AI Agent for Spreadsheets: Analyze Data by Asking

An AI agent analyzes a spreadsheet by writing and running code against it, not by reading cells into a chat window. You ask "which region lost the most revenue last quarter," the agent writes a query, executes it, and returns the answer with the rows it used. The reason that distinction matters: an agent that computes can be checked, and one that reads numbers off a grid and does mental arithmetic cannot.

That difference is the whole guide. Everything below is about getting the first thing and never accepting the second.

Why this is one of the two biggest agent use cases

The data backs this up rather than it being a hunch. LangChain's State of Agent Engineering, a survey of over 1,300 professionals, found research and data analysis is the second most common primary agent use case at 24.4%, just behind customer service at 26.5%. Together those two account for more than half of all primary agent deployments.

The same survey found 57% of respondents have agents in production, quality is the top barrier to shipping (32%), and cost has dropped as a concern — falling model prices moved it down the list. So the problem isn't that agents can't analyze data. It's that people can't tell when the analysis is wrong.

What an agent does that a formula doesn't

If you already have the spreadsheet, why involve an agent at all? Three jobs where formulas lose:

The first row is where most people actually live. You don't want a dashboard; you want one answer, once, and then you want to move on. Building a pivot table to answer a question you'll never ask again is the spreadsheet equivalent of writing a function to add two numbers.

A worked example

Let's take a genuinely messy file — the kind that comes out of a Shopify export, a bank, or a booking system. Suppose it has these columns:

order_date, region, amount, currency, status
"2026-08-01", "north ", 1250.00, "SAR", "completed"
"01/08/2026", "North", 890.5, "SAR", "Completed"
"2026-08-02", "south", 145.75, "USD", "refunded"

Three rows, three problems, and each one breaks a formula:

• The same region appears as north , North, and south — trailing whitespace and inconsistent case.
• Dates are in two formats, one of which is ambiguous depending on locale.
• There's a currency column, so summing amount adds riyals to dollars.
• status has both completed and Completed, and a refund is sitting in the same column as revenue.

A human skims past these. A formula doesn't — it silently returns a number that's wrong in a way nobody notices until a decision depends on it.

Step by step

1. Keep a pristine copy. Give the agent a working copy and never let it touch the original. This is the single most important habit in the whole workflow, and it's the same principle behind why agents should ask before they act.
2. Describe your columns, not just the file. Tell it that amount is gross and status includes refunds, and that currency must be respected. Column meaning is the context that prevents a plausible wrong answer.
3. Ask for the cleaning first, as its own step. "List every distinct value in region and status, and tell me what you're normalizing." Review that before any number is calculated.
4. Then ask your real question. One question per pass, so a wrong answer is traceable to one cause.
5. Demand the rows. "Show the rows that make up that total." An answer you can't trace is a rumor.
6. Spot-check by hand. Pick three rows from the evidence and verify them yourself. Thirty seconds, and it catches whole classes of failure.

The prompt pattern that does the most work here is boring: ask for the method alongside the answer. "Show me the rows, the filter you applied, and how you handled currency." An agent that has to explain its query usually can't quietly guess.

What to check before you trust a number

• Did it exclude refunds? A revenue figure that includes refunded orders is wrong, and it looks completely normal.
• Did it convert currency at some rate it invented? If the agent reports a single total across currencies, ask what rate it used. If it can't name one, the total is fiction.
• Does the row count match? "Total across 412 orders" is checkable. If your file has 1,200 rows, that gap is the story.
• Are the dates parsed the way you think? Ambiguous formats silently move transactions between months, which is exactly how a quarterly comparison flips sign.
• Is the answer suspiciously round or suspiciously clean? Small datasets are where models hallucinate summary statistics — one guide to a popular analysis tool warns directly that too few rows or many missing values produces invented numbers. Ask for the underlying rows when the data is thin.

When to keep using formulas

Agents are not always the right tool, and pretending otherwise costs you accuracy:

• The answer is recurring. If you'll ask the same question every month, build the formula once. An agent that re-derives it monthly will re-derive it slightly differently.
• The data is small and clean. Ten tidy rows don't need a code-executing loop.
• It's a calculation of record. Anything feeding a tax filing, a payroll or an audit belongs in a deterministic formula you can point at. Keep the agent for analysis, not bookkeeping.

For the recurring-and-personal cases, the right split is an agent that finds the insight and formulas that hold the line — the same division of labour that shows up in using an agent for expense reports.

The takeaway

An AI agent analyzes a spreadsheet well when it runs code, shows its working, and hands you the rows — and badly when it reads numbers and reports a total you can't trace. Ask for the method with the answer, check refunds and currency every single time, and keep your formulas for the questions you'll ask again.

The messy sample dataset used in this guide — download the CSVmessy-orders-sample.csv · CSV