Superhuman's refined AI draft feature finally delivers natural email replies

Email overload has long been a fixture of modern work life. Since the recent wave of large language models, many startups have attempted to ease the burden by deploying AI to sort messages and compose responses on behalf of users. Now, email client Superhuman has rolled out an improved version of its auto-draft capability that spots messages requiring replies and crafts responses that feel genuinely human rather than algorithmically sterile.

A meaningful step forward#

Earlier iterations of Superhuman's drafting tools—instant replies and follow-up suggestions—fell flat. The generated messages read like they came from an overeager corporate bot, discouraging actual use. The updated feature is a marked departure. After testing the beta, reviewers found themselves sending multiple auto-generated drafts with minimal tweaking. The system identifies emails worth responding to and constructs replies mirroring your conversational style, while also presenting two alternative versions from which to choose.

In practice, this translated to drafts confirming meeting times or acknowledging press embargoes with just a few keystrokes of adjustment. The AI also deftly declined irrelevant pitches—for instance, declining requests for authored posts with language that felt true to voice.

Room for refinement#

The feature isn't flawless. Early defaults leaned toward accepting invitations, sometimes proposing midnight meetings that required manual course correction. However, the system learns: when similar timing appeared in a subsequent message, the revised draft automatically pushed back, suggesting the unsuitable hour didn't work.

Users can refine the feature by visiting Settings > Personalization to provide biographical details and upload reference materials, helping the AI better calibrate its tone.

Under the hood#

Superhuman co-founder Rahul Vohra revealed that during testing, 40% of generated drafts were dispatched within a single day, and three-fifths required no human adjustment whatsoever. The engine powering this improvement contrasts sharply with earlier efforts built on GPT-3.5, which lacked the sophistication or context capacity of newer systems. "Today, we are using a mixture of models to make this work," Vohra explained, noting that the actual composition taps frontier models from both Anthropic and OpenAI.