Data Analysis with Agents
Let agents clean spreadsheets, run analyses, and build charts — while you stay in control of the conclusions.
You have a spreadsheet and a question. Maybe it's twelve months of sales exports and you need to know which regions are slipping, or a survey with 900 free-text answers and a leadership meeting on Thursday. A year ago you'd either wrestle with pivot tables for an afternoon or wait in line for an analyst. Now an agent can clean the data, run the analysis, and draw the chart in minutes.
The risk is that it will also hand you a confident conclusion that's wrong in a way you can't see: a filter applied silently, duplicates double-counted, a correlation dressed up as a cause. This lesson shows you how to work with agents on data so you get the speed without giving up control: profile first, clean with a paper trail, ask answerable questions, check the charts, and verify before you present.
How agents actually analyze data
When you hand a spreadsheet to a capable agent, it usually doesn't "read" the numbers the way you do. It writes and runs a small program (typically Python with a data library like pandas) that loads the file, computes the answer, and prints the result. Chat apps like Claude, ChatGPT, and Gemini do this in a sandbox when you upload a file; coding agents like Claude Code or Codex do it on your own computer, which is useful for large files or data you'd rather not upload.
This matters for one reason: computed numbers are much more trustworthy than estimated ones. If the agent ran code, the arithmetic is almost certainly right, and any errors are in the logic (which rows, which columns, which definition), which you can inspect. If it just looked at the data and answered, it may be pattern-matching, and the numbers can be plausible but wrong.