse uploaded files (CSV, XLSX, JSON, PDF and others) with Python;
• create PNG charts and Python scripts;
• create and edit .xlsx, .pptx, .docx and PDF files, which you can download or save to Google Drive.

For data that lives in a warehouse rather than a file, see Claude connectors: BigQuery, Snowflake and others have MCP servers Claude can query directly.

1. Check code execution is on. It's on by default for Free, Pro, Max, Team and new Enterprise organisations, under Settings > Capabilities. Owners can switch it off, so if Claude says it can't run code, ask your admin. Excel files need it; without it, Claude can't open .xlsx at all.
2. Upload the file. Anthropic's upload limits allow up to 500 MB per file and 20 files per chat. Its file-creation help gives a lower figure, 30 MB per file, for files inside the code sandbox. If a large file fails, split it or aggregate it first.
3. Ask it to describe the data before analysing it. Columns, types, row count, date range, obvious gaps. Thirty seconds of reading here catches the wrong file, the wrong sheet and the half-loaded export.
4. State the definitions. Don't ask for "ROAS" and hope. Say which value and which cost.
5. Ask for the method alongside the answer. The code, the filters, and any rows excluded.
6. Export. Ask for the result as a spreadsheet or deck, or turn it into an artifact to share.

A prompt that does steps 3 to 5 at once (illustrative):

Attached is our Q3 paid media export. Before any analysis, tell me the columns,
the row count, the date range, and any columns with missing values.

Then calculate ROAS by channel for 1 July to 30 September: total conversion
value divided by total spend, per channel. Don't average daily ROAS values;
compute from totals. Exclude rows where campaign_type is "test".

Show the Python you ran, the number of rows used per channel, and a list of
anything you had to assume.

Note the line about averaging. A mean of daily ROAS values is not the same as total value divided by total spend, and it's an easy thing for anyone, human or model, to get wrong. Our ROAS metric page has the definitions.

Mistakes like that are hard to spot in a finished answer. Count connects to Claude and shows the calculation behind each number, so a mean of ratios is visible before it reaches a report.

Once the code runs, the arithmetic is right. Errors come from what the code was told to compute. Five patterns account for most of what we see:

1. Picking one definition without saying so. "Churn" can mean customers lost or revenue lost. In ChartMogul's worked example, three customers pay $20, $30 and $50; the $50 customer leaves. Customer churn is 33%. Revenue churn is 50%. Same data, both correct, very different board slide. Google defines ROAS as conversion value divided by spend and ROI as profit divided by spend; ask for one and you may get the other.

2. Being thrown by irrelevant detail. Apple researchers found that adding a single irrelevant clause to maths problems cut accuracy by up to 65%; GPT-4o fell from 95.2% to 63.1% (Mirzadeh et al., GSM-Symbolic, 2024). Real exports are full of irrelevant columns and notes.

3. Reading the wrong number from a document. When the data arrives as a PDF report rather than a table, extraction is a weak point. Even given the right page of a filing, GPT-4-Turbo got 15% of FinanceBench questions wrong (Islam et al., 2023). Older model, same lesson.

4. Writing plausible SQL against a real schema. In the original Spider 2.0 paper, the best agent solved 21.3% of enterprise text-to-SQL tasks against 91.2% on the academic Spider 1.0 (Lei et al., arXiv v1, 2024). Agents have improved a lot since, but the wrong join or date column still gives a confident, wrong number.

5. Quiet filtering. Dropping rows with missing values, deduplicating, excluding test accounts. Each is often right. Each changes the answer, and none appears in the summary unless you ask.

Saturated benchmarks make AI look near-perfect. Benchmarks built from real analyst work tell a different story:

These aren't measures of Claude in the chat app, and the models tested aren't all current. They're the best evidence available that multi-step analysis on real data is still hard, even for the strongest systems: roughly one in three analyst tasks missed on the finance benchmark, and about two in three spreadsheet workflows on SpreadsheetBench V2.

Our position: Claude is very good at the mechanics of analysis (loading, cleaning, reshaping, charting) and at explaining what it did. It saves real hours. It doesn't know your business's definitions unless you tell it, and it won't flag the decision it made silently unless you ask. For a direct comparison with OpenAI's equivalent, see Claude vs ChatGPT for data analysis.

Anthropic is plain about the risk: the feature "gives Claude internet access to create and analyze files, which may put your data at risk". A prompt injection hidden in a file or web page could make Claude run untrusted code or send conversation data elsewhere, so Anthropic advises watching what Claude does and stopping it if anything looks unexpected. It mitigates with an injection classifier, per-user isolated sandboxes and resource limits. Public sharing of chats with file outputs is disabled on Free, Pro and Max.

For company data, the practical rule is to upload the extract the question needs, not the whole table, and to use Enterprise network controls if you handle anything regulated.

Anthropic's own help page advises that "you may need to review and refine Claude's outputs" and suggests starting with simple tasks. Fair. What follows is the review we'd do on any number that's going into a report.

1. Rows in, rows out. Compare the row count Claude says it used with the row count of your file. A gap means a filter, and you should know what it was.
2. Read the filters and joins in the code. Not the whole script: search it for the lines that drop, filter, merge or group. That's where the decisions live.
3. Recompute one figure yourself. One cell, in a spreadsheet, from the raw data. If it matches, the rest is probably mechanical. If it doesn't, stop.
4. Ask about definitions: "Which definition of [metric] did you use, and what other definitions could someone reasonably mean?"
5. Reconcile one total against the source system: total spend, total customers, total revenue. If the totals don't match, nothing built on them will.

The deeper problem is that the code and the claims drift apart once Claude writes its summary. Count connects to Claude and splits the answer into claims, each with a reasoning chain showing the query or calculation behind it, and labels the ones that rest on assumptions. Its canvas also runs SQL and Python against your warehouse, so you can open the step that looks wrong and re-run it.

Can Claude analyse data?

Can Claude read Excel files?

How large a file can Claude analyse?

Is Claude better than ChatGPT for data analysis?

Is it safe to upload company data to Claude?