Email clients are increasingly using AI to manage overflowing inboxes by categorizing messages and drafting replies that sound like the user. Superhuman has released a new version of its auto-draft feature that detects important emails and generates replies that aim to sound less robotic.
Background and prior attempts
Superhuman previously offered tools such as Instant replies and follow-up auto-drafts. Those earlier attempts often produced messages that came across as an overly enthusiastic AI salesperson, leading to limited use. In the current beta, however, the author reports that several generated drafts required little or no editing before being sent.
How the new auto-draft works
- The feature identifies emails likely to need a response and drafts a reply based on the user’s tone in prior conversations.
- For each identified message it also creates two alternative variations the user might prefer.
- The system learns from user actions and refines its suggestions over time.
Practical examples from beta use
During testing the drafts included agreeing to an embargo on a pitch to obtain more details, and confirming meeting timing with minimal edits. In another case the feature correctly generated a reply declining to handle a contributed post for TechCrunch (the author notes TechCrunch does not accept contributed posts).
The feature is not perfect: by default it sometimes produced overly positive responses to pitches or agreed to meeting times after midnight. In those cases the user could quickly choose a different generated variation and send that instead.
Learning and personalization
The system adapts to user behavior: after an earlier problem with accepting a post-midnight meeting, a later similar suggestion produced a draft stating that the timing didn’t work. Users can further personalize outputs by visiting Settings > Personalization to add details about themselves and their role, and attach files or links for context.
Technology and company context
Superhuman co-founder Rahul Vohra said during testing that 40% of auto-generated drafts were sent within one day, and 60% of those were sent without any manual editing. Vohra noted that earlier features like Instant replies were built on older models such as GPT-3.5, which were less capable and had smaller context windows. The new implementation uses a mixture of models: the actual writing is handled by frontier models from both Anthropic and OpenAI.
In related corporate news, Grammarly acquired Superhuman last year and then rebranded the company as Superhuman. The company is also developing an assistant called Superhuman Go that carries context across platforms and between apps.
Why it matters
AI-generated first drafts have made it easier for communicators and PR professionals to produce initial messages, increasing inbound email volume. While the author is not comfortable fully handing inbox management over to AI, the updated auto-draft could help users reply to more people more quickly when lengthy typed responses aren’t necessary.
