Grab the four Google Sheets prompts that do the grunt work, and the one test for when to keep the formula
Google slid a spreadsheet builder into the plan you already pay for, and the trick is knowing the exact jobs it does well before it quietly falls apart on the big ones.
TLDR: Gemini in Google Sheets can now build and edit whole spreadsheets from a plain-English prompt, and for most Business and Enterprise plans it’s already switched on inside what you pay for every month. It’s genuinely good at a short list of bounded jobs, categorizing a messy column, cleaning a list, drafting a formula, standing up a small dashboard off your own data. It falls apart the second you point it at a big multi-file, cross-app, auditable model. Below are four prompts you can paste into a sheet today, and a two-second test for which jobs to hand it and which ones to keep in a formula.
Almost every operator I work with has one spreadsheet a person babysits every single week. Expense categorization that somebody retypes by hand, a lead list that shows up messy from three sources, a little P&L view that a bookkeeper rebuilds every month because the export never lands clean. It’s the least glamorous work in the building, and it’s usually either eating an owner’s Sunday or getting paid for by the hour.
That’s the work Google just quietly moved into a spreadsheet you’re already paying for. Back on April 22 Google turned on Gemini in Google Sheets, the thing that builds and edits entire spreadsheets from a sentence, and it’s been rolling wider through the summer, with the data-entry piece, Fill with Gemini, picking up 11 more languages in July (Google Workspace Updates).
For the last three years I’ve watched the tools that used to earn a monthly retainer get pulled, one by one, into software operators already have open. This is that pattern landing on the spreadsheet. The catch is that Gemini in Sheets is great at a narrow band of work and quietly bad outside it, and if you don’t know where that line sits you’ll either miss the free win or trust it with a number the bank is going to read.
By the end of this you’ll have four prompts you can paste into a sheet today, and a two-second gut check I’m calling The Bounded-Job Test for deciding which spreadsheet jobs to hand Gemini and which ones to keep in a formula.
Who this is for: operators running on Google Workspace (Business or Enterprise) who touch spreadsheets weekly, or pay someone who does.
Here’s what you’re walking out with:
The free win most operators are sitting on: the spreadsheet builder is already inside Business Standard and up, no new line item, no separate tool.
Four copy-paste prompts: categorize, clean, formula, and a small dashboard off your own numbers, the jobs it nails.
The Bounded-Job Test: a two-second check that keeps you from handing it the one task that’ll bite you.
The failure line a finance guy already hit: a real tester ran the big cross-app forecast and it fell over, so you don’t have to learn that one the hard way.
What just shipped, in plain terms
Gemini in Sheets does two things that used to need real spreadsheet chops. It builds, so you can say “build a P&L dashboard from my sales and service data” and it drafts a plan, pulls the numbers, and lays out formatted tables, pivots, and charts. And it edits in the side panel, so “add scorecards and a bar chart above my inventory data” happens next to your sheet instead of making you rebuild anything.
The part that makes it more than a toy is what Google calls Workspace Intelligence, which lets Gemini reach across your Gmail, Drive, and Chat to pull context into the sheet instead of you copy-pasting between an inbox and a tab. On Google’s own spreadsheet benchmark the model hit a 70.48% success rate on messy real-world tasks, which they frame as near-expert and which, in operator terms, means it’s right most of the time and wrong often enough that you still have to look (The Keyword).
Where’s the money angle. It’s bundled. Gemini in Sheets is included in Business Standard and Plus, Enterprise Standard and Plus, the AI add-ons, and the consumer Google AI Pro and Ultra plans. If you’re on one of those, you’re already paying for this, it’s just sitting in the side panel behind the Gemini icon waiting for a prompt. The one thing worth flagging is limits: Google’s promotional higher limits ran through July 15, and after that per-user caps apply, with paid AI add-on seats getting more headroom, so heavy users hit a ceiling light users never will.
The four jobs to hand it, with the exact prompts
The jobs Gemini in Sheets does well share a shape: bounded scope, your own data, and an answer you can eyeball in a few seconds to know it’s right. Here are the four I’d reach for first, each with the prompt, what a good result looks like, and the moment to stop and check.
Job 1: Categorize a messy column
The move: point Fill with Gemini at a column of raw text, transaction memos, lead sources, support tags, and have it write a clean category next to each row. Select your data, open Fill with Gemini in the side panel, and give it the rule.
In a new column next to this transaction list, label each row as one of:
Software, Payroll, Contractors, Ads, Rent, Travel, Meals, Other.
Base it on the vendor name and memo text. If you're unsure, use "Other"
and do not invent a new category.
What a good result looks like: clean single-word labels that match the eight buckets you named, “Other” showing up on the genuinely ambiguous rows instead of a guess. Google says Fill with Gemini runs about 9x faster than doing 100 cells by hand, and for this kind of labeling that’s roughly the lift you’ll feel.
The failure mode: on long columns the labels start to drift, a category like “Service” mutates into “service issues” a few hundred rows down, or it quietly invents a ninth bucket you didn’t ask for. That’s your signal the run got too big, chop it into batches of a couple hundred rows and add “do not invent a new category” to the prompt, which I already baked in above for that reason.
Walk away with: a categorized column in seconds instead of an afternoon, on the kind of data you’d otherwise pay someone hourly to tag.
Job 2: Clean up a list that came in three different shapes
The move: hand it a column that three sources filled in three different ways and have it standardize the format without you writing a nest of formulas.
Standardize this "Company" column: strip "Inc", "LLC", "Ltd", and
trailing punctuation, fix obvious capitalization, and put the cleaned
name in the next column. Leave the original column untouched.
What a good result looks like: “acme co., llc” and “ACME CO” both land as “Acme Co” in the new column, and the original stays put so you can compare. Telling it to leave the source column alone is the whole game, you always want the before and after side by side.
The failure mode: it “helpfully” merges two companies that were genuinely different, or rewrites a name it didn’t recognize. Eyeball the rows where the cleaned value differs a lot from the original, that’s where a real error hides, and it’s a five-second scan when the columns sit next to each other.
Walk away with: a deduped, consistent list you can run a mail merge or an import against, without babysitting a formula chain.
Job 3: Draft or explain a formula in words
The move: describe the calculation you want in plain English and let it write the formula, or paste a formula you inherited and ask what it does. This is the one even the skeptics agree is solid.
Write a formula for cell H2 that returns the gross margin percent using
revenue in column D and cost in column E, formatted as a percentage,
and returning blank if revenue is zero. Then explain in one sentence
what it does.
What a good result looks like: a working formula plus a one-line plain-English explanation, so you’re not pasting something you can’t read. The explanation is the tell, if it can describe the formula in a sentence that matches what you asked, it usually built the right thing.
The failure mode: it returns a formula that runs without erroring but calculates the wrong thing, off-by-one on a range, wrong column reference. This is why you ask for the explanation, if the sentence doesn’t match your intent, the formula won’t either, and you catch it before it’s buried in a model.
Walk away with: formulas you’d have Googled for twenty minutes, written and explained in one pass.
Job 4: Stand up a small dashboard off your own numbers
The move: give it a clean tab of your own data and ask for a one-screen summary. Keep it to data that already lives in the sheet, not a cross-app pull.
Using the data on the "Sales" tab, build a one-page summary above the
data with: total revenue, revenue by month as a bar chart, top 5
products by revenue, and month-over-month growth. Keep it on this sheet
and don't change my raw data below.
What a good result looks like: a tidy header block with a couple of scorecards and one clean chart, your raw data untouched underneath. For a quick read on how a month went, this is the job it was built for, Google’s own example is literally a small-business P&L view.
The failure mode: ask it to reach into three other files or your inbox to build the same thing and reliability drops off a cliff, which is the exact wall the next section is about. Keep the dashboard job pointed at data that’s already in the sheet and it holds up.
Walk away with: a shareable one-screen view of a month, built in the time it takes to describe it.
The Bounded-Job Test, and where this thing breaks
Every job above passes the same three-part check, and that’s the test worth memorizing. Before you hand Gemini a spreadsheet task, ask: is the scope bounded to one sheet or one column, is it working off data you already have in front of you, and can you eyeball the answer in a few seconds to know it’s right. Three yeses, hand it over. A no on any of them, keep it in a formula or keep it human.
The reason that test matters is that the failure isn’t hypothetical, and someone already ran it into the ground so you don’t have to. Wouter Born, a CFO-tech investor who writes these tools up for finance teams, tried the exact workflow Google’s architecture is supposed to nail: three store managers email in monthly forecast updates with attachments, and Gemini reads the emails, pulls the assumptions, opens the files, maps them to the corporate model, and flags what needs sign-off. The dream cross-app job. His verdict:
“The vision is right. But the execution isn’t there yet... The complex, multi-file, cross-app workflow that makes Google’s architecture special isn’t reliable enough for finance yet.” (AI CFO Office)
He didn’t write it off, and neither would I. He landed right where the Bounded-Job Test lands, on what to use it for:
“For now, if your team runs on Google Workspace, use Gemini for the things it does well: formula generation, data enrichment, quick analysis questions, and spreadsheet cleanup. Those features work and they’re included in what you’re already paying for.” (AI CFO Office)
Two more honest tradeoffs before you lean on it.
It’s got no formula trail. When Gemini fills a number, that number isn’t the output of a formula you can click into and trace, it’s a generated value sitting in a cell. That’s fine for a lead category, it’s a real problem for anything an accountant, a lender, or a tax reviewer has to trace back to a source. For auditable numbers, keep the formula and let Gemini draft or label around it, don’t let it be the thing that produces the figure of record.
And it runs out of room. Push a task past a few hundred cells in one go and it tends to stall, refuse, or hand back a partial answer, there are real session limits under the hood on top of the per-user caps that kicked in after July 15. Big jobs need chopping into bounded ones anyway, the same discipline the test pushes you toward.
Failure modes, in one place
You point it at the auditable P&L. A generated number with no formula trail is the one thing you don’t want a lender reading. Keep those in formulas, use Gemini to label around them.
You run it on 2,000 rows at once. Labels drift, categories wander, output comes back partial. Batch it into a couple hundred rows at a time.
You ask for the big cross-app pull. Read three inboxes and four files in one shot, and it falls over, that’s the Born test. Keep the job inside one sheet.
You trust it without the eyeball. If you can’t verify it in a few seconds, it wasn’t a bounded job.
Where this fits the bigger picture
The operators I work with who get real value out of AI aren’t the ones chasing every launch, they’re the ones who’ve gotten sharp about which jobs to hand a machine and which ones stay human, and it’s the same muscle whether the tool is Sheets, your CRM, or your inbox. A free spreadsheet builder is a nice win. Knowing the line between the grunt work you give it and the number you protect is the part that compounds.
That judgment, which repeatable jobs to productize into an AI workflow and which to leave alone, is most of what I trade notes on with operators inside the Abra AI community, where the skill files and workflow builds live. If you’d rather map it against your own stack directly, that’s what an AI Clarity Call is for.
Recap
Gemini in Sheets is already in the plan most operators pay for, and it’s good at bounded jobs: categorizing a column, cleaning a list, drafting a formula, building a small dashboard off your own data. Run those with the four prompts above, and run everything else through the Bounded-Job Test, bounded scope, your own data, an answer you can eyeball. When a task fails it, keep the formula and keep the human.
Tomorrow: the one inbox job I’d hand an AI before I’d hand it a spreadsheet.
Andrew
P.S. A few worth pulling up next:
The tool-compression pattern this fits into: Run the 30-minute Canva test that shows whether your ad creative retainer still earns its invoice
The Google-side listings version of the same story: Run the three-prompt Google Business Profile check
The context brief that makes any of these prompts land better: Grab the 5,000-character ChatGPT brief
Where the workflow builds and skill files live: whop.com/abra-ai
Map it to your own stack: muddventures.com/book