The payables clerk has a 38-tab workbook, a 4pm deadline, and a headache building behind her left eye. Column headers moved between tabs in March. Two vendors are spelled three different ways. Somewhere around row 4,112 a credit note has been entered twice.
So she does what a lot of people do now: uploads the file and asks ChatGPT to sort it out. Sometimes it works remarkably well. Other times it hands back a tidy-looking summary that is quietly, catastrophically wrong.
That gap is the thing worth understanding before you give a ChatGPT agent for spreadsheets any real work. Not the demo version. The version where the model writes a formula pointing at column J when your data sits in column K, and nothing about the output looks strange.
What makes it an agent rather than a chatbot
Asking ChatGPT to explain INDEX MATCH is a conversation. Giving it a file and a goal is a job. “Merge these four exports, flag every invoice where the total doesn’t match the purchase order, and give me a clean summary tab” is the kind of request that pushes the model into planning mode: it writes code, runs it in a sandbox, opens your workbook, manipulates rows, and returns an output.
Under the hood, the heavy lifting is Python. Library calls to pandas and openpyxl, with the model writing and re-writing the script until something runs without throwing an error. That single fact explains almost everything else in this article. It means the tool is genuinely strong at code-shaped data work, and it means the tool has no idea whether the answer is correct. It only knows whether the code executed.
The work it genuinely handles well
Anyone who spends a Friday afternoon on spreadsheet housekeeping should try this. The wins are real and they show up fast.
Reshaping messy imports
I fed it a 12,000-row ERP export where dates appeared in five formats, sku codes had trailing spaces, and the region field was a free-text box someone had clearly been abusing since 2019. It normalised all of it in one pass and produced a mapping log showing every change. Fifteen minutes of prompting against roughly three hours of manual work.
Writing and explaining formulas
This is the most underrated use. Ask for a formula in plain English and you get one, complete with a sentence on why the absolute references matter. Better still, paste in a nested monster inherited from a departed colleague and ask it to unpick the logic line by line. That’s a genuine time saver in any inherited workbook.
Chaining steps across files
Monthly reporting is mostly the same seven operations in the same order. Pull three files, join on a key, apply a mapping table, calculate a variance column, pivot, format, export. An agent can run that sequence and hand back a draft, which then only needs your review rather than your whole afternoon.
Requests that tend to land cleanly:
- “Deduplicate on invoice number, keep the row with the latest date, and tell me how many you removed.”
- “Unpivot the twelve month columns into a single date and amount column.”
- “Add a column comparing actual to budget and flag anything beyond 10% either way.”
- “Group by cost centre and give me total spend, row count, and the largest single line.”
- “Rewrite this formula so it works when rows are inserted above row 2.”
Where it goes wrong, and how silently
This is the part the demos skip. The failures are rarely dramatic. You don’t get a crash; you get a plausible number.
Rows vanish on joins. A merge on customer ID will happily drop every unmatched record without comment. If 400 rows had a trailing space in the key, they’re gone, and the totals still foot because the code did exactly what it was told.
Values get hardcoded. You asked for a variance column. It gave you numbers instead of formulas, so next month’s update breaks the whole sheet and nobody notices until the board pack is printed.
Rounding drifts. Floating-point arithmetic plus an agent’s cheerful confidence is a bad combination in a currency column. Sums that should match to the penny come back off by 0.0000001, or off by a rounding choice nobody documented.
Big files time out. Workbooks in the hundreds of megabytes, or with heavy pivot caches and external links, will hit sandbox limits. The agent may return a partial result rather than admit defeat.
Live documents aren’t live. Uploading a snapshot of a Google Sheet means working on that moment in time. Nothing syncs back, so a second person editing the original creates an immediate conflict.
The checks that keep your numbers honest
None of the above is a reason to avoid the tool. It’s a reason to build a review habit around it, the same way you’d review a junior analyst’s first pass.
Always work on a copy, never the source file. Ask for the plan before the output, in plain language, so you can spot a wrong assumption before it becomes 40 rows of wrong arithmetic. Then ask the question that catches most problems: how many rows went in, how many came out, and what was excluded? A good agent answers that happily. A bad answer is your signal to stop.
Reconcile one control total independently. If the original file shows £482,190 in column G, check that number against your cleaned output with a calculator, not with the model. Then eyeball twenty random rows by hand. It sounds tedious. It takes four minutes and it has caught a silently dropped join for me more than once.
For anything that feeds a report, ask for a change log tab listing every transformation applied. That artifact is worth more than the cleaned data, because it’s what lets someone else reproduce your result.
Where the human stays in the loop
Judgement calls don’t delegate well. Deciding whether a duplicate invoice is a data error or a legitimate split payment requires context the model doesn’t have. So does deciding which of two conflicting vendor records is the real one.
The pattern that holds up is agent for the mechanical work, human for the definition of done. If you want a clearer picture of where that boundary sits across different tasks, this breakdown of what ChatGPT agent mode can actually do and where you should step in is a sensible starting reference.
Regulated teams have an extra layer to think about: audit trails, data residency, and the awkward fact that a sandbox is not your document management system. There’s useful precedent in this piece on making AI operational in constrained public sector environments, which deals with exactly the problem of getting useful automation past a compliance review.
When a dedicated agent beats a general one
ChatGPT is a generalist that happens to be good with files. If spreadsheets are the whole job rather than one task among many, purpose-built platforms start to pull ahead, mainly because they hold a connection to your data source instead of a one-time upload.
Platforms designed for repeatable workflows let you define the pipeline once and run it on a schedule, with the joins and validation rules locked in place rather than re-improvised each session. That’s the model described in this guide to building a digital workforce with Relevance AI, and it solves the hardcoded-values problem by design.
Worth noting too: a spreadsheet task is often one step in a longer chain — pull the data, calculate the variance, email the regional manager, log the exception. If that’s your reality, a roundup of personal AI agents that actually finish the job will save you some trial and error, because very few of them handle the last mile well.
Start with the tab you resent most
Don’t begin with the workbook that matters. Begin with the one you dread, the reconciliation you keep putting off until Thursday. Run it through the agent, then run it again manually and compare. Time both. Note where the outputs diverge and why.
That exercise tells you three things you can’t get from a product page: how much time you actually save, which failure modes your particular data triggers, and whether you’d be comfortable signing your name to the result. Two or three cycles of that and the workflow becomes obvious. Some tabs go straight to the agent with light review. Others stay manual, because the setup cost outweighs the gain.
The clerk with the 38 tabs won’t get her afternoon back entirely. But a cleaned, reconciled version of that workbook at 2pm, with a change log she can defend, beats a heroic manual slog that finishes at 4:30 and never gets checked properly. That’s the realistic promise of an agent for spreadsheets. Not autonomy. A faster first draft, plus a paper trail.

