Connecting an AI tool to your inbox means handing it some of the most sensitive data you have: client conversations, invoices, password resets, things people told you in confidence. You should know exactly what happens to that mail before you click “Connect Gmail”. This post walks through it plainly. The full detail, kept up to date, lives on our Trust & security page.
Your email is not used to train AI
We don't train models on your email, and we don't sell it or share it for advertising. To categorise, prioritise and draft, an email is sent to Anthropic's commercial Claude API. Under Anthropic's commercial terms, API inputs and outputs are not used to train their models.
What the AI receives is the email being processed: subject, sender and the body up to about 4,000 characters. When it writes a draft it may also get matching snippets from your knowledge base and, if you connected a calendar, a few free time slots. It doesn't get your whole mailbox.
If you add your own Anthropic API key, those calls go to Anthropic under your account instead of ours.
What we can access, and when we use it
With Gmail we ask for read access (to triage incoming mail, and read Sent mail for follow-up reminders and learning your writing style) and modify access (to apply “Email Triage/…” labels, archive, and send replies). Modify access can't permanently delete mail. Calendar is optional and limited to free/busy: no event titles, attendees or descriptions.
Having permission isn't the same as using it. Nothing changes in your mailbox until you turn on mailbox sync, and nothing is sent without your click unless you switch on Autopilot for a category. You can revoke access at any time from Settings or from your Google or Microsoft account permissions page. The exact permission list for both providers is on the Trust page.
What we store, and for how long
- Sender, subject, date and the first 10,000 characters of each email body, so the dashboard, drafts and corrections work.
- The AI's results: category, priority, reasons, summaries and drafts.
- Your settings, rules, and any knowledge-base text or pages you add.
- No attachments, and no calendar data beyond the times offered in a draft.
Data is kept while your account is active. Disconnect a mailbox or delete your account and the related data is permanently removed within 30 days.
How it's protected
OAuth tokens, and your own API key if you add one, are encrypted at rest with AES-256-GCM and only decrypted in memory when they're used. All traffic is over HTTPS, and we never see your Google or Microsoft password. Our servers and database are hosted by Contabo in the EU.
To be precise about the limits: that encryption covers credentials. Stored email text and AI results are kept in our database behind access controls; they aren't individually encrypted the way credentials are.
Who else is involved
Besides Anthropic, a short list of subprocessors handle specific jobs: Google and Microsoft (mailbox access), Stripe (billing; we never see card numbers), Brevo (account emails and the daily digest), Cloudflare (delivery, DDoS protection, cookie-less analytics) and Contabo (hosting). The subprocessor table lists exactly what data each one sees.
You can read every prompt
Every prompt we send to the AI is published, word for word, on our transparency page. If you want to know what the AI is told to do with your email, you can read the actual instructions rather than a summary of them.
What we don't claim
We don't hold SOC 2 or ISO 27001 certification, and we'd rather say so than imply otherwise. Email Triage AI is run by TelaLoom, a small Australian business. If your organisation needs a security questionnaire answered, send it through the contact form and a person will reply.