How Automail works
An AI email assistant that drafts in your voice.
What it does
Automail watches your inbox in the background, classifies every incoming message by intent, urgency, and risk, and drafts a reply in your own voice. You review, tweak if needed, and send. Default posture: draft, never auto-send. You're always the last hand on the keyboard.
Voice training
Automail embeds the emails you've already sent — by default the past year, but you can extend to 3 years or all-time — into a vector index. When it drafts a reply, it retrieves the 3–5 most similar past messages (same recipient gets a score bump, recency gently weighted) and uses them as style references in the prompt. The model picks up your idioms, sign-off, formality calibration, and register — casual with friends, formal with clients — automatically. No model fine-tuning, just retrieval.
Decision detection
Some emails ask for input only you can give — "what do you want for dinner", "which slot works", "should we go with A or B". Auto-replying to these is dangerous. Automail's classifier flags them and routes them to a special screen where you pick a suggested answer (or type a short one), and the AI writes a full reply around your choice in your voice.
Safety
- Default to draft. Auto-send is opt-in per low-risk persona, never for high-risk.
- Body-hash check at send. If the body changed since you opened it, send fails.
- 8-second deferred send with undo on every email.
- Audit log on every action — who, when, what changed.
- Private notes the AI shows you live in a separate surface and are never included in the outgoing body. Three-layer separation (column · server action path · UI).
- Swipe-to-dismiss requires a two-step gesture and a 5-second undo window.
What we ask for
Read + send access to your mailbox. We sync the last 90 days of your inbox by default for classification, and a configurable window (default 1 year) of your sent folder for voice training. Spam, promotions, social, and updates are filtered out unless you opt them in.
What we don't do
- We don't use your mail to train any model. Embeddings stay in your database.
- We don't sell or share your data. Single-tenant by default.
- We don't auto-send anything unless you explicitly enable auto-send on a low-risk persona.
- We don't claim HIPAA, SOC 2, or GDPR compliance — and won't unless we're actually certified.
Stack
Next.js + TypeScript on a single DigitalOcean droplet. Postgres + pgvector for storage and retrieval. Redis + BullMQ for background jobs. Better Auth for OAuth. OpenAI for classification, drafting, and embeddings. All open-source / commercially permissive licenses.