Spreadsheets and data tools

The category where verification is easy, and that changes everything

Most AI work is hard to check. Data work is the exception: if the model writes code, you can read the code, run it, and compare the answer with something you already know. That makes this the category where these tools give the most reliable return, and it should shape how you choose among them.

The consequence is a strong preference: prefer a tool that shows you the code over one that shows you only the answer.

Four shapes on offer

AI inside the spreadsheet. Formula suggestions, natural-language column operations, summarising a range. Convenient for small jobs, and it operates on your live sheet, which is worth respecting.

Code interpreter in a chat product. Upload a file, ask a question, get an answer plus a chart plus the code that produced it. Currently the best general-purpose option for anybody who does not write code, precisely because the code is visible.

A notebook with an assistant. Jupyter or a hosted notebook with completion. More control, more setup, the same visibility.

SQL assistants. Turn a question into a query against a database. The query is the thing to read; a wrong join is invisible in the result and obvious in the SQL.

The free path

LibreOffice Calc for the spreadsheet. Python with pandas in a free notebook: Google Colab and Kaggle both offer free hosted notebooks, and a local Jupyter costs nothing. A free coding assistant such as Continue with a local model handles the writing. DuckDB runs SQL over CSV files directly, free, on your own machine, without a database server.

That combination does everything the paid data tools do for anyone working at the scale of a spreadsheet, and it keeps the data local.

The failures specific to data work

The silently dropped row. A filter or a join that removes records nobody counted. Always print the number of rows before and after every operation. This one habit catches more errors than any other.

The wrong column. Two columns with similar names, and the model picks the other one. Reading the code catches it; reading the chart does not.

Numbers that are text. A column of amounts stored as strings, with currency symbols or thousands separators, sorts alphabetically and sums to zero. Very common in exported data.

Dates. Ambiguous formats, mixed conventions in one column, and time zones. 03/04/2026 is two different days depending on where the file came from.

The confident average. A mean computed over a column containing sentinel values like -1 or 9999 for “unknown”. The answer is a number, and it is meaningless.

Silent type coercion. A join on an identifier where one side is numeric and the other is text produces zero matches, and the code runs without complaint.

The habit that makes it safe

Ask three questions of every result, and ask them out loud:

1. How many rows went in, and how many came out?
2. Does the answer match something I already know? One value you can verify by hand.
3. What would this look like if the code were wrong? If you cannot tell, add a check that would distinguish the two.

Then read the code. Even without knowing the language, the column names and the filter conditions are readable, and that is where the errors are.

Where the AI-in-the-spreadsheet products differ

If you do stay inside a spreadsheet, one distinction is worth knowing. Some products generate a formula you can inspect and keep; others compute a value and paste it in. The first leaves an auditable sheet that recalculates when the data changes; the second leaves a number nobody can explain in six months.

Prefer the formula every time, even when the value is what you wanted. A spreadsheet full of pasted constants is the same problem as an answer without code, and it is harder to spot because the sheet looks normal.

Where these tools are excellent

Exploratory work on a file you have never seen. Ask what is in it, what the distributions look like, what is missing, what looks odd. That is an hour of tedious work compressed into five minutes, it is exactly the kind of task where a wrong answer is visible, and it is the best argument for the whole category.