Quick answer: AI assistants such as ChatGPT, Claude, Gemini and Copilot are genuinely useful for spreadsheet work — explaining a formula you inherited, drafting one from a plain-English description, or cleaning up a messy text column. They are also wrong often enough that checking their output is part of the job, not an optional extra. The dangerous failures rarely produce error messages; they produce formulas that run cleanly and return a plausible, incorrect number. Work on a copy of the sheet, reconcile totals before and after, and verify each function against official documentation before the result informs a real decision.
Spreadsheets deserve special scepticism because of what sits on top of them. Prices quoted to customers, VAT returns, staff hours, stock reorders: for many small businesses, the spreadsheet is where numbers become decisions. A chatbot answer that is wrong wastes a minute; a spreadsheet total that is quietly wrong can misprice a job or misstate your accounts. That is why "the AI wrote it and it looked right" is not a safety standard — and why the routine below exists.
Where AI genuinely helps in a spreadsheet
Used for the right jobs, an assistant saves real time in Excel and Google Sheets:
- Explaining an inherited formula. Paste something like
=SUMIFS(D2:D201,E2:E201,"Paid",B2:B201,">="&DATE(2026,1,1))from a former colleague's file and ask what it does. Modern assistants are good at translating formula syntax into plain language, including the parts that matter, such as the&that joins a comparison to a date. - Drafting a formula from a description. "Total column D where column E says Paid and the date in column B is this month" is usually enough for a usable first draft.
- Cleaning messy text columns. Splitting "Smith, J." into surname and initial, standardising phone number formats, or stripping stray spaces — the assistant can suggest an approach and the formula to go with it.
- Navigating built-in features. "How do I remove duplicates without deleting the rows next to them?" gets a correct step-by-step answer faster than most searches.
Notice what is missing from that list: anything whose result must be silently trusted. That is not a flaw of one tool. These systems generate likely answers, not verified ones — and the verification step stays with you.
A small example that shows the failure mode
The example below is constructed for illustration, but it mirrors what comes out of real invoicing and booking tools.
You export 200 invoices: invoice number in column A, date in B, client in C, amount in D, status in E. The status column is free text, so the export contains Paid, PAID, paid — with a trailing space the tool inserted — and Pending.
You ask an assistant: "Write a formula that totals the amount in column D for rows where column E says Paid." It answers:
=SUMIF(E2:E201,"Paid",D2:D201)
The formula runs without errors and returns a number. It is also quietly wrong. SUMIF criteria are case-insensitive, so PAID is counted — but paid with a trailing space is a different piece of text from Paid, and those rows are excluded without any warning. Your total is short by exactly the invoices the export labelled inconsistently. Nothing looks broken.
What catches this is a reconciliation, and it takes seconds. Count the rows that mean Paid regardless of formatting:
=SUMPRODUCT((TRIM(UPPER(E2:E201))="PAID")*1)
Compare that with =COUNTIF(E2:E201,"Paid"). If the first number is bigger, rows are being missed — and the total that counts them all is:
=SUMPRODUCT((TRIM(UPPER(E2:E201))="PAID")*D2:D201)
Often the better fix is not a cleverer formula but cleaner data: add a helper column with =TRIM(UPPER(E2)), fill it down, and point the simple SUMIF at the clean column. Fix the data, and the simple formula becomes trustworthy.
The checking routine
- Ask for the reasoning with the answer. "Explain your assumptions and how I could test this" turns a confident monologue into something checkable. An assistant forced to state its assumptions sometimes corrects itself before you run anything.
- Never work on the live file. Duplicate the sheet, run the AI-assisted formula on the copy, and compare results before anything touches your real data.
- Reconcile before and after. Row counts, totals, counts of unique values — note them before the change and check them after. The example above was caught by a reconciliation, not by an error message.
- Test the edges on purpose. Blank cells, duplicated keys, numbers stored as text, values with trailing spaces. Ten deliberate edge cases tell you more than a hundred well-behaved rows.
- Verify unfamiliar functions against official documentation. Assistants occasionally produce plausible but non-existent functions or arguments. Google publishes a function list for Sheets and Microsoft maintains an alphabetical reference for Excel functions; both take seconds to search.
- Hand-check a small sample. Pick three rows, work out by hand what the result should be, and compare. If your hand calculation and the formula disagree, trust the hand calculation until it is proven wrong.
This is the same discipline I apply in my own spreadsheet and automation work: understand what the tool produced, where its limits are, and whether the people who rely on the result can check it too.
Three details that trip people up
Argument separators. With UK regional settings, Excel and Google Sheets separate formula arguments with commas. On systems set to other regions — Portugal among them — the same formulas expect semicolons: =SUMIF(E2:E201,"Paid",D2:D201) becomes =SUMIF(E2:E201;"Paid";D2:D201). If a pasted formula errors immediately and the syntax looks fine, check the separators.
Dates that are not dates. Exports often store dates as text. Sort by the date column: if the order is nonsense, you are sorting text, and date-based formulas will misbehave. Watch the format itself, too: 03/04/2026 is 3 April in the UK and Portugal but 4 March on systems set to US format. Spot-check the earliest and latest rows before trusting any monthly totals.
Lookups that silently approximate. VLOOKUP and its relatives use approximate matching when you leave out their final argument, and an approximate match on unsorted data can return the wrong row with no error. Always pass FALSE (or 0) explicitly for an exact match.
The privacy step people skip
An invoice export is a client list: names, sometimes email addresses, addresses, phone numbers. Pasting it into an AI assistant shares that data with a third-party service. Three habits keep this sensible:
- Replace identifying columns before pasting. Client A, Client B, Client C. The formula does not care what the client is called.
- Send only what the task needs. If the question is about amounts and statuses, the client column can stay out of the prompt entirely.
- Check what the service does with your content. Data-use terms differ between tools and between plan tiers — read them for the tool you actually use.
Anonymising first also makes the technical checking easier: you are testing structure and logic, not handling live records.
When you do not need AI
- The task is one
SUMorAVERAGEaway. - Sorting, filtering, Remove Duplicates, Text to Columns, Pivot Tables: these features are built in, and finding the menu option is faster than holding a conversation.
- You already know the formula. Asking is slower than typing.
- The data is too sensitive to share even anonymised. Ask about the approach in the abstract, or do it by hand.
AI earns its place when the logic is fiddly, the syntax is unfamiliar or the cleanup is repetitive — not as a default for every cell.
Two questions that come up
The formula returned a number and no error — isn't that proof it worked? No. Errors announce themselves; wrong answers do not. A returned number is an output, not a verification. The reconciliation in step 3 is what turns "it ran" into "it is right".
Is it safe to paste my accounts into an AI chatbot? Treat a pasted table like an email to an outside company: send the minimum, strip identifying details, and check the tool's data-use terms first. If the data cannot be anonymised, describe the columns in words and ask for the formula pattern instead of pasting the contents.
One task, checked properly
If you would rather work through one real spreadsheet task with someone beside you, that is what the Personal AI Session is for: 60 minutes online, one-to-one, on one task you prepare and anonymise in advance — a cleanup like the example above, a formula you inherited, or a report you rebuild by hand every month. It costs £75 per session (89 € in Portugal), the final total for the scope listed, and it includes a short written recap within two working days plus a follow-up clarification email thread. Requesting a session creates no booking or payment obligation.
Still deciding? Book a free 30-minute call — no preparation needed, and if training is not the right fit, we will say so.
Related reading: common AI mistakes small businesses make.
Last updated: 22 October 2026