Superhuman’s Agentic Labs AI Auto-Draft Is the First That Feels Like Me
You know that cold dread when your inbox hits 50+ unread and the usual AI replies sound like they’re reciting a corporate memo written by a 2nd-year intern? Superhuman just fixed that—with an auto-draft feature that’s so good, co-founder Rahul Vohra told me 40% of drafts get sent within a day, and two-thirds of those sail through without so much as a keystroke.
This isn’t the first time Superhuman tried to automate replies. Back with Instant replies on GPT-3.5, the tone was stiff, the context shallow, and the hallucinations… well, let’s just say there were a few embargo confirmations I’d rather forget. But the new implementation? That’s where agentic labs AI shifts from buzzword to real utility.
I’ve been using the beta for ten days, and I’ve lost count of how many pitch rejections, meeting confirmations, and embargo responses shipped exactly as written. Not because I’m lazy—because the system learned my voice, my boundaries, and even how to say “no” without sounding rude.
So what changed? Why does this one feel like me? And why are so many people already trusting it with high-stakes replies?
It’s Not One Model—It’s a Mixture
Rahul Vohra puts it plainly: “Today, we’re using a mixture of models to make this work. The actual writing is done by frontier models from both Anthropic and OpenAI. So we’re applying the maximum amount of intelligence and context to this that we possibly can.”
That’s a big leap from the old Instant replies, which relied on GPT-3.5 and suffered from tight context windows and a lack of nuance. Now, Superhuman’s auto-draft can cross-reference your past conversations, your tone, your recent replies—even your calendar and CRM data—to produce something that sounds authentically you.
Think of it less like a template engine and more like an intern who’s read your entire email history, attended every team meeting you’ve missed, and knows exactly when to push back—or when to just hit send.
Real-World Behavior: When It Works—and When It Doesn’t
Here’s what happened on Tuesday:
- A startup founder pitched me a story for TechCrunch (we don’t accept authored posts). The draft said, “Thanks so much! I’m not the right person for that, but let me point you to our contributor guidelines.”
- Another pitch asked for a last-minute interview about an embargoed story. The auto-draft agreed, with timing that would’ve cut into my sleep. But I didn’t send it—I picked the second variation, which politely said, “Happy to cover this once the embargo lifts—here’s what I can commit to.”
The key isn’t just accuracy. It’s contextual awareness.
Vohra shared a telling stat: during beta, 60% of auto-drafted replies that users sent needed no editing at all. But they didn’t just copy-paste blindly—they trusted it because the variations gave them safe alternatives when the first draft crossed a line.
In short: it’s not about replacing your typing. It’s about replacing the 5-second hesitation before you type the third draft of “Sure thing!”
It Learns the Hard Way—By You Correcting It
The feature gets smarter because you correct it.
One morning, I approved a midnight meeting. The next day, when another pitch suggested 12:30 AM, the draft automatically said, “That timing doesn’t work for me—I’m available Tuesday at 3 PM instead.”
That’s the power of reinforcement learning, but without any engineering work on my part. Just send a few replies, reject one variation, and the next time it knows: midnight = hard no.
The same thing happened with tone. Early drafts sometimes sounded too eager, like a sales rep reading from cue cards. But after I rejected one such email, every follow-up auto-draft dialed back the enthusiasm just enough—keeping it friendly but not overeager.
Superhuman Go: The Unified Agent Around Auto-Draft
Let’s be honest—auto-draft alone isn’t enough. If it lives only inside Superhuman Mail, you’ll still miss 70% of your replies when you’re on Slack or Notion.
That’s where Superhuman Go steps in.
Launched last year after Grammarly acquired Superhuman and rebranded the company as Superhuman, Go is the unified AI assistant that lives wherever you work. It surfaces suggestions across tools, pulls in context from your email, calendar, and docs—and yes, it’s the same auto-draft engine working everywhere.
Think of Go as your co-pilot: it doesn’t change how you work; it just makes every tab, doc, and message smarter. When you’re prepping for a 1:1, Go reminds you of your last discussion and work you promised to finish. When a customer thread goes cold, it suggests the right next step without you leaving the chat.
The result? Auto-draft isn’t a bolt-on feature anymore. It’s baked into your whole digital workflow—inside email, outside email, and everywhere in between.
Personalization—Because Your Voice Isn’t Generic
The final piece is personalization. You can head to Settings > Personalization and add details about your role, preferred phrasing, tone, and even relevant files or links.
That’s how the system knows when to sound executive (“Let’s sync up shortly.”) versus casual (“Down for a quick coffee?”). And if your tone shifts over time—say, you go from early-stage founder to VP—the assistant adjusts automatically.
What’s cool is that Superhuman doesn’t force one voice. It learns yours and mirrors it, line by line, until the auto-draft feels less like a machine and more like your faster, smarter self.
The Bottom Line: It Works—If You Let It
I used to roll my eyes at every “AI replies” claim. Most sounded like generic boilerplate with extra steps.
Superhuman’s auto-draft is different. It gets better every time you use it—and uses a mixture of Anthropic and OpenAI models to do it right. It knows when to say no, how to phrase a rejection without burning bridges, and when to just send the damn email.
You don’t have to trust it blind. But if you let it learn your style, you’ll find yourself sending 60% of drafts with zero edits… not because you’re outsourcing your voice, but because for the easy stuff, it’s finally your voice—not an AI’s attempt at it.
Want to try it? Install Superhuman Go and give the beta a shot. Just keep an eye on your midnight meeting invites.