AI Spreadsheet Tools 2026: Rows, Bricks, Copilot Compared

AI spreadsheet tools let you describe a calculation in words and get a formula, a chart, or a cleaned column back. They come in three kinds: assistants that write formulas and edit your sheet (Excel Copilot, Gemini in Google Sheets), functions that call a language model from inside a cell, and AI-native products that build analyses and dashboards (Rows, Bricks, Quadratic). This guide compares them using vendor pages read in October 2026, explains the one big change of the year (Microsoft removed Excel's in-cell COPILOT function), and shows how to check an AI-written formula before you trust it.

---

The short answer

Tool	Type	Best for	Pricing (checked or reported, October 2026)
Excel with Copilot	Assistant and agent inside Excel	Teams already on Microsoft 365	Needs a Copilot license or an eligible Microsoft 365 plan; check current price
Google Sheets with Gemini	Assistant plus an AI cell function	Teams already on Google Workspace	Part of Workspace plans per Google
Rows	AI-native web spreadsheet	Pulling data from SaaS tools	Free; Plus $8 per user monthly; Pro $79 per month plus $8 per user
Bricks	AI analyst and dashboard builder	Non-technical teams that want charts from files	Free tier and paid plans per its site
Numerous.ai	Add-on that adds an AI function	Batch text tasks in Sheets or Excel	About $10 a month billed annually
Quadratic	Spreadsheet with Python, SQL, and AI	Analysts who want code and a grid	Free; Pro $18 per user monthly, billed annually

If you already live in Excel or Sheets, the built-in assistant is the lowest-friction place to start, and switching to a new spreadsheet product only for AI rarely pays off. The exceptions are teams that pull a lot of data from outside services, where Rows or Quadratic can save real plumbing work.

---

Three ways AI shows up in a spreadsheet

The marketing blurs these together, and they behave differently.

1. The assistant. You type a request ("add a column that flags orders over $500") and the tool writes a formula or edits the sheet. The model works once, you see the result, and from then on the sheet recalculates with ordinary formulas. This is the safest pattern because the output is a formula you can read. It works like a driving instructor who points at the turn: you still steer.

2. The AI cell function. A function such as Google's AI function or Numerous's =AI() calls a model from a cell, so the model's answer becomes the cell's value. This is useful for text tasks across many rows, such as classifying feedback or extracting a company name. It is also the riskiest pattern, because a language model is not deterministic: the same prompt can return different text on different runs, and the answer is regenerated whenever the sheet recalculates. It works like asking an intern to fill in each row from memory.

3. The analyst. A tool reads your file, decides what to compute, and builds charts, tables, or code. Rows' AI Analyst, Bricks, and Quadratic fit here, as does Excel's Agent Mode. The value is speed on exploratory questions. The risk is that the work happens in steps you did not write, so you need to inspect them.

A spreadsheet is an unusual place for a language model because spreadsheets are built on determinism: the same inputs always give the same output, and that is why people trust them. A model generates likely text, which is close to the opposite property. Good tools keep the model on the writing side of the line (producing formulas and code that run deterministically) and treat model output placed directly in cells with caution.

---

Excel and Copilot

Microsoft's own help page describes Copilot in Excel as able to edit worksheets (adding and renaming sheets, changing ranges, applying formatting and data validation), generate formulas and calculations, summarize data, create and edit charts and PivotTables, sort and filter, search the web with citations, import data from other workbooks, and run custom skills for repeatable tasks. It has three modes on Windows, Mac, and the web: edit (the default), plan (which proposes steps for you to confirm before changes), and chat only, which does not change your workbook (Microsoft Support). The page notes that Copilot is not included with every Microsoft 365 subscription and that availability depends on your organization's settings.

Agent Mode

Agent Mode, which plans and carries out multi-step work in a workbook, became generally available on Excel for the web in December 2025 and on Excel for Windows and Mac on January 27, 2026. A guide from Glide, which sells a competing app builder, reports that it requires a Microsoft 365 Copilot license, a Microsoft 365 Premium subscription, or a Personal or Family plan that uses AI credits, that it is not available in the EU or UK, and that users can choose OpenAI or Anthropic models or an automatic mode. The same guide notes two limitations that matter in practice: Agent Mode changes the workbook in real time with no preview step, and it works only within the open file (Glide). Treat those as a competitor's description and confirm against Microsoft's documentation, but the practical advice holds regardless: work on a copy, not the only version of an important file.

The COPILOT function is gone

Microsoft previewed a =COPILOT() worksheet function that sent a prompt and cell references to a model and returned the answer in a cell. Microsoft's support page now states that, starting September 14, 2026, the function is no longer available. Results already calculated stay in the file as cached values, but when a cell containing the function recalculates, it returns #NAME?. Microsoft says Copilot in Excel remains available through the Copilot pane. The function had been a preview feature available only through the Frontier and Microsoft 365 Insider programs (Microsoft Support).

This is a useful case study. A workbook that relied on a preview-stage AI function broke when the feature was withdrawn, and any tutorial recommending it became outdated. When you evaluate an AI feature for work that must keep running, ask whether it is generally available or a preview, and keep a way to reproduce the result without it.

Pricing. Microsoft's pricing page did not load when we checked, so we do not quote a figure. Confirm the current price with Microsoft or your reseller.

---

Google Sheets and Gemini

Google's page on spreadsheet AI lists what Gemini in Sheets does: create tables from a prompt, generate formulas, build pivot tables, apply find-and-replace and conditional formatting, complete patterns with Smart Fill (enhanced Smart Fill can populate text fields), identify trends and outliers, clean data, create charts, and answer questions about your data. It also describes an AI function you can use from any cell to generate, summarize, or categorize text across many rows. The page says premium AI, including Gemini in Sheets, is now part of Google Workspace plans, without listing which tiers include which features (Google Workspace).

The AI function is the cell-level pattern described above, so the same caution applies: outputs can vary, and recalculating regenerates them. For repeatable work, ask Gemini to write the formula and paste the formula in, rather than leaving model calls in a large range.

Scale. Google's documentation says a spreadsheet can hold about 20 million cells or 100 MB (Google Docs Editors Help). Excel allows up to 1,048,576 rows and 16,384 columns per worksheet and 32,767 characters in a cell (Microsoft Support). If your file approaches those limits, a spreadsheet may be the wrong container, and a database or a dataframe is a better one. Our guides to SQL for data science and pandas cover the next step up.

---

Rows

Rows is a web-native spreadsheet built around connecting to outside services and analyzing the result with AI. Its pricing page, read in October 2026, lists:

Plan	Price	AI tasks per month	Integration accounts	File import limit
Free	$0	5	10	1 MB
Plus	$8 per user monthly ($6 annual)	200	10	5 MB
Pro	$79 per month plus $8 per user ($59 plus $6 annual)	1,000	100	100 MB
Enterprise	Custom	Unlimited	Unlimited	100 MB

Each message to the AI Analyst uses one AI task, as does each file processed by AI Vision, such as a PDF or image. Per-cell enrichment, such as AI sentiment scoring in a formula column, counts against a separate monthly limit (500 on Free and 1 million on Plus and Pro). Data table refresh is manual on Free, daily on Plus, and every minute on Pro (Rows pricing).

Strengths. A clean web interface, native integrations, and live data tables that refresh on a schedule.

Weaknesses. Complex Excel workbooks, with macros and elaborate formulas, may not carry over cleanly, and the Free plan's five AI tasks run out quickly.

Bricks

Bricks describes itself as an AI data analyst that turns data into dashboards and reports, with an AI-powered spreadsheet for preparation. You describe tasks in plain language, such as calculations, lookups across tables, cleaning, and formatting, and describe charts you want. It exports PDF reports, PowerPoint presentations, or shareable dashboard links, and supports real-time team editing. The page mentions importing CSV and Excel files, and says Bricks is free to use for teams of any size with paid plans for more features, usage, and support (Bricks). We could not read a price list, so check the plans page.

Fit. Teams that want charts and summaries from a file without building a pivot table. Limit. An analyst tool decides what to compute. If you do not already know what a correct answer looks like, you cannot tell whether it picked the right measure.

Numerous.ai

Numerous adds an =AI function to Google Sheets and Excel. Its pricing page lists a Personal plan at $10 a month billed annually, with allowances for characters of input and output and for formula generations, and a seven-day free trial. The page is sparse on the function's syntax and models, and on which plan supports Excel (Numerous pricing).

Fit. Batch text work across hundreds of rows, such as classifying support tickets or extracting a field from messy notes. Caution. This is the cell-function pattern again: results can vary, and each recalculation can use allowance. A common safeguard is to run the function once, paste the results as values, and keep the prompt in a note.

Quadratic

Quadratic is a spreadsheet where cells can run Python, SQL, JavaScript, or ordinary formulas, with an AI that writes the code and explains it. Its site lists connections to databases such as Postgres, MySQL, Snowflake, and BigQuery, support for the Model Context Protocol so other AI agents can read and write sheets, and sheets of up to 100 million rows. It names three AI agents (an auditor, a data augmentor, and an error fixer) and says you can open any cell to check the AI's work (Quadratic).

Its pricing page lists a free Personal plan with limited AI usage, Pro at $18 per user per month with $20 in monthly AI credits, Business at $36 with $40 in credits, and a custom Enterprise plan with a self-hosting option. Prices shown are billed annually (Quadratic pricing).

For people who already use Python, the appeal is that the AI produces code you can read and rerun, which is closer to the assistant pattern than the cell-function pattern. See our comparison of DuckDB and pandas for related tools.

---

Side-by-side comparison

Tool	Pattern	Output you can inspect	Main risk
Excel Copilot (edit and chat)	Assistant	Formulas and edits in the sheet	Edits applied directly in edit mode
Excel Agent Mode	Analyst	Steps shown as it works	No preview step; works on the open file
Gemini in Sheets (assistant)	Assistant	Formulas and tables	Plausible but wrong formulas
Gemini AI function	Cell function	Text only	Output can change on recalculation
Rows AI Analyst	Analyst	Tables, charts, messages	AI task limits on lower plans
Bricks	Analyst	Charts and reports	Chosen measure may not match your intent
Numerous.ai	Cell function	Text only	Variable output, allowance use
Quadratic	Analyst with code	Code in cells	Code you do not read still runs

---

What spreadsheet errors cost, and what AI changes

Spreadsheet errors are old, well documented, and easy to underestimate. Two cases show how small mistakes travel.

Reinhart and Rogoff. In 2013, Thomas Herndon, Michael Ash, and Robert Pollin examined an influential economics paper on public debt and growth. They found, among other problems, that the authors' Excel spreadsheet left out the last five rows, which covered five countries. Correcting that alone changed the average growth rate for high-debt years from -0.1 percent to 0.2 percent, and with their other criticisms applied the figure became 2.2 percent (New Statesman). Reinhart and Rogoff acknowledged the coding error while disputing the critics' broader conclusions. The mistake was a formula range that stopped too early, exactly the kind of error an assistant could introduce as easily as a human.

Gene names turned into dates. A 2016 study screened 35,175 supplementary Excel files from 18 journals and found gene-name errors in 987 files from 704 articles, and more than 20 percent of the supplementary files in some journals had errors. Excel with default settings converted names such as SEPT2 into "2-Sep" and MARCH1 into "1-Mar." The authors reported that the problem grew by more than 15 percent a year over five years (University of Reading Open Research blog, summarizing Ziemann and colleagues in Genome Biology). No AI was involved; the spreadsheet guessed at data types and changed values silently.

What AI changes is speed and volume. An assistant makes it easy to produce a dozen formulas in a minute, and each formula is a small bet that the range, the criteria, and the data types are right. The error types stay the same (wrong range, wrong criteria, silent type conversion). Only the rate at which unchecked formulas appear goes up.

---

How to check an AI-written formula

Suppose you ask an assistant: "Calculate the average order value for customers acquired through Google Ads." It might return:

=AVERAGEIFS(D2:D1001, B2:B1001, "google_ads")

In AVERAGEIFS, the first argument is the range to average (here, order values in column D), and the following pairs are a criteria range and a criterion (column B equals "google_ads"). The formula is plausible. Whether it is correct depends on facts the model may not have: whether the data really spans rows 2 to 1001, whether the source column says "google_ads" or "Google Ads", and whether blank or refunded orders belong in the average.

A short routine catches most problems:

1. Read the formula against the headers. Confirm each range matches the column you mean and covers every data row.
2. Test on a small known sample. Filter five matching rows by hand, compute the average with a calculator, and compare.
3. Check the edges. Test a row at the top, a row at the bottom, a blank, and an entry with different capitalization or trailing spaces.
4. Compare two methods. Calculate the same figure with a PivotTable or a different formula; the two should agree.
5. Look for silent conversions. Dates, leading zeros, long ID numbers, and gene-like names are the usual casualties when text is read as numbers or dates.
6. Record the prompt and the date, especially for numbers that go into a report.

Our overview of what data cleaning involves explains why the cleaning step eats so much analysis time, and why spreadsheet automations break at scale covers the same failure modes when formulas run on a schedule.

---

Where AI helps and where it does not

Helps. Writing formulas you could write but would have to look up, explaining an inherited sheet ("what does this nested IF do?"), cleaning text columns, standardizing categories, drafting charts, and producing a first-pass summary of a table. These are tasks where you can see the output and test it.

Helps with supervision. Large-scale classification with a cell function, where you can sample 50 rows to estimate accuracy. If the sample shows 90 percent agreement with your own labels, you know roughly what you are accepting.

Does not help. Models with many interdependent assumptions (a financial forecast), where a wrong assumption is invisible in the output. Regulated work where auditors need to see exactly how a number was produced. Files near the size limits above. And any task where you cannot say what the correct answer looks like. The model will return something confident either way.

For analytics that start from a spreadsheet but outgrow it, see our overview of AI data analysis tools, which covers tools that go beyond the grid, and how data visualization can mislead for the chart side of the problem.

---

Privacy and governance

Spreadsheets hold payroll, customer lists, and financial plans. Before pointing an AI feature at them, answer four questions from the vendor's documentation, not from marketing:

• Does the data leave your tenant or workspace, and where is it processed?
• Is it used to train models, and can that be turned off for your plan?
• How long are prompts and results retained?
• Can an administrator restrict the feature by user group or sensitivity label?

Cell-function add-ons send the text of cells to a model provider, so each prompt can carry the data in the referenced cells. Agent-style features that edit a workbook directly deserve a rule of their own: work on a copy and keep version history on. Our article on how local AI tools protect privacy explains the trade-offs of keeping data on your own machine.

---

What to pick

If you are	Start with	Then consider
A Microsoft 365 organization	Copilot in Excel (chat mode first)	Agent Mode on copies of files
A Google Workspace organization	Gemini in Sheets	Numerous.ai for batch text tasks
A startup pulling data from many SaaS tools	Rows	Quadratic if you need SQL and Python
A non-technical team that wants dashboards	Bricks	Keep one person who can verify the numbers
An analyst who works in Python	Quadratic	A dataframe workflow outside the grid
A solo user on a budget	Free tiers of Rows or Quadratic, or a general chatbot	See our ChatGPT alternatives guide

Do not pay for several overlapping tools. Pick one, learn its limits, and keep a plain-formula fallback for anything that has to keep working.

---

Frequently asked questions

Q: Did Microsoft remove the COPILOT function from Excel? A: Yes. Microsoft's support page says that starting September 14, 2026, the function is no longer available. Previously calculated results remain as cached values, but cells that recalculate return #NAME?. Microsoft says Copilot in Excel remains available through the Copilot pane.

Q: Will AI replace spreadsheet skills? A: No. AI shortens the path from a question to a formula, but you still need to know what a correct result looks like and how to test one. Knowing the structure of your data is what lets you catch a wrong range or a silent type change.

Q: How accurate are AI-generated formulas? A: Often correct for common patterns, and sometimes wrong in ways that look reasonable. Test on a small sample you can compute by hand, check edge cases, and compare with a second method before a number goes into a report.

Q: Is the Gemini AI function safe to use across thousands of rows? A: It works for text classification and extraction, but outputs can vary between runs and regenerate on recalculation. Sample a few dozen rows to estimate accuracy, then paste results as values if you need them stable.

Q: Is Excel Agent Mode safe on important files? A: One third-party guide reports that it edits the workbook in real time with no preview step. Use it on a copy, keep version history on, and review the changes it reports.

Q: Is it worth switching from Excel to Rows or Quadratic? A: Only if your work depends on live data from outside services or on SQL and Python inside the sheet. For ordinary spreadsheet work, the built-in assistant in the tool you already use costs less to adopt.

---

The takeaway

The useful question about an AI spreadsheet tool is not how clever the model is but what it hands back. A formula, a piece of code, or a chart built from visible steps can be read, tested, and rerun. A string of text dropped into a cell by a model cannot be checked the same way, and it can change the next time the sheet recalculates. The history of spreadsheet errors, from a formula range that ended five rows early to gene names read as dates, suggests that the dangerous mistakes are quiet ones. Tools that produce inspectable output and habits that test it against a known answer keep the speed of AI without giving up the property that made spreadsheets trustworthy in the first place.

---

Read next:

• Best AI Data Analysis Tools 2026
• Best AI Tools for Everyone 2026
• Google Sheets Automation Guide
• Best ChatGPT Alternatives 2026
• What Data Cleaning Actually Involves
• Why Spreadsheet Automations Break at Scale

Last verified: October 22, 2026. Prices, plan limits, and feature availability change often; confirm them on each vendor's site.