Data analysts use AI to speed up the parts of the job around the analysis: profiling and cleaning a file, writing formulas and SQL, drafting charts and writing findings for non-technical readers. They keep accuracy by making the AI check the data before it answers, asking for formulas or code instead of typed-in numbers, and tying every key total back to the source, the routine in Analyze a Spreadsheet.
The accuracy risk is real but specific. AI rarely gets a SUMIFS wrong when it writes the formula. It goes wrong when it quietly drops rows, counts duplicates twice, guesses what a column means, or reports a number it calculated in its head instead of in code.
Where do analysts use AI day to day?
- Scoping. Turning "can you pull some numbers on churn?" into a clear question with a definition, a time frame and a comparison.
- Data profiling. Columns, types, row counts, date ranges, duplicates, blanks and odd values, before any analysis.
- Cleaning. Standardizing dates, fixing category spellings, removing subtotal rows from exports.
- Formulas and queries. Writing and debugging Excel formulas, SQL and pandas from a plain description.
- Exploration. Suggesting cuts of the data and chart types worth trying.
- Communication. Findings for a manager, metric definitions, and documentation for a recurring report.
The prompts for data analysts include ready-made versions of most of these, from writing SQL from a plain description to sanity-checking results.
How do I keep AI analysis accurate?
- Profile first, ask second. Ask what each column means, the row count, the date range, duplicates and blanks, and nothing else yet.
- Make the cleaning decisions yourself. AI can find the duplicates. You decide whether they are real orders or a double export.
- Ask for formulas, not values. In a workbook, formulas let anyone trace a total. In a chat, ask it to run code for calculations instead of estimating.
- Tie out. Check one or two totals against the source system before anything goes to a stakeholder.
- Make it say what it assumed. Ask for a list of assumptions and anything it could not determine.
- Keep the raw file. Work on a copy, so you can rerun from the original.
What does this look like in a real lesson?
In Analyze a Spreadsheet, Maple Hollow Garden Center has six months of sales, about 1,200 orders, in a CSV. You upload it to ChatGPT and ask only for the columns, row count, date range and any duplicates or blank cells. It finds three orders that appear twice and 12 rows with no channel. You tell it to remove the duplicates, label the blanks "Unknown" and give you the cleaned file, and you open that file to look at it. Only then do you ask for monthly revenue and the fastest-growing category, get a line chart you can download, and finish with three takeaways, each with its number, plus one thing to double-check.
In Use case: messy export to clean Excel workbook, Claude takes Willow Creek Plant Shop's messy till export (two title lines, dates in three formats, a subtotal line, a repeated row, "house plants" spelled three ways) and returns a workbook with a Data tab, a Summary tab and a chart. The reply lists every fix, including two duplicate rows it removed. In Excel, cell B4 and the grand total in E8 show SUMIFS and SUM formulas, so changing a sale updates everything.
When an answer looks wrong, Catch Made-up Facts Before They Reach Customers teaches the fix routine: point at the exact problem, add constraints, ask what it assumed and check facts against a source.
Which tools do analysts use?
- ChatGPT for uploading CSV and Excel files, running analysis in code and producing charts.
- Claude for building real Excel workbooks with working formulas through its built-in Excel skill.
- Julius AI for asking plain-English questions of a spreadsheet and getting charts and forecasts.
- Copilot in Excel or Gemini in Sheets if your company already licenses them, so the data stays in your workspace.
For a tool-by-tool view, see which AI tool can analyze an Excel spreadsheet.
What mistakes should analysts avoid?
- Asking for insights before profiling.
- Accepting cleaned data you have not opened.
- Copying a number from the chat into a report without a formula or code behind it.
- Uploading sensitive data to an unapproved tool. Remove personal data your rules say should not leave the building.
- Letting a chart decide the story. Check the axis, the filters and the time frame.
Key takeaways
- Use AI around the analysis: profiling, cleaning, formulas, queries and write-ups.
- Profile first, then decide cleaning yourself.
- Ask for formulas or code so every number is traceable.
- Tie key totals back to the source before you share.
Frequently asked questions
- Can AI replace SQL for analysts?
It writes and debugs SQL quickly, but you still need to read the query and check the result, especially joins and filters. Knowing SQL makes AI more useful, not less.
- How big a file can I upload to ChatGPT or Claude?
Both handle typical exports of thousands of rows, with limits that depend on the plan. For very large data, work in your database or warehouse and use AI for the queries.
- How do I know the AI did not drop rows?
Ask for the row count before and after each step, and compare with the source. If the counts do not reconcile, ask it to list the rows it removed.