Superhuman's AI email assistant is getting good enough to actually use
A new auto-draft feature uses frontier AI models to write replies that match your tone. In testing, 60% of generated drafts were sent without any editing at all.
Key points
- Superhuman, the email app owned by Grammarly, launched a new auto-draft feature in beta that writes replies in your personal tone.
- During testing, 40% of AI-generated drafts were sent within one day, and 60% of those were sent without any manual edits, according to Superhuman co-founder Rahul Vohra.
- The feature uses frontier AI models, meaning the most capable and up-to-date models available, from both Anthropic and OpenAI to write draft replies.
- Earlier Superhuman reply features relied on GPT-3.5, an older and less capable model, which Vohra says produced stiffer, less natural-sounding text.
- Users can personalise how the AI writes by going to Settings, then Personalization, and adding details about themselves and their role.
For anyone who spends part of their working day buried in email, the promise has always been the same: let AI handle the easy replies so you can focus on the hard ones. Superhuman, the email app, is now making a serious attempt to deliver on that promise.
The company launched a beta of its revamped auto-draft feature this week. It reads through your inbox, works out which messages actually need a reply, and then writes one that aims to sound like you, not like a form letter. It also offers two alternative versions, so you can pick the tone that fits.
First reported by TechCrunch AI, the feature learns from past conversations to mirror how you write. It improves over time based on which drafts you accept, edit or reject. In one example during beta testing, the system agreed to a meeting scheduled after midnight. After the user sent a correction, the feature began declining similar requests automatically.
The step forward here is the underlying technology. Superhuman's co-founder Rahul Vohra said earlier reply tools were built on GPT-3.5, an older AI model with a narrower understanding of context. The new feature runs on a mix of the latest models from Anthropic and OpenAI, both companies that build some of the most capable AI systems currently available.
"The actual writing is done by frontier models from both Anthropic and OpenAI," Vohra said. "We're applying the maximum amount of intelligence and context to this that we possibly can."
Should ordinary email users care about this?
Not yet, but it's worth watching. The feature is currently in beta, meaning it is still being tested and is not available to everyone. When it does roll out broadly, it will sit inside the Superhuman app, which already requires a paid subscription. Grammarly, the writing-assistant company, acquired Superhuman last year.
For those who do get access, the practical upside is real. Short replies, meeting confirmations, standard declines: these are exactly the emails that eat time without requiring much thought. The feature handles them passably and is getting better.
The feature is not fully reliable. It sometimes generated overly positive responses or agreed to unsuitable meeting times during early testing. Checking the draft before you hit send is still essential.
Superhuman is also building a broader assistant called Superhuman Go, designed to carry context across multiple apps, not just email. That product is still in development.
