Most small businesses have one person who spends half a day every month turning spreadsheet exports into a report. Sales by rep, jobs by status, overdue invoices by customer. AI tools can now do a large share of that work. They also get numbers wrong in ways that are easy to miss, and a wrong number in a report is worse than a late report.

This is what Microsoft Copilot and ChatGPT do well with spreadsheets, where they fail, and how to set up a monthly report you can trust.

What each tool does with a spreadsheet

Copilot in Excel works inside the workbook. It needs a paid Microsoft 365 Copilot license, roughly $20 to $30 per user per month on top of your existing Microsoft 365 plan, and the file has to be saved in OneDrive or SharePoint. You ask in plain English and it adds formula columns, builds pivot tables and charts, and highlights rows that meet a condition. The result stays in Excel as formulas you can click on and check.

ChatGPT works on a copy. You upload the file, it writes and runs code against the data, and it gives back a summary, charts, or a new spreadsheet. Plus costs $20 a month. The Business plan runs about $25 to $30 per user and keeps your data out of model training by default, which matters if the sheet has customer or payroll data in it.

Pick one report you build every month and note how long it takes by hand. That is your baseline for judging either tool.

What they do well

  • Cleanup. Splitting full names, fixing date formats, trimming stray spaces, flagging duplicates. Tedious by hand, fast and reliable with AI.
  • Formulas. Describe what you want and get the formula back, with an explanation. This alone saves the average Excel user hours of searching.
  • First draft summaries. Totals by month, top ten customers, a chart of revenue by service line. Good for a first look at an export you have not seen before.
  • The written part. Once the numbers are settled, both tools turn a table into three paragraphs of commentary for an owner or a board in under a minute.

Start with cleanup and formulas. The output is easy to check and the time saved shows up the first day.

Where the numbers go wrong

The errors are rarely arithmetic. They come from the tool misreading the sheet or guessing what you meant.

  • Messy layout. Merged cells, two header rows, subtotal rows in the middle of the data, and blank rows all cause trouble. A subtotal row counted as a sale doubles your revenue.
  • Unstated definitions. Ask for "sales last quarter" and the tool decides for itself whether refunds count, whether to use invoice date or payment date, and whether your quarter is calendar or fiscal. It will not always tell you what it chose.
  • Numbers stored as text. A column where some amounts are text gets partly skipped. The total looks plausible and is short by whatever was skipped.
  • Large or multi sheet files. Past tens of thousands of rows, or when the answer needs data joined across tabs, both tools get less dependable. ChatGPT may quietly work on a sample, and Copilot may decline or answer from part of the data.
  • Answers without calculation. If a chatbot reads a pasted table and replies with a figure without running any code, treat that figure as an estimate. Language models are poor at adding long columns in their head.

Before you ask for any analysis, format the data as a proper Excel table: one header row, one record per row, no merged cells, no subtotals. That 15 minutes removes most of the failures above.

A checking routine that takes ten minutes

  1. Reconcile one total. Check the grand total against a number you already trust, such as the figure in your accounting system. If those two disagree, nothing else in the report matters yet.
  2. Ask for the method. Have the tool state the row count it used, the filters it applied, and the date range. In Copilot, read the formula. In ChatGPT, open the code it ran.
  3. Spot check three lines. Pick three customers or three reps and add them up by hand with a filter.
  4. Write the definitions into the request. "Sales means invoiced amount, net of credit notes, by invoice date, calendar quarter." Save that wording and reuse it every month.

Put the reconciliation step in writing for whoever builds the report, and make it a rule that no AI built figure goes to a client, a lender, or the tax return without it.

When to stop using a chatbot for it

If the same report runs every month from the same export, asking a chatbot to rebuild it each time is the wrong setup. Every run is a new chance for it to interpret the data differently. Use the AI once to help build a fixed workbook with a data tab, pivot tables, and formulas. After that, the monthly job is to paste in the new export and refresh. The numbers come from formulas that do the same thing every time, and the AI is left with the commentary.

List your recurring reports and mark which ones are the same every month. Those are candidates for a fixed workbook, and the rest are fine to do ad hoc with a check.

If you want help deciding which reports are worth setting up this way, the first call is free and takes about 30 minutes. Bring one report and the export it comes from. Book the call here.