Superhuman released a new version of its auto-draft feature that identifies important emails and writes reply drafts that sound less robotic than previous versions. According to TechCrunch tests, 40% of generated drafts are sent within a day, and 60% of those without manual edits.

The feature analyzes tone from past conversations and generates three variants to choose from. In the journalist’s tests, drafts correctly handled cases like embargo confirmations, meeting scheduling, and declining out-of-scope requests, with recognizable errors (absurd times, overly positive tone) that could be fixed by selecting a different variant.

Why this matters. It’s an example of AI applied to a repetitive task where the cost of error is low and time savings are obvious. The test shows the feature learns from use: after correcting a response about an absurd meeting time, subsequent suggestions account for that feedback. It doesn’t solve the problem of difficult emails—those require the two-step playbook—but it can speed up courtesy responses and administrative confirmations where content is predictable.

If you want to try it. The feature is available in Superhuman and customizes in Settings > Personalization, where you can add information about your role and context to improve draft tone.

In detail

What came before

Superhuman had already tried similar features with Instant Replies and follow-up auto-drafts, but according to TechCrunch testing those versions produced text that sounded like “an overly enthusiastic AI salesperson.” Previous versions used GPT-3.5, a model with reduced context windows and more limited capabilities than current frontier models.

What changes now

The new implementation uses “a blend of models” from Anthropic and OpenAI for actual writing, according to co-founder Rahul Vohra. The feature works on three levels:

1. Automatic selection: identifies which emails might need replies
2. Multi-variant generation: produces three different drafts for each selected email
3. Learning from use: adapts subsequent suggestions based on user corrections and choices

The TechCrunch journalist’s field test shows concrete results: usable drafts for embargo confirmations, detail requests, meeting scheduling. Typical errors—overly positive tones on unsolicited pitches, post-midnight times proposed for meetings—were visible and quickly correctable by selecting another of the three variants.

Stated limitations

The feature isn’t designed to replace human judgment on complex or sensitive emails. The journalist explicitly writes: “I don’t have the confidence to completely hand the reins to AI for managing my inbox.” The utility lies in reducing time on responses where content is largely predictable.

The 60% sent-without-edits percentage needs context: we’re talking about administrative emails, confirmations, courtesy responses. It doesn’t include emails where tone, substance, or implications require attention.

The architecture behind it

Vohra explains that choosing multiple models instead of one addresses a practical problem: “We’re applying the maximum amount of intelligence and context possible.” This likely means each phase (classification, generation, tone adaptation) uses the most suitable model for that specific task, rather than loading everything onto a single model.

Superhuman’s acquisition by Grammarly in 2025 brought resources to develop Superhuman Go, an assistant that spans multiple platforms while maintaining context. Auto-draft fits this direction: the idea is the assistant knows what you did in other apps and uses that context to write better.

What this doesn’t conclude

This feature doesn’t solve information overload: responding faster to more emails might simply increase the volume of conversations you need to follow. And it doesn’t replace judgment about when not responding is the right choice.

For emails where tone matters, where you risk compromising a relationship, or where stakes are high, the difficult email playbook remains the most reliable approach: first decide the points to cover and tones to avoid, then let AI draft the response.