Superhuman's AI auto-draft now sounds human enough to send

Email client Superhuman has long struggled with an AI problem that plagues many productivity tools: auto-generated replies that sound robotic and fail to capture the sender's voice. In the past, features like Instant replies and follow-up auto-drafts were built on older models such as GPT-3.5, which had smaller context windows and less nuanced understanding of tone. Users largely ignored these earlier attempts because the generated emails sounded like an overly enthusiastic AI salesperson pitching a deal.

The new auto-draft feature works differently. It identifies which emails merit a response and generates a draft based on your tone from previous conversations, then offers two additional variations you might choose instead. In testing, the feature demonstrated surprising effectiveness: according to Superhuman co-founder Rahul Vohra, 40% of auto-generated drafts were sent within one day, and of those, 60% were sent without any manual editing—a dramatic jump in usability compared to earlier iterations. The author of the article reported sending generated drafts with little to no editing in several real-world cases: confirming meeting times, agreeing to embargo requests, and declining requests to write authored posts for TechCrunch.

The feature is not infallible. By default, it often generated overly positive responses to pitches or agreed to meetings scheduled at post-midnight times, requiring users to select an alternative from the offered variations. However, Superhuman built in a learning mechanism: after the author rejected a midnight meeting suggestion, when a similar late-night meeting was proposed days later, the feature automatically generated a draft declining the timing. This learning loop gives the feature a collaborative feel rather than a black-box automation.

The improvement in quality stems from Superhuman's technical approach. Vohra explained that the company is now using "a mixture of models" to power the feature, with the actual writing done by frontier models from both Anthropic and OpenAI, applying "the maximum amount of intelligence and context" possible. This is a marked shift from older LLM-based replies. Users can customize the feature's behavior by navigating to Settings > Personalization and adding details about themselves, their role, and providing files or links for additional context. Superhuman, which was acquired by Grammarly last year and rebranded under Grammarly's ownership, is now building a broader assistant called Superhuman Go that will carry context across platforms, though the auto-draft feature itself remains email-focused for now.

Superhuman's challenge has always been the uncanny-valley problem of AI-generated email: it automates a task people do constantly, but when the output sounds off, it breaks trust and wastes time in editing. Previous features like Instant replies suffered from sounding like an overly enthusiastic sales bot, so adoption remained low despite the time-saving potential. The new version tackles this by anchoring drafts to the user's own writing style—analyzing tone from past conversations rather than generating generic corporate-speak. By offering three variations and learning from user rejections (Vohra's example of the midnight-meeting fix illustrates this), Superhuman shifts from a fire-and-forget automation to a collaborative drafting tool that improves with use. This matters because the email inbox is becoming a bottleneck precisely as AI makes it easier for others to send more email; Superhuman positions itself as the relief valve for people receiving thousands of messages monthly but who need more control than a fully automated system would allow.