From: redpointai
Rahul Vohra, CEO of Superhuman, an email tool, discussed the challenges and strategies for competing with established email providers like Gmail and Outlook, particularly in the context of AI integration [00:00:21].
The Enduring Nature of Email
Journalists have perennially claimed the death of email, but it has persisted [00:01:35]. This persistence is less about email itself and more about the fundamental role of the email address [00:01:45]. When joining a company, an individual receives an email address because it provides a company-owned way to address them, unlike personal cell numbers [00:01:56]. Email addresses serve as an underlying piece of infrastructure for logging into various solutions like SSO (Single Sign-On) [00:02:30].
While email addresses are here to stay, the nature of email itself is evolving, primarily driven by AI and collaboration [00:02:40].
Superhuman’s Approach to Incumbent Competition
Superhuman’s strategy against email incumbents like Outlook and Gmail focuses on exploiting their inherent limitations:
Targeting Underserved Segments
Incumbents are forced to create one-size-fits-all solutions due to their massive user bases (e.g., Microsoft Outlook with 400 million+ paid seats, Gmail with over a billion users) [00:48:17]. A startup can identify an underserved, economically powerful segment of this market and build a superior product specifically for them [00:48:37]. This shifts the challenge from acquiring an unfair distribution advantage to simply building a better product [00:48:55]. Superhuman initially focused on founders and VCs, then expanded to leaders, managers, and any “outbound professional” (salespeople, recruiters, consultants, etc.) [00:52:02].
Leveraging Product Speed
Incumbents struggle with product speed [00:49:15]. Superhuman was launched on the premise of being the fastest email experience, aiming for instantaneous response, sub-100 millisecond search, and allowing users to get through emails twice as fast, reply one to two days sooner, and save four hours or more per week [00:49:18]. The foundational speed of the app, built entirely in JavaScript (a significant architectural shift from traditional client-server applications), is incredibly difficult for incumbents to replicate after decades of development [00:49:50].
Design Philosophy
According to Conway’s Law, the structure of a product often reflects the organization that built it, rather than being logical for the user [00:50:06]. Incumbents like Google (Gmail) exhibit this, with product features being siloed based on internal team structures rather than user flow (e.g., Calendar being separate from Mail) [00:50:18]. Startups, being leaner, can prioritize a seamless, user-centric design [00:17:15].
Superhuman’s design principles include:
- “When you want it, and out of the way when you don’t”: This applies to features, ensuring a minimal interface [00:19:34].
- Minor UX tweaks at scale: Accumulating hundreds of small, seemingly minor UX improvements creates a fundamentally different product experience [00:33:03]. For example, switching from arrow keys to Tab and Enter for instant replies made the feature 10 times more usable and faster [00:22:50].
AI as a Competitive Edge
Superhuman’s AI strategy has been crucial in its growth:
Phased AI Feature Rollout
Superhuman followed a three-phase approach for AI feature development [00:06:06]:
-
On-Demand AI Features:
- Write with AI: Launched in June/July (last year relative to recording), this feature allows users to jot down a few words and generate a full email that matches their voice and tone. It also offers options to shorten, lengthen, improve, or fix writing, and change language [00:04:19]. These features are easier and cheaper to build/run as users must actively remember to use them [00:06:26].
- This phase helped assess technology understanding, feature quality, and user reception [00:06:43].
-
Always-On AI Features:
- Auto Summarize: Launched in November (last year relative to recording), this feature provides a one-line summary above email conversations that updates instantly with new emails. A shortcut reveals a bullet-point summary [00:04:42]. Users often read the summary first, sometimes negating the need to read the full email [00:05:17].
- Instant Reply: Released a few weeks prior to the discussion, this feature presents draft replies for every incoming email, allowing users to simply edit and send, or sometimes send without edits [00:05:30]. Users of Instant Reply write emails twice as fast, demonstrating significant business impact [00:05:57]. This feature also provides “inspiration to reply” by showing three possible response options, even if not fully used [00:09:04].
- These features are more ambitious, difficult, and expensive due to their continuous operation at massive scale [00:07:05]. Superhuman has processed 4 billion emails since launching Auto Summarize and Instant Reply [00:07:28].
-
Agentic AI Future:
- The long-term vision involves AI agents, potentially multiple per user, operating autonomously [00:07:57].
- These agents would receive goals (not just tasks), plan and break them down into subtasks, handle ambiguity by asking questions, interrogating systems (like CRMs or internal APIs), and even interact with other agents [00:08:08].
- This will free up users to be more creative, strategic, and impactful [00:08:32].
Development Strategy
Superhuman prioritized AI as an “existential” project, leading to a “Theta mode” operating model where executives are deeply embedded in the project, treating it as critical [00:14:07].
- Prompt Engineering over Fine-Tuning: Superhuman aims to avoid fine-tuning models as much as possible due to time, cost, and transferability issues across model versions [00:25:05]. Instead, they push prompt engineering “as far as it will go” [00:25:29], using multi-shot learning based on a user’s past emails to match their voice and tone [00:25:39]. Their prompt for Instant Reply is notably complex and large [00:23:22].
- Robust Testing: Successive model versions can break existing prompts, necessitating a robust regression testing and evaluation framework [00:26:17]. They use Brain Trust for this, and even use LLMs to assess their own output (e.g., checking if replies address the right topic/person) [00:26:48].
- Model Selection: Superhuman primarily uses OpenAI models, believing them to be 6-9 months ahead in model quality, committed to aggressively lowering costs, and providing excellent collaborative support [00:29:28]. Model speed is critical; for features like Instant Reply and Auto Summarize, peak throughput is more important than raw intelligence, making models like GPT 3.5 Turbo more feasible due to higher tokens-per-minute limits compared to GPT-4 [00:31:39].
Pricing Strategy for AI
Superhuman has maintained its $30/month price point for eight years, despite adding significant AI features [00:38:10]. The revenue previously allocated to onboarding and training has been shifted to funding AI development [00:33:59]. The expectation is that model costs will continue to decrease, eventually enabling edge LLMs running on devices [00:30:03].
Educating Users on New Workflows
A key challenge for AI startups is teaching users new workflows [00:34:36]. Superhuman’s trick is to:
- Start with Always-On Features: Features like Auto Summarize and Instant Reply naturally expose users to AI’s benefits without requiring them to “remember” to use them [00:36:09]. Such “100% reach features” are rare and invaluable [00:35:21].
- Connect Always-On to On-Demand Features: For example, after an Instant Reply is inserted, buttons like “lengthen” or “custom” can appear, leading users to the full “Write with AI” prompt, which then provides a keyboard shortcut for future use [00:36:25]. This subtly guides users to more advanced features.
- Provide Intuitive UI: Offering options and buttons for AI interactions, rather than solely relying on chat-like prompts, gives users more control and improves usability [00:37:11].
Investment Perspective: Taking on Incumbents
From an investor’s standpoint, Rahul Vohra is generally bearish on de novo AI companies that lack a distribution advantage [00:40:58]. Building both a new application and integrating AI effectively is extremely difficult, akin to running two companies simultaneously [00:41:14].
He is bullish on companies that possess a distribution advantage (e.g., existing contacts, viral loops, or a track record of launching successful consumer apps), as the investment is then in the founder’s ability to create distribution and retain users, with AI serving as a hook [00:41:42].
Incumbents, despite their size, are still “encumbered” [00:47:24]:
- They move slowly [00:47:27].
- They are limited by their need for one-size-fits-all solutions [00:47:30].
- They struggle with product speed [00:49:15].
- Their design is often a reflection of their organizational structure (Conway’s Law), not optimal user experience [00:50:06].
- Internal incentives often reward launching new features regardless of performance, leading to feature bloat and lack of coherence (e.g., Google’s various communication apps) [00:51:20].
These weaknesses provide significant opportunities for startups to compete [00:51:37].
The Agentic AI Future
The future of AI involves multiple AI agents for each individual [00:57:57]. This raises a “trillion-dollar question” about how this ecosystem will come together [00:44:37]. Superhuman believes its strategic advantage lies in its high user engagement (3 hours a day for 50%+ of users), making it a primary “pane of glass” [00:44:51].
Rahul envisions:
- Companies building agents specific to their domain (e.g., an email agent to triage, draft, schedule, and send emails, interacting with other agents) [00:45:45].
- A centralized service (like ChatGPT or another platform) that federates these agents, handling authentication and communication between them [00:42:28].
- User interaction with this agent ecosystem via an interface within Superhuman, or even eventually by voice [00:46:06].
This future will lead to an “Enlightenment” period, where technology allows humans to focus on higher-level, creative, and strategic tasks, transforming what we perceive as “work” into something that might resemble “leisure” to future generations [00:57:50].