Google Sheets AI, the =AI() Function for Finance
Google put Gemini inside Sheets as a formula. You can type =AI(“categorize this expense”, A2) into a cell and get an answer back, which sounds like the end of a lot of tedious work.
It is, for one specific kind of task. It is also the most dangerous-looking thing Google has shipped into a spreadsheet, because it looks like a formula and does not behave like one.
Here is what it does, the four properties you need to know before you put it near a model, and the finance job it is genuinely excellent at.
Three different things share the name
When people say Google Sheets AI they mean one of three things, and they behave nothing alike.
| What | Where it lives | What it is for |
|---|---|---|
| The Gemini side panel | A panel next to your sheet | Asking questions, explaining formulas, talking through an approach |
| The =AI() function | In a cell, like any formula | Generating, summarizing and categorizing text at volume |
| Building a sheet from a prompt | A conversation that produces structure | Standing up a tracker or a template from scratch, added April 2026 |
The middle one is the interesting one and the one this page is mostly about.
Four things to know about =AI() before you use it
You can write it as =AI() or =GEMINI(). Both run Gemini against whatever cells you point at.
- It does not recalculate, which is the big one. You have to click Generate and Insert, or Refresh and Insert, every time. A sheet full of =AI() calls will hold last week’s answers against this week’s data until somebody refreshes them.
- Responses are text only. Google is explicit about it. This is not a calculator and it was never meant to be one.
- The result is attributed to you in version history when it inserts. If your review process runs on who changed a cell, this reads as a human edit.
- It batches to 350 cells at a time, and there are longer-term rate limits that lock you out for 24 hours if you push through them.
Take those four together and the shape becomes clear. This is a bulk text-processing tool that happens to live in a cell, and treating it as a formula is how people get hurt.
One more, from Google’s own documentation: they explicitly say not to rely on these outputs as financial or professional advice. That is boilerplate, and it is also worth reading twice given where you are about to point it.
The finance job it is genuinely excellent at
Categorization at volume. That is the use case, and it is worth the setup on its own.
Every finance team has a column of free-text descriptions that somebody maps by hand. Expense line items. Vendor names spelled four ways. Customer feedback that needs a sentiment. Transaction memos that need a GL code.
That work is slow, it is boring, and it is exactly what a language model is for.
=AI("Categorize this expense into one of: Travel, Software, Professional Fees, Facilities, Marketing, Other. Return only the category name, nothing else. If the description does not clearly fit one, return Other.", A2)Two details in that prompt matter more than the rest. Constraining the output to a fixed list stops it inventing a nineteenth category on row 400. And telling it what to do when the answer is unclear stops it guessing, which is the difference between a column you can review and a column you have to redo.
Run it, spot-check thirty rows, then paste the results as values. Do not leave live =AI() calls in a sheet anyone else will open, because the next person will assume the numbers refresh.
The prompts, written out
The AI library for finance teams
The categorization prompt above, the mapping prompts, and the review checklist for AI-generated columns. Free, and it lands in your inbox in about a minute.
What it must never touch
Arithmetic. Any of it.
The rule that applies to every model in every spreadsheet applies double here, because =AI() sits in a cell and looks exactly like SUMIFS sitting next to it. A number produced by =AI() is a piece of generated text that happens to contain digits. It did not add anything up.
So: have it write the formula, and let Sheets do the calculation. Ask it in the side panel for the QUERY or the SUMIFS you need, put that in the cell, and keep =AI() for the text columns.
The same discipline, applied on the Microsoft side, is in Python in Excel.
Building a sheet from a prompt
Since April 2026 Gemini will build and edit whole spreadsheets from a conversation. Describe a tracker and it produces the structure, the headers, the formulas and the formatting.
For a blank-page problem this is genuinely useful. A project budget tracker, an expense log, a simple forecast layout, all in about a minute instead of twenty.
What it is not is a way to build your actual model. Anything with dependencies across tabs, anything auditable, anything someone will still be maintaining in two years wants a person deciding the structure. Use this for the scaffolding and take over early.
Sheets or Excel for this
If your company is on Workspace, you use Sheets and the question does not arise. But the honest comparison, for anyone with a choice:
| Google Sheets | Excel | |
|---|---|---|
| Bulk text categorization | Better. =AI() over a column is hard to beat. | Possible, more setup |
| Real modeling | Weaker at scale and past a few hundred thousand rows | Still where finance models live |
| Collaboration | Better, and it has been for a decade | Improving, still second |
| AI inside the workbook | Good and improving fast | ChatGPT for Excel and Claude for Excel are further along |
My split: text work in Sheets, model work in Excel, and stop trying to make either one do the other’s job. AI in Excel covers the Excel side, and Google Gemini for finance is the wider Gemini picture including Gemini Notebook.
Apps Script still matters
None of the above schedules anything. If you want a report to build itself on the first business day and land in someone’s inbox, that is still Apps Script, and it is still the most underused thing in Workspace.
The change is that you no longer have to write it. Describe the trigger and the steps to a model, paste the result into the script editor, and read it before you run it. Two questions: what does it read, and what does it write.
For the wider automation picture, finance automation is the map, and automating the month-end close covers the close specifically.
Where to start
Find the column you map by hand every month. Copy 100 rows into a fresh sheet, run the categorization prompt against them, and check every one against what you would have written.
Count the disagreements. Under five and you have found a real hour back every month. Over twenty and your categories are ambiguous, which is worth knowing for reasons that have nothing to do with AI.
