Google Sheets AI Agent: What It Costs and What Breaks

Somebody types "Delhi". Somebody else types "New Delhi". A third person types "delhi " with a space on the end. Your pivot table now shows three cities, and you are the one who straightens it out on Monday before the numbers go anywhere.

That is the chore. The question is which piece of software takes it off you.

Three completely different products answer to "Google Sheets AI agent" right now, and they do not do the same job. One writes a formula for you. One is a script you build on a Saturday and then own forever. One runs the job without you building or maintaining anything. Almost every confused question I get about spreadsheets and AI lives in that gap.

I read Zapier's, Lindy's and Relevance AI's pricing pages on 16 September 2026. None of the pages ranking for this search tells you what any of it costs. So the prices are here, dated, sources at the bottom.

Three different things people mean by a Google Sheets AI agent

Ask five people what they want here and you get three products back. The words are identical every time. The bill, the effort and the thing that shows up on Monday are completely different.

The AI inside the sheet

Google has put AI in the spreadsheet itself. There is a side panel that answers questions about the file you have open, and generation inside a cell so a formula or a first draft of a column arrives without you writing it.

It is genuinely useful, and it is the one option here that may already be sitting inside a subscription you pay for. Google's pricing page returned no machine-readable price when I checked on 16 September 2026, so I will not quote you a seat price I could not read.

What it will not do is run. You open the file, you ask, you accept the answer. Close the tab and nothing happens.

Build it once with a script

This is what most of the tutorials ranking for this query teach you to do. Open Apps Script, paste in code that calls a model, add a time trigger, and the sheet starts doing a thing on its own.

It works. I have no complaint about the approach, and for one narrow job it is the cheapest automation you will ever own.

The cost shows up later. Somebody holds the API key, notices when a model provider changes a response format, and is free the week the script quietly stops. If that somebody is you and you enjoy this sort of thing, close the tab and go build it.

Something that runs when nobody is watching

The third one is an agent proper. Your Google account is connected, it wakes on a schedule or a trigger, it reads the rows, it does work in other tools, and it writes results back where you can see them. What an AI agent actually is covers the general shape if the word is new to you.

This is the only one of the three that takes the chore away without you building or maintaining anything. It is also the one you are most likely to be billed for separately, and those two facts are the same fact.

On Gravity, Sheets connects through an ordinary Google sign-in. No API key to paste, no service account file to store, nothing sitting in a cell where the next person to open the sheet can read it.

One number on how central the spreadsheet is to all of this. Of the 88 agents in Gravity's launch catalogue, 55 touch Google Sheets. Only Gmail shows up in more of them. Whatever people say they want automated, a sheet is usually somewhere in the middle of it.

Quick comparison

Every figure here is what that vendor's own pricing page showed on 16 September 2026; cells marked as unreadable or not recorded are stated as such. Zapier's $19.99 is the monthly rate; the same page shows $160 a year billed annually, or $13.33 a month. Relevance AI publishes no readable self-serve number and Google Workspace's page returned no machine-readable price, so neither is guessed at here. Prices move.

Cleaning the column somebody typed badly

This is the job that works best today and the one people ask for first. The agent reads a range, applies a rule, and either fixes the cell or flags the row for a person. Normalise "Delhi" and "New Delhi" to one spelling. Strip the trailing spaces. Force one date format across a column that currently has three. Drop the second copy of a duplicate, or keep both and colour them so somebody can decide.

What decides whether this works is the rule, not the model. "Clean this up" is a wish. "Any value in column C that is not one of these fourteen approved values, copy that row to a Review tab and leave the original exactly where it is" is a rule, and a rule is what survives contact with a real sheet.

Write yours down before you go shopping. If you cannot write it in one sentence, no tool on this page will guess it. The step-by-step version of this job has the rule shapes that hold up over a few months.

A weekly summary, written back to its own tab

Every Monday somebody reads last week's rows and types four numbers into a message. Twenty minutes. The same twenty minutes, every week, forever.

A scheduled agent does this the boring way: read the rows inside a date range, compute the totals, write a dated block into a Summary tab, stop.

The write-back is the part that matters, and the part in-sheet AI does not give you. An answer in a chat panel is something you have to go and fetch. A dated row in a tab is a record your accountant can open in March. Ask any tool you are weighing whether it writes into the sheet or only talks about it.

Keep the summary on its own tab and never let the agent touch the source rows. Two reasons. You can delete a bad summary without losing data, and your version history stays readable on the day something goes wrong.

When a row should become an email

A lot of sheets are queues pretending to be spreadsheets. Overdue invoices. Leads waiting on a second follow-up. Documents promised on Friday.

The agent's job here is small and specific: pick the rows that match, fill the template, send, then write the date it sent into a column so it never sends the same message twice.

That last clause is the entire design. Without a sent-at column you will eventually email one customer four times in a week, and the apology costs more than the automation ever saved you.

On Gravity you can require your approval before an agent messages a customer. The inbox side of this job has its own post: the Gmail agent walkthrough goes through what approval looks like in practice. Slack behaves the same way, with a channel in place of an address.

The status column flips to Paid

The other pattern is a trigger rather than a schedule. Somebody edits one cell and something should follow: the chasing stops, a task closes, a message goes out to whoever was waiting on it.

Ask one question before you buy anything for this job. How often does it look?

Some watch the edit and react in seconds. Others poll on an interval, which makes your "instant" reaction only as fresh as the last poll. Both designs are fine, but the interval is rarely on the marketing page, so ask before you buy.

What tends to happen next is that the sheet quietly becomes the control panel for a process living in four other tools. The same pattern turns up in a database, where the row you edit is a record; the Notion agent post covers that version of it.

What a Sheets agent still cannot do

Nobody selling one of these writes this section, so here it is, starting with the failure people hit first and recognise last. Read your whole sheet and reason about all of it. A model reads what fits in the space it has, and a wide sheet fills that space far faster than a narrow one. There is no published row count where it breaks, because the number depends on your columns. The dangerous part is that it does not refuse: it answers with confidence about the slice it managed to read. Give it a filtered view, a named range or a date window instead of the whole file.

Know what your columns mean. A header that says "Amt2" tells the agent exactly as much as it would tell a new hire on day one. Rename your headers into English before you blame the model.

Survive you reorganising the sheet. Insert a column on Tuesday and anything that addresses cells by position starts writing into the wrong place. Rules that reference header names hold up. Rules that reference column F do not.

Undo itself. An agent that writes bad values across 4,000 rows has made 4,000 edits, and Sheets version history will get you back at the cost of an afternoon. This is the real argument for pausing on anything customer-facing, and for keeping writes on a Summary tab rather than in live data.

Rescue a sheet that should have been a database. If your formulas already take a second to recalculate, an agent sitting on top inherits every problem the file already had.

One distinction is worth keeping straight, because the tutorials blur it constantly. An AI that writes you a formula speeds up your own work. An agent that runs the job every Monday takes the job off you, and that is the one with a monthly bill.

Which should you start with?

Start from how often the chore comes back. That sorts this faster than any feature list.

Once, or twice a year: use the AI inside the sheet, because nothing else here is worth a subscription for a job you do in March.

One narrow job and a developer on the team who enjoys this sort of thing: Apps Script, for the price of an afternoon and the responsibility of owning it.

The sheet is one hop from another app and you want to draw that hop yourself: Zapier, at $19.99 a month for Professional, or $160 a year billed annually, as read in September 2026.

Per-seat pricing suits a small team and you are happy to buy it that way: Lindy, at $29.99 per user a month as of September 2026. Multiply by the number of people before you decide anything.

Your company has a procurement process and a security questionnaire: Relevance AI, which publishes no self-serve price on its pricing page. No self-serve price on the page usually means a sales call, so budget for one.

The file is an .xlsx living in OneDrive: different post, different connection story. The Excel version of this question is the one to read, because the chore is identical and the plumbing is not.

You want the chore gone and you have no interest in building anything: that is what Gravity is for. The first agent is free with no card. After that Pro is $5 a month during the alpha; You can buy more usage if you run out. Gravity is in private alpha, so you apply rather than sign up.

Still comparing: the no-code AI agent platforms roundup covers the ones that hand you a builder, which is the fork in the road this whole post is about.

My own bias, said plainly: for a chore that repeats every single week, I would pick the option where nobody builds anything. Agents listed by profession is the fastest way to check whether yours has already been written.

FAQ

Sources