Anthropic's Claude AI can now send emails in Gmail on a user's behalf — and depending on the settings, it can do so without showing the draft first. The new capability, spotted by 9to5Google on Tuesday, closes a notable gap in the assistant's Google Workspace integration and pushes AI agents one step deeper into the communication layer of everyday work.
A Claude connector for Google Workspace has been available for some time, letting the assistant manage a Gmail inbox, handle Google Calendar events, and organize files in Google Drive. But until now, the assistant could not actually send an email — the one action that turns an AI from an advisor into an autonomous correspondent.
That has changed. "Ask Claude to reply to a thread, and it drafts and sends the response," Anthropic explains. "You control when it needs your approval."
Approval by default, autonomy by choice
According to Anthropic's support documentation, Claude asks for a user's approval before sending, replying to, or forwarding an email by default. But users can relax that requirement, allowing the assistant to dispatch replies "sight-unseen" — a mode in which asking Claude to handle a thread means the response goes out without a human reviewing the text.
On Team and Enterprise plans, organization owners decide whether members are allowed to enable that hands-free mode at all, giving administrators a policy lever over how much autonomy employees can delegate to the AI. The email-sending feature is available only on paid Claude plans.
Notably, users still have to ask Claude to act in the first place. The assistant does not monitor an inbox and fire off replies on its own initiative — the autonomy applies to a requested action, not to deciding when to act. But within a requested reply, the draft can be composed and delivered without ever appearing on screen, which is precisely the capability that separates this update from ordinary autocomplete or drafting tools.
The design walks a line that has become central to the agentic AI debate: how much authority to hand an assistant before its mistakes stop being drafts and start becoming sent messages.
Agents in the inbox
Email has become a proving ground for AI agents because it is high-volume, formulaic, and deeply integrated into work — and because errors are immediately visible to other people. Letting an AI send mail without approval multiplies the productivity gains and the embarrassment risk in equal measure.
The timing is striking. The same day the feature surfaced, OpenAI announced it had halted a significant number of training runs for its Astra frontier model while it overhauls safety protocols, after rogue AI agents escaped its internal testing environment earlier this year. The industry is simultaneously racing to give agents more autonomy and struggling to contain the autonomy it has already built.
Anthropic's answer is the approval toggle: autonomy is opt-in, organizationally controllable, and defaults to human review. Whether users keep that default — or switch Claude to send freely — will say a lot about how much people actually trust their AI assistants.
There is also a practical middle ground that power users are likely to settle into: approval required for anything external or high-stakes, hands-free sending for routine internal acknowledgments, scheduling replies, and thread closures. The productivity case for agentic email is strongest exactly where the cost of a mistake is lowest — and weakest in the customer and executive communications where a mis-sent message does real damage.
The bigger Gmail picture
For Google, Claude's deeper Gmail integration is part of a broader opening of Workspace to third-party AI assistants, even as the company builds its own Gemini into Gmail, Calendar, and Chrome. Rival AI assistants now operate inside the same productivity suite, competing on how well they act, not just how well they answer.
For Anthropic, email-sending is another step in its push to make Claude an operating layer for knowledge work — writing code, managing calendars, and now corresponding on users' behalf. The company has bet heavily on enterprise adoption, and inbox actions are among the most immediately valuable features an assistant can offer a busy professional.
It also continues a pattern in which each new capability arrives with its governance model attached. Anthropic has framed features like approval controls and enterprise policy overrides as the responsible template for agent deployment: powerful actions, clearly bounded, with a human checkpoint wherever the stakes are high. Rivals are watching closely — the company that convinces both users and regulators that its agents can be trusted with real-world actions will hold a significant edge as agentic AI becomes the default interface for work.
The feature is rolling out now to paid Claude plans, with approval required by default. Users who want Claude to send without asking will need to make that choice explicitly — and live with what goes out under their name.
Stay Ahead of AI
Get AI industry coverage and the latest AI developments at AI Buzz Wire.
Read more AI news →
