Is it safe to give an AI assistant access to your email? A 9-point checklist

AI email assistants need to read your messages to be useful. That's a lot of trust. Before you click “Allow”, these are the questions worth asking, and the answers you should expect.

The 9-point checklist

1. Does it connect through the official permission screen?

You should see Google's or Microsoft's own consent screen listing exactly what the app can do. Never type your email password into a third-party site.

2. Which permissions does it ask for?

Reading and sending is reasonable for an assistant that drafts replies. Deleting mail, changing settings or managing your contacts usually isn't needed. Fewer permissions is better.

3. Has it been verified by Google?

Apps that read Gmail use a Google “restricted scope” and must pass Google's OAuth verification and an annual independent security assessment (CASA) before they can serve the public. Until then, Google shows a warning screen and caps the number of users. Ask where the app is in that process.

4. Can the AI send anything on its own?

This is the most important question. An assistant that can send without your approval can be tricked: an incoming email can contain hidden instructions (“prompt injection”) that try to make the AI forward data or reply to someone. The safe design is that the AI only drafts, and a human presses send, every time.

5. Is your data used to train AI models?

Look for a clear “no” in the privacy policy, both for the app and for the AI provider it uses. Major AI providers' commercial APIs generally don't train on customer data by default, but check the specific provider.

6. How are your access tokens stored?

The tokens that let the app read your mail should be encrypted at rest and never exposed to the browser. A good security page will say how.

7. Who can see your messages?

Ask whether employees can read your mail, and under what conditions. “Only with your permission, for support” is the answer you want.

8. Can you delete everything?

You should be able to disconnect any account, export your data and delete your workspace yourself, without emailing support. Also check that you can revoke access from the provider's side (myaccount.google.com/permissions for Google).

9. Are the security claims real?

Be wary of badges like “SOC 2” or “bank-grade encryption” without a report or details behind them. A trustworthy company also tells you what it hasn't done yet.

Red flags

• Asks for your password, or logs in to LinkedIn or WhatsApp “as you” through a browser.
• Offers an “auto-reply with AI” switch that sends without review.
• No named AI provider or sub-processor list.
• No way to delete your data yourself.

How Senvoro answers these

Senvoro connects through Google, Microsoft and Slack's own permission screens with read and send permissions only, encrypts tokens with AES-256-GCM, and processes text with Anthropic's Claude, which doesn't train on data sent through its commercial API. The AI can only draft: every message needs you to press “Approve and send”, and there is no auto-send setting. Google's Gmail verification isn't complete yet, and we say so. The full details, including what isn't done yet, are on the security page.