Spreadsheets are where most business logic quietly lives, and where AI assistants either shine or quietly corrupt a formula. On October 5 to 6, 2026, Anthropic put Claude inside Google Docs, Sheets and Slides as a public beta, available on all paid Claude plans. It sits in a sidebar, reads the file you have open, edits it in place and, by default, asks you to approve each change.
This guide covers what it does, how to set it up, where it fits against other options, and a short checklist for using it on files that matter. We worked from the vendor listing and press coverage, not a long hands-on test, so treat quality claims as things to verify on your own files.
Anthropic announced the launch from the official @claudeai account, in a post that passed 5.5 million views: "Claude now works inside Google Docs, Sheets, and Slides, and those files also open inside Claude. In Google Workspace, Claude sits in a sidebar next to your file, reads what you have open, and edits it in place. You can approve each edit before it lands." The announcement video is below; the original post is on X.
Note the second half of that announcement: those Google files also open inside Claude, so the integration runs in both directions, from Workspace into Claude and from Claude into your Google files.
TL;DR: the questions people are asking
| Question | Short answer |
|---|---|
| What is it? | A Claude sidebar inside Google Docs, Sheets and Slides. |
| Who gets it? | All paid Claude plans, in public beta. |
| How do I install it? | Google Workspace Marketplace, then Extensions, Claude, Open Claude. |
| Does it edit in place? | Yes, with approval per edit available. |
| What does it do in Sheets? | Formulas, pivot tables, native charts, new tabs, Python-backed cleanup. |
| Biggest risk? | A confident but wrong formula or a silently changed range. |
| How to undo? | Version history in Google Workspace. |
What it does in each app
Google Sheets is the headline. Reported capabilities include:
- Writing formulas, including multi-step ones you would otherwise assemble from documentation.
- Building pivot tables, the quick summaries of raw rows that many people avoid.
- Creating native Sheets charts rather than pasted images, so they update with data.
- Adding new tabs for summaries or scratch work.
- Using Python for tasks that formulas do badly, such as merging tables, fuzzy matching or fixing malformed data, then writing the cleaned result back into the sheet.
Google Docs and Slides get in-file editing and drafting help. In Docs that means rewriting, restructuring and summarizing a document in place. In Slides it means helping with layout, wording and structure. Coverage described the Slides side as "smarter," with details less specific than for Sheets, so test it on a real deck before relying on it. If your main need is generating decks from scratch, compare it with dedicated tools, as in our look at Gamma 5.
How to set it up
- Install from the Google Workspace Marketplace. Search for Claude and add it. Your Workspace admin may need to allow it.
- Sign in with a paid Claude plan when prompted.
- Open a file and go to Extensions, then Claude, then Open Claude. The sidebar appears.
- Keep per-edit approval on while you get used to it.
- Optionally connect the Docs, Sheets and Slides connectors (beta) in the Claude app, which let Claude create and edit Google files from a chat rather than from inside the file.
The connector route builds on the earlier Claude Gmail and Drive connectors. The sidebar is the more direct, in-context version.
What it is good for, and where to be careful
A practical way to think about it is by task type.
| Task | Good fit? | Caution |
|---|---|---|
| Writing a lookup or conditional formula | Yes | Check edge cases and blank cells |
| Building a pivot table from clean data | Yes | Confirm the source range |
| Making a chart | Yes | Check axes, units and labels |
| Cleaning messy text columns | Yes, via Python | Spot-check a sample of rows |
| Financial models with chained assumptions | Use with care | Wrong formulas can look plausible |
| Anything with personal data | Check policy first | Data handling and sharing settings |
| Rewriting a Doc | Yes | Keep the original in version history |
The common failure of any assistant in a spreadsheet is a formula that returns a number that looks right and is not. A mismatched range, an off-by-one row or an assumption about blank cells can pass a glance and fail an audit.
A review checklist for AI-edited spreadsheets
- Read every formula it writes at least once, and ask it to explain any you do not understand.
- Test with known inputs. Put in a few rows where you know the answer and compare.
- Check ranges. Look for ranges that stop one row short or include headers.
- Inspect totals against an independent source, such as a report or a manual sum.
- Look at edge cases, such as blanks, text in numeric columns, duplicates and dates in different formats.
- Use a copy of the sheet for large restructuring, and keep the original untouched.
- Review the diff. Version history lets you compare before and after.
- Keep approvals on for shared files so a colleague's data is not changed silently.
How it compares with the alternatives
Claude is arriving in a crowded space.
- Gemini in Workspace is built into the suite and has the deepest hooks, which makes it the default for many organizations.
- ChatGPT added its own Docs, Sheets and Slides integration in August, which we covered in ChatGPT's Google Docs, Sheets and Slides integration.
- Cursor shipped Workspace plugins for a different audience, described in our Cursor Google Workspace plugins post.
- Claude Cowork merges chat, docs, slides and design in Anthropic's own environment, as in our Cowork coverage.
The honest way to choose is a bake-off on your own spreadsheets. Pick three tasks you actually do, such as a monthly summary, a messy import and a chart, run them on each tool, and compare accuracy and the time you spend fixing output. Model choice, price and admin controls then break ties.
Admin and security considerations
If you manage a Workspace, treat this like any new add-on. Review the permissions it requests, decide who can install it, and check data-handling terms for files with regulated or customer data. Test with non-sensitive files first. Ask whether logs of AI edits are available for audit, and whether you can restrict it to certain organizational units. For agents more broadly, our opinion piece on agent teams explains why permission design is becoming the real bottleneck.
Prompts worth trying in Sheets
A few concrete requests show how to get value without handing over judgment.
- "Add a column that flags rows where the invoice date is more than 30 days before today and the status is not paid, and explain the formula."
- "Build a pivot table of revenue by region and month from A1 to F500, then a line chart of the monthly totals."
- "This column mixes dates in three formats. Standardize to ISO dates in a new column, and list any rows you could not parse."
- "Find duplicate customers by email, ignoring case and spaces, and show me the groups before changing anything."
- "Check my totals tab against the raw data and tell me where they disagree."
Notice the pattern: ask for an explanation, ask it to show a sample before changing data, and ask it to list what it could not handle. Those habits catch most errors early. For Docs, ask for a summary of changes after a rewrite. For Slides, ask for speaker notes, which expose weak logic faster than layout does.
Rollout advice for teams
Start with a pilot group that works in spreadsheets daily, such as finance, operations or marketing analytics. Ask them to log time saved and errors found over two weeks. Decide upfront which files are off limits, such as payroll or anything with regulated personal data, until data-handling terms are reviewed. Pair the rollout with a short guide on verifying AI-written formulas, since the habit matters more than the tool. Revisit the decision after the beta ends, when pricing and terms may change.
Common failure modes to expect
Expect a few repeat offenders. Assistants sometimes reference the wrong sheet when a workbook has several similar tabs, so name the tab in your request. They can misread merged cells and header rows, which breaks pivots, so unmerge and clean headers first. Locale differences matter too: a formula using commas as separators fails in a workbook set to a locale that expects semicolons. Volatile functions that recalculate constantly can slow a large sheet, so ask for non-volatile alternatives when performance matters. Finally, very large ranges may be truncated when passed to Python for cleanup, so verify row counts before and after any bulk operation.
What this means for what you build or pay
If you already pay for a Claude plan, the sidebar costs nothing extra to try, which lowers the bar for a trial. If your team lives in Sheets, the likely payoff is the unglamorous work: formulas, pivots and cleanup that eat hours. The risk is the same as with any assistant, trusting output you did not verify. Treat it as a fast junior analyst whose work you review, not as an auditor.
Related reading
- Claude Gmail and Drive connectors
- ChatGPT's Google Docs, Sheets and Slides integration
- Claude Cowork: chat, docs, slides and design
- Cursor's Google Workspace plugins
- Gamma 5: agent-driven presentations
- Agent teams and the permission bottleneck
Primary: the @claudeai announcement post and video (October 6, 2026) · Anthropic's Claude for Google Workspace listing and help article (updated October 5, 2026) · 9to5Google, Digital Trends and other launch coverage
Details are accurate as of October 7, 2026 and come from the vendor listing and press coverage, not a long hands-on test. Features, plans and data-handling terms may change during the beta; verify before enabling for sensitive files.
