Is It Safe to Let AI Read Your Email? A Transparent Answer

For Gmail, Sifta requests: gmail.readonly (to read emails), gmail.modify (to archive emails when you instruct it to), gmail.compose (to draft replies you request), and userinfo.email (to identify your account). For Outlook: Mail.ReadWrite (read, archive, and draft), User.Read (identify your account), offline_access (maintain the connection), openid, and email. Sifta never requests permission to send email autonomously — every send action requires your explicit approval.

Sifta uses two providers. Email classification and the 2-3 sentence summaries delivered to you via iMessage are generated by OpenAI's GPT-4.1-mini. A lightweight noise-detection pass uses OpenAI's GPT-4.1-nano to filter clear spam and newsletters before the main classifier runs. When you send a message back to Sifta in iMessage — asking it to search your inbox, explain a thread, or take an action — those conversational replies are generated by Anthropic's Claude Haiku model. Both providers' API usage terms prohibit using customer API inputs to train their general models.

Yes. Sifta stores email metadata and body text in a Supabase PostgreSQL database to enable features like inbox search, agent context, and multi-turn conversations. All database tables have row-level security (RLS) enabled — your data is only accessible to your authenticated account and cannot be queried by any other user. Your Google and Microsoft OAuth tokens are stored in encrypted form.

No. OpenAI's API data usage policy prohibits using inputs submitted via the API to train OpenAI models without explicit opt-in. Anthropic's policy likewise prohibits using API inputs for training without consent. Sifta has not opted in to either program. Your emails are not used to train any AI model.

The gmail.compose scope grants Sifta technical permission to create and send emails from your Gmail account. The Mail.ReadWrite scope grants the equivalent for Outlook. Sifta's application code uses these scopes only when you explicitly instruct it to draft or send a reply — the AI will not initiate any send action on its own. If you are concerned, you can review your authorised scopes at any time in your Google Account or Microsoft account settings and revoke access entirely.

For Gmail: visit myaccount.google.com, click Security, then 'Third-party apps with account access', find Sifta, and click Remove Access. For Outlook: visit account.microsoft.com, click Privacy, then 'Apps and services that can access your data', find Sifta, and remove it. Access is revoked immediately — Sifta stops monitoring your inbox the moment the OAuth permission is removed.

When you cancel, Sifta stops monitoring your inbox. Your email data (body text, metadata, AI summaries) remains in Sifta's database with row-level security isolation intact until you request deletion. To permanently delete your account and all associated data, email support@joinsifta.com. Revoking OAuth access through Google or Microsoft settings stops inbox monitoring independently of — and immediately, without waiting for — account deletion.

Your email data is processed by automated systems (the AI classifier and conversational agent). Human Sifta employees do not access your email content during normal operations — row-level security at the database layer enforces this technically, not just by policy. If you report a specific support issue and voluntarily share an email excerpt as part of that report, the support team member handling your case may see what you have explicitly provided.