Claude moved into Google's house. Mind the default permissions.

On Tuesday, Anthropic put Claude inside Google's most-used enterprise apps. The new Claude for Google Workspace add-on is in public beta on all paid plans, including Enterprise, and it works in both directions: open a Google file and Claude sits in a sidebar that reads the document, sees what you selected, and edits it directly; or start in Claude, paste in a file link or ask it to create one, and work on the resulting Doc, Sheet, or deck from the chat side. Access follows your existing Google sharing permissions. There is no separate price. It is included with paid plans.

This is Anthropic's first-party integration. Third-party Workspace add-ons using Claude models already existed; what changed is that the vendor now ships it, supports it, and extends its enterprise controls to it: the Compliance API, customer-managed encryption keys, and OpenTelemetry audit exports all carry over, per Anthropic. Admins can deploy the add-on to a whole Workspace domain or selected groups.

The capabilities are the ones operators actually want. In Docs, Claude fixes sentences and headings in place, with bigger rewrites arriving as proposal cards you apply or dismiss. In Sheets, the genuinely useful part: formulas, pivot tables, native charts, and, for the messy jobs, pulling ranges into Python, cleaning the data, and writing it back. In Slides, new slides built from the deck's existing layouts, with Claude checking its own output for overlapping objects and unreadable text. Connectors and Skills come along: Anthropic's example is a quarterly business review deck assembled from Salesforce history and Drive call notes via a saved Skill.

Here is the part to read twice. Per Reworked's read of Anthropic's documentation, installing the add-on defaults to full read, write, and delete access to your Google files, with no visible option to narrow that at the start. Claude asks for permission before each action, and the default edit mode is "Ask before edits," which presents changes as approval cards. But there is also an "Accept all edits" mode that lets Claude work a task to completion without stopping. That toggle is the entire governance model for a tool with delete rights on your company's documents, in beta, with Anthropic's own docs carrying the precaution that "in a busy file, check the result." Some features may be limited or unreliable. That is the standard beta caveat, and this time it sits on top of write access to the shared drive.

The competitive picture got strange this week. Gemini has been native in Workspace for over two years. Microsoft's Copilot does multi-step editing inside Word, Excel, and PowerPoint, and VentureBeat notes Microsoft says its agentic Office editing uses Anthropic models. So Anthropic is now competing with Microsoft at the application layer while supplying models to Copilot, competing with Google inside Google's own apps, and already shipping a generally available Claude integration for Word, Excel, and PowerPoint, with Outlook still in beta. The industry has converged on where the next fight happens: inside the documents employees already live in (see our note on Copilot's autopilot billing). The model is no longer the product surface. The file is.

The operator decision: pilot this on "Ask before edits" and keep it there until you have seen a quarter of real usage. Do not flip "Accept all edits" fleet-wide in beta. Before the sidebar reaches finance or legal documents, confirm the trusted-app posture with your Workspace admin, check what the audit export actually captures, and test what happens when Claude and a human edit the same file at once. The per-action approval model is the right default (the same boundary we argued for in the Copilot CLI attack writeup), and the delete grant is the part that makes the default load-bearing. Anthropic built the guardrail. Whether your deployment keeps it on is now your call.