Superhuman's AI auto-draft lets you reply without typing

Most AI email drafts sound like an overly excited real estate agent or a rigid corporate bot. You spend more time deleting the draft and rewriting it yourself than you would have spent starting from scratch.

Email client Superhuman is rolling out an overhaul of its auto-draft feature, and it might actually change how we handle inboxes. Instead of churning out generic templates, the app now analyzes incoming emails, identifies which ones need a reply, and writes responses that mirror your personal tone based on past conversations.

Frontier models and three ways to reply

Under the hood, Superhuman completely ditched older systems like GPT-3.5. Co-founder Rahul Vohra said the new feature runs on a mixture of models, leaning heavily on the latest frontier models from both Anthropic and OpenAI to maximize context and intelligence.

Whenever an important email arrives, the client drafts a primary response alongside two alternative variations. In early beta tests, it handled tasks like agreeing to pitch embargoes, confirming meetings, and declining out-of-scope requests. Vohra revealed that during testing, 40% of generated drafts were sent within a day, and 60% of those were sent without any manual editing at all.

Users can also feed the system more context by heading to the settings page to add details about their job, upload reference files, or drop in links. The app also learns from your corrections over time. When a tester rejected a draft that accidentally agreed to an absurd midnight meeting, the system learned its lesson and automatically declined the next late-night invitation.

Why it matters

We think Superhuman's smartest design choice here is providing multiple variations instead of assuming AI knows best. By keeping a human in the loop to click and approve, it avoids the creepy uncanny valley of fully autonomous email while still shaving real minutes off routine replies.

If your daily work involves triaging dozens of logistics and scheduling emails, this is genuinely worth testing. It will not write your sensitive strategy memos, but it shines at clearing out repetitive administrative chatter where you just need to say yes, no, or confirm a time.

That said, the trap is over-trusting the defaults. The system still leans toward being overly agreeable, occasionally agreeing to crazy meeting hours or pitches by default. You still have to keep an eye on it. There is also a broader irony at play: AI email pitches created the modern inbox flood in the first place, and now we are paying premium software to let another AI reply to them. Superhuman, backed by parent company Grammarly and working on a broader Superhuman Go assistant, is betting that the winner of the AI workplace race is whoever captures this context layer first.