The Three-Hour Battlefield: Rahul Vohra on Why Email Survives and Agents Will Inherit the Inbox

Vohra argues that predictions of email's death miss the underlying reason email persists: the email address is the first thing a company gives a new employee because the company needs a way to address that person with an identifier it owns. A personal cell number belongs to the employee, but an email address belongs to the organization. The address is also the login for SSO and many other systems, making it foundational infrastructure rather than just an inbox.

Email itself will change, Vohra says, along two visible themes: AI and collaboration. On the collaboration side, Superhuman has added team read statuses, team reply indicators, team snippets and team scheduling. He highlights features users do not have to remember to use, such as seeing when a founder has opened a colleague's email, and says shared threads with comments are coming soon.

Vohra describes three flagship AI features. Write with AI lets users jot a few words and get a full email in their own voice and tone, then shorten, lengthen, improve, fix or translate it; it launched around June or July of the prior year. Auto Summarize puts a one-line summary above a conversation that updates as new emails arrive, with a shortcut to a bullet-point summary. Instant Reply, released a few weeks before the interview, drafts replies so users can edit and send, and Vohra says users write twice as fast with it.

The company's roadmap followed two phases. Phase one was on-demand AI features like Write with AI: easy to build, cheap to run because users must remember to invoke them, and useful for testing whether users love the capability. Phase two is always-on AI like Auto Summarize and Instant Reply, which are more ambitious, more expensive and operate at massive scale—Superhuman has processed 4 billion emails since launching them, compared with the 500,000–600,000 emails in the N1 corpus.

Phase three is the agentic future. Vohra imagines multiple mostly autonomous agents per person, able to take goals rather than tasks, plan and break goals into subtasks, work serially or in parallel, resolve ambiguity by asking questions, interrogate internal APIs or CRMs, and interact with other agents. He frames that shift as freeing people for more creative, strategic and impactful work.

Vohra says Superhuman began its generative AI journey in February after ChatGPT made clear that work would change, but the decision was not immediately obvious. He credits time away at David Hornik's Lobby conference in Hawaii—an intimate gathering of founders and VCs focused on the future—with giving him the perspective to prioritize AI. Back in the Bay Area, he concluded everything was about to change rapidly.

Prioritization was unusually hard because Superhuman normally evaluates features by user value or business value. With AI, the team could not predict how well the features would work or how users would feel, and it could not build a conventional business case. So Vohra took a philosophical stance—that LLMs would change everything, not overnight and not obviously—and used founder fiat to set direction, repeating the story many times.

With newly hired president Paul Tessier, Vohra created two operating models. Alpha mode is the default: teams are autonomous, receive goals and are held accountable. Theta mode embeds an executive, who does individual-contributor work day to day, because the project cannot fail and must move very fast. Superhuman ran Theta mode for AI starting in February with roughly one designer, one marketer and two or three engineers, working through basic questions like whether AI should live in a sidebar, be conversational or be embedded in the product surface.

Write with AI launched that summer and went viral in a way Superhuman had not experienced for years. Vohra posted a graph of new monetized retained seats per week that he jokingly called a thirst trap, and he says AI made the company top of mind again. He also offers a startup lesson: after avoiding cash problems and co-founder fallouts, the thing that kills companies is running out of steam. His advice, borrowing from Paul Graham, is to just keep typing.

Instant Reply required non-obvious product decisions. Vohra initially argued replies should only appear after a user starts replying, because most emails do not get replies and Superhuman had worked hard to keep the interface minimal. He then realized half the feature's value is inspiration: users often need to see three ways to respond before they can overcome the activation energy to reply at all. The team also worked to keep the feature tucked away until wanted and out of the way when not.

Reply length was another surprise. Write with AI generates medium-length replies, but Instant Reply felt wrong when it inserted a paragraph after a fast Tab-and-Enter interaction. Vohra realized that when users move quickly, the reply should match that speed, so Instant Reply inserts one or two sentences and often does not fill the screen. The interaction itself also changed: early versions used arrow keys, but switching to Tab and Enter made the feature roughly 10 times faster and more usable because arrow-key motions are elbow and shoulder movements that people avoid.

The prompt behind Instant Reply is enormous—Vohra says OpenAI thought Superhuman might have the largest prompt among its partners—and it explicitly avoids corporate jargon and phrases like “we apologize for the inconvenience.” Superhuman avoids fine-tuning because it takes time, requires assembling and tagging data, and does not transfer across model generations. Instead it pushes prompt engineering and multi-shot learning using the user's own emails, then relies on BrainTrust for regression tests and evals when prompts or models change.

Vohra says even successive versions of the same model can break a prompt, which forced Superhuman to build robust evals. The team uses LLMs as judges, asking whether Instant Reply addresses the right topic, the right person and the right length; early failures included replies like “thank you Alice for congratulating Bob” on long company-wide threads. On model choice, Vohra believed OpenAI was six to nine months ahead on quality but somewhat behind on speed, and he valued OpenAI's collaboration and aggressive cost reductions. For Instant Reply and Auto Summarize, throughput and latency matter more than raw intelligence, so Superhuman uses GPT-3.5 Turbo rather than GPT-4 because GPT-4 reserved instances have lower tokens-per-minute limits and the feature is extremely bursty—especially when everyone queues email for 8:55 on Monday morning.

Superhuman has charged $30 a month for eight years, with lower prices for teams or contracted accounts, and Vohra says pricing is difficult but the product is far better than it was eight years ago. In the early days, extra revenue went into concierge onboarding and training; today the ratio has flipped from roughly 85% concierge and 15% self-service to 85% self-service and 15% concierge. Superhuman has shifted those dollars into AI and into building what Vohra calls the fastest, most powerful AI email experience.

Pricing for AI is an open project. Vohra says Superhuman is working through how to price Enterprise, Superhuman for Sales, a possible platinum edition and AI itself, and he hopes to have an answer in a few months. Superhuman for Sales includes HubSpot and Salesforce integrations and a recent-opens feed. Vohra says emailing someone within five minutes of an open makes a reply about 20 times more likely, as long as the sender does not make it weird; good salespeople, recruiters, founders and investors know how to use that signal properly.

The enterprise motion is real. Vohra says a tier-one strategy consulting firm bought Superhuman for 1,500 managing-director partners after Superhuman showed those partners were replying to emails 8 to 13 hours faster. Enterprise buyers also ask about InTune integration, data loss prevention and stronger anti-phishing features. Vohra acknowledges that pricing is hard across all these streams but sees clear business impact.

As an investor, Vohra says he and Todd were bearish on most de novo AI companies because many pitches amounted to “the app we were going to build anyway, now with AI,” which means building two companies at once. He would not want to build a new email client today and also try to make it the best AI email client; that would require raising hundreds of millions and running two companies. He is bullish when founders have a distribution advantage—contacts from prior work, a distribution insight, a viral loop or a track record—because the real investment is in the founder's ability to create and retain distribution.

Vohra is most excited about the agentic ecosystem, though he does not know what it will look like. He expects some service—possibly ChatGPT or something else—to federate a person's agents with an authentication handshake. In a family-planning example, he imagines asking his executive AI when open enrollment is; it would know the roster of his other agents, authenticate to the Rippling agent, ask about open enrollment, share pregnancy-planning details and return a recommended health plan instantly, without meetings or phone calls.

The trillion-dollar question is how the ecosystem of agents comes together and who owns the user relationship. Vohra points to Superhuman's strategic advantage as one of the few three-hour-a-day products: email, chat and messaging are where professionals spend huge amounts of time. He does not think Superhuman must own every agent; instead, every company may build an agent for its domain, and Superhuman could be the pane of glass where users interact with an email agent that triages, drafts, schedules, sends and talks to other agents—through a sidebar or eventually voice.

Vohra agrees incumbents have a greater advantage in AI than in past cycles, but not so great that founders should avoid them. Incumbents must build one-size-fits-all products: Office 365 has more than 400 million paid seats and Gmail has more than a billion users, which by definition means a broad solution. A founder can target an underserved but economically powerful subsegment, build a better product for it and turn an uncontrollable market into a controllable product problem.

Speed is a structural weakness for incumbents. Superhuman launched on being the fastest email experience, with instantaneous response and search and sub-100-millisecond user interactions; those speed claims became the foundation for getting through email twice as fast, replying one to two days sooner and saving four hours or more each week. Vohra notes that Superhuman looks like a web app but was the first email client written entirely in JavaScript, a switch that would be very difficult for an incumbent after 20 years as a client-server application.

Design is another opening. Vohra invokes Conway's law: a product's structure often reflects the organization's structure more than what is logical for users. He recalls looking at Google Workspace and wondering why Calendar was not in the left-hand navigation; the answer was that Calendar lived in the grid of nine squares because different product managers owned different areas. He argues that launching gets rewarded inside large companies, which helps explain repeated attempts at Google Meet or Hangouts, even when the products do not cohere for users.

On market size, Vohra says Superhuman progressed from founders and VCs to leaders and managers to any outbound professional—salespeople, recruiters, BD people, consultants and community evangelists. The market math is simple: about a billion professionals spend an average of three hours a day reading and writing email, which is 3 billion hours a day and more than a trillion hours a year. At $30 a month, 300,000 subscribers is a $100 million run-rate business; in the a16z boardroom, he asked whether anyone doubted Superhuman could reach 300,000 subscribers, and no one did.

Vohra calls AI job displacement both overhyped and underhyped. It is underhyped for entry-level customer support and sales roles, where an AI that is only 80% as good as a human but costs 10% or less will be an easy trade-off for many companies—and may improve the consumer experience over today's robotic voice support. It is overhyped if people think those roles are uniquely exposed: he says there is nothing unique about customer service or sales except headcount, and the same pressure will reach his job as CEO.

He frames the shift historically. A thousand or 1,500 years ago, subsistence farmers or hunter-gatherers would look at modern jobs and say people are just playing; many people today would say the same about venture capital or podcasting. Vohra expects people 50 years from now to look primarily engaged in what we perceive as leisure, though it will not feel like leisure because humans are hardwired to work, stress, compete and play status games—a point he attributes to Bertrand Russell.

Asked which other company would be interesting to build generative AI features at, Vohra names Slack or Microsoft Teams because they are also three-hour-a-day products. He would not start with AI there; he would first build a better workplace product, then make it AI-enabled. His Slack critique is that conversations are ordered by when they started rather than by the last reply, there are no subject lines unless users write them, and users cannot mark things done or snooze them. His own Slack might have 100 to 150 unread channels with interjumbled conversations and inconsistent thread discipline, so he opens it, closes it and looks at his calendar instead.

Vohra still sees email's weaknesses: messages and subject lines are immutable, and the architecture is fully decentralized, with everyone storing every copy, which makes shared threads and comments hard to build. But email offers subject lines, done, snooze, assign, sharing and inbox zero, and he would combine the best of email and chat into a workplace experience designed for thoughtful asynchronous discussion with a choice of modality. He also imagines a future-of-work agent that knows what is important across Slack, email and the calendar. His own coping mechanism is to stop worrying about what to do next and trust his calendar and his executive assistant, who tracks company priorities and his preferences; the mission behind Superhuman, he says, is helping professionals feel happier, more productive and closer to their potential.