01
Three permission layers
Per-app opt-in, per-recipient opt-in, per-message confirmation. Compose As cannot fire a draft if any of the three is missing — and you can revoke any layer instantly.
Cross-app feature
AI that writes in your voice. Only when you ask.
Compose As learns your tone from the things you've actually written, asks before composing, shows the draft, and lets you edit. It never sends. It never reads your bank. It never speaks for you in someone else's voice.
“The first AI writing layer designed to refuse jobs.”
02
Voice-modeled, not voice-cloned
Trained on the messages you've actually written across OneTruth Reach + your linked outboxes. Stays in your voice band even when you're tired and would have written something blander.
03
Money is excluded by design
Compose As is enforced server-side never to draft on financial conversations. Apologies for missing a payment, requests for splits, replies to billers — all human-only.
04
Compose As ≠ send for me
Every draft requires your explicit send. Even reflexive replies pass through your eyes first. The button is always 'Review and send,' never 'Send now.'
05
Receipts on every draft
Every Compose As output writes to your audit log: prompt, model version, draft text, edited text, send timestamp. You can prove what you wrote vs what the model proposed.
06
Voice belongs to you
Your voice profile is encrypted at rest, never used to train shared models, and exportable + deletable from Account → Danger zone whenever you want.
Compose As is the engine. Reach is its first home; Together gets it for shared notes; Work gets it for client briefs. Money never gets it.