Superhuman has rolled out an upgraded auto‑draft tool that crafts personalized email replies using advanced AI models. In testing, 60% of drafts were sent without any manual edits, promising a sizable boost in daily productivity.
मुख्य बिंदु (Key Takeaways)
- Superhuman's new auto‑draft sends 60% of drafts without edits.
- The feature generates two alternative responses for user selection.
- It leverages frontier models from Anthropic and OpenAI for richer context.
Since the explosion of large language models (LLMs), firms have raced to tame overflowing inboxes by deploying AI that both categorises messages and drafts replies that sound like the user. In this landscape, Superhuman has introduced a refreshed auto‑draft capability that not only spots important emails but also writes replies in a tone that feels less robotic.
From Early Experiments to a Refined Approach
Earlier attempts such as “Instant Replies” and “Follow‑up Auto‑Drafts” often produced overly enthusiastic, sales‑y responses that many users ignored. The latest iteration differs: it creates two additional variations for each draft, giving users the flexibility to pick the most appropriate tone. During a beta trial, the author dispatched several emails that required little to no editing.
How the Engine Works
The system first determines which messages likely need a reply, then analyses past conversations to emulate the user’s voice. Alongside the primary draft, it offers two alternative phrasings—examples include quietly agreeing to an embargo request or confirming a meeting time with minimal tweaks.
Real‑World Performance
In practice, the feature generated a clear refusal to author a post for TechCrunch (which doesn’t accept such submissions) and even drafted a positive response to a pitch that the author would normally decline. Occasionally, it suggested meeting times past midnight, but the alternative drafts allowed quick correction. Importantly, the model learns from these interactions; after the midnight‑meeting mishap, a subsequent similar suggestion was met with a draft stating the timing “doesn’t work for me.”
Looking Ahead
Co‑founder Rahul Vohra disclosed that 40% of auto‑generated drafts were sent within a day, and 60% of those required no manual editing. Earlier features relied on older models like GPT‑3.5, which had limited context windows. The current rollout blends frontier models from both Anthropic and OpenAI, maximising intelligence and contextual awareness. Building on this, Superhuman is also developing “Superhuman Go,” an assistant that carries context across platforms, further streamlining the user’s workflow.
