Superhuman's Auto-Draft = Where AI Writing Is Headed?

 The bar for "AI-written email that doesn't sound like AI" has been embarrassingly low for years. Superhuman's new feature is one of the first credible signs that bar just moved. 

Key Points

• The feature: Superhuman's redesigned auto-draft identifies important emails and writes replies matched to the sender's tone, offering two alternate drafts alongside the primary suggestion.
• The adoption number: Superhuman says 40% of auto-generated drafts were sent within a day during testing, and 60% of those went out with zero manual editing.
• The model stack: The feature runs on a mix of frontier models from both Anthropic and OpenAI, a departure from earlier versions built on older, smaller-context models like GPT-3.5.
• The learning loop: The system adjusts based on corrections, catching a repeated late-night meeting request after being corrected once.
• The company context: Superhuman, acquired by Grammarly in 2025 and now operating under the Superhuman name, is building toward a cross-platform assistant called Superhuman Go.

Superhuman's New Auto-Draft Feature

Superhuman released a new version of its auto-draft feature, building on earlier attempts like instant replies and follow-up drafts that TechCrunch's Ivan Mehta described as sounding like an overly enthusiastic AI salesperson in past iterations. The redesigned version identifies which incoming emails likely need a response, then generates a primary draft written in the recipient's own tone based on past conversations, plus two alternate variations to choose from instead.

In hands-on testing, Mehta reported sending several of these drafts with little to no editing, including replies agreeing to embargo terms on a pitch and confirming meeting times. The feature isn't flawless: it defaulted to a positive response on a pitch and once agreed to a post-midnight meeting time before Mehta corrected it, after which the system stopped suggesting similarly timed meetings.

Why The Model Mix Matters More Than The Feature Itself

The most telling detail in this release isn't the UI, it's the model architecture behind it. Superhuman co-founder Rahul Vohra told TechCrunch that earlier versions of the auto-reply feature ran on older models like GPT-3.5, constrained by both intelligence and a smaller context window. The new version runs on what Vohra described as a mixture of frontier models from both Anthropic and OpenAI, chosen specifically to maximize intelligence and context for the writing itself.

That's a meaningful shift in how a mainstream productivity product is built. Rather than picking one vendor and optimizing around its limits, Superhuman is treating model selection as a live decision, routing the actual writing task to whichever frontier model handles it best. For a feature whose entire value proposition rests on sounding like a specific human rather than a generic assistant, that model flexibility appears to be doing real work.

Why This Reads As Next-Level LLM Behavior

What separates this from years of failed "smart reply" attempts is the combination of three things working together rather than in isolation: tone-matching pulled from real conversation history, multiple draft variations instead of one forced answer, and a correction loop that actually holds. The midnight meeting example is the clearest evidence of that loop functioning as intended. The system didn't just accept a correction once, it generalized the correction to a similar future request, which is the behavior people have been promised by AI assistants for years without reliably getting.

Vohra's adoption numbers back this up. A 40% same-day send rate, with 60% of those requiring zero edits, is a meaningfully different outcome than earlier auto-reply tools ever reported, and it suggests the tone-matching is clearing a bar users didn't think was clearable a year or two ago.

What Marketers And Growth Teams Should Take From This

Email volume driven by AI-assisted outreach is only going up, and this feature is a direct response to that same pressure from the receiving end.

• Expect faster response cycles from prospects: If tools like this become standard, the person receiving your outreach may be replying through an AI draft rather than typing from scratch, which changes how quickly and how generically people respond to cold email and pitches.
• Rethink what makes outreach worth a real reply: A tool that auto-declines low-value pitches, as it did with an authored-post request in Mehta's testing, means generic, templated outreach is more likely to get filtered before a human ever reads it.
• Watch model-mixing as a pattern, not a novelty: Superhuman routing tasks across Anthropic and OpenAI models based on strength is a preview of how more marketing and growth tools will likely be built going forward, and it's worth factoring into how your team evaluates the growth strategy behind any AI-assisted workflow you adopt.

Assistant, Not Autopilot

Even Mehta, hands-on with the beta, was clear that he isn't ready to hand his entire inbox to AI. That distinction, assistant rather than autopilot, is likely where this category settles for the next stretch. The value isn't full automation, it's removing the friction of typing out routine replies while keeping a human in the loop for judgment calls.

Superhuman's broader ambition, a cross-platform assistant called Superhuman Go carrying context between apps, suggests the company sees this auto-draft feature as a first step rather than the destination. If email replies can get this close to indistinguishable from a person's actual voice, the next question is which other repetitive, context-dependent writing tasks follow the same path.

If your team is trying to figure out where AI-assisted communication actually saves time versus where it just moves risk around, that's the exact kind of evaluation our AI marketing services team helps clients work through.