This column started in June with ChatGPT arriving in Google Sheets. Nineteen weeks later the other large AI company has done the same thing, with more care about permissions, and the result is worth a careful look from anyone whose business lives in a Google spreadsheet.

Claude for Google Workspace is a single add-on from the Google Workspace Marketplace. Install it once and a Claude sidebar appears in Docs, Sheets and Slides. In Anthropic’s own description, the existing Claude connectors let you search your Drive from the Claude app, and “the extension works the other way around: Claude comes to the file you’re already in.” It is in beta for the Pro, Max, Team and Enterprise plans, which is to say every paid plan, and it works in Chrome, Edge and Safari.

Sheets: write formulas, build models, clean data. Docs: draft, rewrite, redline. Slides: build
decks, restyle slides, use speaker notes. It reads the file you have open, and only that file. (Branded graphic by PGH Consulting, LLC)

What it does in Sheets is the part a finance person will care about. Anthropic’s examples are the jobs that eat a Friday afternoon: build a three-statement model on a new tab from an assumptions sheet, explain how the total in a cell is calculated and flag anything hard-coded, fix a lookup that returns an error, turn a column of dates in three formats into one, add a sensitivity table and a chart. In Docs it drafts, rewrites and cuts long documents and, per the Marketplace listing, can redline a contract with change tracking. In Slides it builds a short deck from speaker notes and restyles existing ones.

Ask before edits, the recommended default: shows changes in plain language, waits for you to
click Allow, best for finance workbooks. Accept all edits is faster but gives less control and still pauses for
security-review items. (Branded graphic by PGH Consulting, LLC)

The switch to leave on

Every AI tool this column has covered has come with a question about what it is allowed to touch. This one answers it in the first setting you see.

The default mode is called Ask before edits. Anthropic’s wording: Claude “shows you what it is about to do, in plain language, and waits for you to click ’Allow.’” There is a second mode, Accept all edits, which lets it make ordinary content changes without stopping. Even in that mode it still pauses for anything that touches links, external images, web addresses or someone’s access to the file, and those pauses carry a Security review label with, in Anthropic’s words, “no ’always allow’ option for them.”

Some things it will not do in either mode. It cannot share the file, give anyone access or change who owns it. It cannot create scheduled automations or load arbitrary web pages into your file. Anthropic says these requests are refused even if you ask, which is the right design for a tool that will be given to junior staff.

For a finance workbook, leave Ask before edits on. The cost is a click per change. The benefit is that every formula it writes into your model is a formula you saw before it landed, and that is the difference between a tool that helps you build a model and a tool that rebuilds your model while you were making coffee.

Scope: the open file only, no Drive, Gmail or Calendar access unless separately connected.
Training: not used. Retention: deleted within 30 days. Attribution: edits under your name. Terms:
commercial for Team and Enterprise, consumer for Pro and Max. (Branded graphic by PGH Consulting, LLC)

What it reads, and what happens to it

The data questions this column keeps asking have plain answers here, and they are better than most.

Scope: Claude reads the file you have open and only that file. Unless you have separately turned on a Drive connector, it cannot see other files in your Drive, your email or your calendar, and the add-on asks Google for only two permissions that touch your files: to show the sidebar, and to view and manage the document, sheet or presentation it has been installed in. It does not request Drive, Gmail or Calendar access.

Training: Anthropic states that it does not use this content to train its models.

Retention: inputs and outputs are deleted from Anthropic’s systems within 30 days. One wrinkle for larger firms: the add-on does not inherit a custom retention setting your organisation may have negotiated, so if you have one, this beta does not yet honour it.

Attribution: edits Claude makes are recorded by Google under your name, so they appear in version history like anything you typed yourself, and a colleague sees the edit but not your chat. That is worth telling your team before someone discovers it in a shared file.

Terms: on Team and Enterprise plans the add-on runs under Anthropic’s commercial terms and data processing agreement. On Pro and Max it runs under the consumer terms. If staff will use it on company files, the company plan is the one to use, which is the same conclusion this column reached about every tool it has covered.

Start with the workbook you already distrust: ask Claude to walk you through one key total and
flag every hard-coded value. Install on the company plan, confirm Ask before edits, give it one cleaning
job, tell collaborators edits appear under your name. (Branded graphic by PGH Consulting, LLC)

A first job worth an hour

Do not start with the three-statement model. Start with the workbook you already distrust.

Open the budget or forecast you inherited, the one with the tab nobody opens, and ask the sidebar the question Anthropic itself suggests: walk me through how this total is calculated and flag anything hard-coded. You will get a plain-language chain from the result back to the inputs, and a list of the cells where someone typed a number over a formula in 2024 and never told anyone. That is an audit you would otherwise pay for or skip.

Then one more: take an export with dates in several formats and ask it to normalise them, with Ask before edits on, and watch what it proposes. If it handles that cleanly on your data, it has earned the next job. If it does not, you have learned that in twenty minutes rather than in a board pack.

What the beta will not do yet

Anthropic lists the limits plainly, and the ones that matter here are these.

It works only on the open file. It cannot open, create or copy other files, cannot export a PDF, and the sidebar has to stay open while it works. Google stops any single operation at six minutes, so a very large workbook may need the job broken up.

In Docs it cannot read, reply to or resolve comments, and its edits land as direct edits rather than suggestions, which matters if your review process runs on suggestion mode. In Sheets it cannot read Connected Sheets tabs that pull live data from BigQuery, and it cannot create triggers or scheduled refreshes. In Slides the charts it makes are pasted as pictures, not linked charts, so they do not update when the numbers change.

It does not remove the checking. A survey of 270 chief financial officers and senior finance leaders at large US companies published this week found that finance teams now spend about a quarter of the working week verifying or correcting AI output, and that only one in twenty would send AI-produced figures to a board without a human review. A sidebar that writes the formula in front of you makes that check faster and easier. It does not make it optional.

And it is a beta. Firefox is out, voice is out, and Anthropic says enterprise controls such as zero data retention and third-party hosting are not available in this release. Read the help page, not the headline, before you roll it out to a team.

What to try this week
1. Install the add-on from the Google Workspace Marketplace on the company plan, not a personal one, and confirm Ask before edits is on. Do not change it for finance files.
2. Open the workbook you trust least and ask Claude to walk you through one key total and flag every hard-coded value. Read the list.
3. Give it one cleaning job on an export, dates or supplier names, and score it the way you scored the model bake-off: correct, followed the instruction, how much editing.
4. Tell anyone who shares your files that Claude’s edits appear in version history under the editor’s name, so nobody is surprised.
5. If you are on Team or Enterprise, ask whoever administers Google Workspace whether Marketplace add-ons are allowed, because that setting, not your Claude plan, decides whether staff can install it.

Claude for Google Workspace is in public beta; features, plan eligibility, data terms and limits are as published on Anthropic’s help centre as at 7 October 2026 and may change. The CFO survey figures are from Datarails’ 2026 survey of 270 US CFOs, published 6 October. An AI that edits in front of you still needs you to look.

Peta-Gaye Hardy is the founder of PGH Consulting, LLC, where she helps finance and operations teams adopt AI in practical, low-risk ways. She writes the weekly AI in Finance & Business column and is based between Jamaica and the United States. Learn more at www.pghconsultinggroup.com. Follow on Instagram and YouTube @pghconsultinggroup, and connect on LinkedIn at linkedin.com/in/peta-gaye-hardy.

Disclosures: This article is informational and does not constitute investment, tax, legal, or accounting advice. Readers should consult a qualified professional before acting. Product features, plan eligibility, data-handling terms and limitations are as published by Anthropic at the time of writing and are subject to change; readers should confirm the current position on the vendor’s own pages before installing. Survey figures are as published by Datarails and have not been independently verified. AI tools can produce errors, and every figure, formula, clause, or claim they produce should be verified against a source before being shared or acted upon. The author has no commercial relationship with Anthropic, Google, Datarails or any product mentioned and was not compensated by them. All product names, logos, and trademarks are the property of their respective owners and are referenced for editorial purposes. The examples described are illustrative and do not depict any real business.