From: redpointai
Superhuman, a “lightning fast email tool,” has strategically integrated Artificial Intelligence (AI) to redefine the email experience, impacting everything from individual productivity to enterprise workflows [00:00:23]. The company’s journey with AI has been phased, leading towards a future where AI agents play a central role in managing digital communication [00:03:02].
The Enduring Role of Email
Despite perennial claims of its demise, email remains a fundamental piece of digital infrastructure [00:01:40]. This is primarily due to the email address serving as a unique identifier owned by companies, unlike personal cell numbers [00:01:44]. Email addresses are often the login for crucial systems like Okta or Rippling [00:02:30]. While the email address itself is here to stay, the email experience is poised for significant change, driven by AI and collaboration [00:02:40].
Superhuman’s AI Evolution: A Phased Approach
Superhuman’s approach to integrating AI has followed a distinct three-phase strategy [00:04:06].
Phase 1: On-Demand AI Features
The initial phase focused on “on-demand” AI features that users explicitly activate [00:06:26].
- Write with AI: Launched in June or July of the previous year, this feature allows users to jot down a few words, and Superhuman generates a full email, matching the user’s voice and tone from previously sent emails [00:04:19]. It also includes options to shorten, lengthen, improve, or fix writing, and even change languages [00:04:32].
- These features are relatively easy and cheap to build and run, serving as a way to test user adoption and technology effectiveness [00:06:33].
Phase 2: Always-On AI Features
Building on the success of on-demand features, Superhuman moved to “always-on” AI capabilities that work continuously for the user [00:06:56]. These are more ambitious, difficult, and expensive to run due to their constant operation at massive scale [00:07:05].
- Auto Summarize: Launched in the previous November, this feature provides a one-line summary above email conversations that updates instantly as new emails arrive [00:04:42]. It can significantly change workflows, as users often read the summary first and sometimes don’t need to read the full email [00:05:19].
- Instant Reply: Released a few weeks prior to the discussion, this feature drafts replies for every incoming email, allowing users to simply edit and send, or sometimes send without any edits [00:05:30]. Users of Instant Reply write emails twice as fast, demonstrating significant business impact [00:05:57]. The mere presence of suggested replies helps users make decisions faster [00:08:48]. This feature often enables users to reply to emails they might otherwise have ignored [00:09:26].
The scale of these features is immense; since launching Auto Summarize and Instant Reply, Superhuman has processed 4 billion emails, compared to a typical LLM training corpus of 500,000 to 600,000 emails [00:07:21].
Phase 3: Agentic AI Future
Superhuman envisions a future with “agentic AI,” where AI agents operate autonomously around a user’s email [00:03:02].
- These agents will be mostly autonomous, capable of handling goals (not just tasks), planning, breaking goals into subtasks, resolving ambiguity by asking questions or interrogating other systems (like internal APIs or CRM), and even interacting with other agents [00:08:08].
- This will free users to be more creative, strategic, and impactful [00:08:32]. For example, an executive AI could coordinate with a Rippling AI to answer complex HR questions instantly [00:43:24]. Superhuman believes its significant daily usage (3 hours/day for 50%+ users) positions it as a prime “pane of glass” for orchestrating these agents [00:44:51].
Product Design and Development Strategy
Superhuman’s rapid advancements in AI product development stem from a specific organizational and design philosophy.
Prioritization and Operating Models
The decision to heavily invest in generative AI in February of the previous year was not immediately apparent, requiring a “philosophical stance” that LLMs would change everything [00:11:00].
- Challenge: It was hard to quantify the value or business case of these features early on [00:12:09].
- Solution: Founder fiat was used to prioritize AI, supported by a new operating model developed with President Paul Tessier [00:12:48].
- Alpha Mode: Default mode, teams operate autonomously, accountable to goals (normal business) [00:13:40].
- Theta Mode: For existential projects, an embedded executive (e.g., CEO) does individual contributor work, with the expectation that the project cannot fail [00:13:52]. Superhuman operated in Theta mode for AI [00:14:14].
Design Principles for AI Features
Early in the development of super intelligence, there was no clear playbook for AI product design (e.g., sidebars vs. conversational interfaces vs. in-product) [00:14:48].
- “When you want it, out of the way when you don’t”: This core principle guided the design of features like Instant Reply [00:19:34]. For Instant Reply, the decision to show suggested replies before the user starts typing was counter-intuitive but crucial for inspiration, despite initial concerns about clutter [00:18:12].
- Matching Interaction Speed: Instant Reply generates very short (1-2 sentences) snappy responses to match the user’s desire for quick movement, avoiding the need to proofread long drafts [00:21:07].
- UX is Key: The interface design, such as using “Tab and Enter” instead of arrow keys, was critical to making Instant Reply 10x faster and more usable [00:21:07]. This focus on seemingly minor UX tweaks, done hundreds of times, creates a fundamentally different product [00:23:03].
Technical Considerations
Prompt Engineering vs. Fine-Tuning
Superhuman largely avoids fine-tuning models due to time, cost, and the inability to transfer learned data across model versions [00:25:52]. Instead, they push prompt engineering “as far as it will go” [00:25:29].
- Their prompts are remarkably large, with OpenAI noting Superhuman as a partner with one of the biggest prompts [00:26:22].
- They use multi-shot learning (providing examples of the user’s past emails) to ensure the AI sounds like the user and avoids corporate jargon [00:24:44].
Evaluation and Model Choice
- Robust Evals Framework: Superhuman uses Brain Trust for a regression testing and evals framework, which is crucial because even successive versions of the same LLM can break existing prompts [00:26:41]. LLMs are even used to assess themselves (e.g., checking if replies address the right topic/person) [00:27:08].
- OpenAI Partnership: Superhuman chose OpenAI models because they believe OpenAI is 6-9 months ahead in model quality [00:29:28].
- Speed and Cost: While OpenAI has a lead in quality, speed is also paramount for real-time features like Instant Reply [00:29:39]. OpenAI’s commitment to aggressively lowering costs was also a factor [00:29:49].
- Collaboration: Superhuman benefits from deep collaboration with OpenAI, including engineering support for complex prompt challenges [00:30:31].
- Throughput over Intelligence: For features like Instant Reply and Auto Summarize, peak throughput and low latency are more critical than the sheer intelligence of the model, which is why GPT-3.5 Turbo is currently preferred over GPT-4 due to its higher tokens per minute limit [00:31:31]. Massively accelerated hardware like Groq will further enable new use cases [00:32:46].
Business Model and Pricing
Superhuman charges a premium ($30/month retail) [00:33:14]. The revenue generated, which once went towards intensive onboarding and training, is now largely redirected to funding AI development [00:33:32].
- Value Proposition: Superhuman focuses on delivering significant value, such as saving users four hours or more per week and enabling replies one to two days sooner [00:33:59].
- Enterprise Use of AI and Model Specialization: The company is expanding with “Superhuman for Sales” (with HubSpot and Salesforce integrations, and a “recent opens” feed that boosts reply rates significantly) and an Enterprise line of business (e.g., closing a deal with a tier-one consulting firm for 1,500 users, enabling partners to reply 8-13 hours faster) [00:38:28].
Lessons for AI Startups
Taking on Incumbents
While incumbents generally have a greater advantage in the AI cycle than in previous tech cycles, it’s still possible to compete [00:46:36].
- One-Size-Fits-All Weakness: Incumbents (like Outlook or Gmail with billions of users) offer one-size-fits-all solutions, creating opportunities for startups to build better products for underserved, economically powerful market segments [00:48:14].
- Product Speed: Incumbents struggle with product speed, being encumbered by legacy systems [00:49:12]. Superhuman, built entirely in JavaScript, achieved instantaneous response and search times, a difficult feat for older client-server applications [00:49:17].
- Design and Organization: Conway’s Law often means incumbent product design reflects internal organizational structure rather than optimal user experience [00:50:06]. Startups can build more cohesive and user-centric designs [00:51:03].
Investment Perspective
- Bearish on De Novo AI: It’s difficult to build a new application from scratch while also making it “amazing with AI,” effectively running two companies [00:40:55].
- Bullish on Distribution: Investment interest is high for founders with a distribution advantage (contacts, viral loops) or a strong track record of retaining users; AI can be a “hook” but is often incidental to the core distribution capability [00:41:42].
- Market Size: Superhuman’s market size (1 billion professionals spending 3+ hours/day on email) made it an easy investment case, demonstrating how serving a small segment of a massive market can still lead to a multi-billion dollar company [00:54:15].
The Future of Work and AI
Overhyped vs. Underhyped AI
The idea that AI is coming for our jobs is both overhyped and underhyped [00:55:43].
- Underhyped: The speed at which AI will impact certain entry-level roles (e.g., customer support, sales) is underestimated [00:56:02]. An AI that is 80% as good as a human but 10% of the cost will be adopted en masse [00:56:13].
- Overhyped: The notion that AI only targets these roles and won’t affect higher-level jobs (including CEO roles) is overhyped [00:56:55]. AI will eventually be able to perform most tasks better than humans [00:57:03].
This shift will redefine “work.” Just as people from a thousand years ago would view modern jobs as “playing,” future generations might view our current work as leisure [00:57:08]. Humans are hardwired to work and engage in status games, meaning “leisure” activities in the future will still feel like work to them, even if they appear silly to future observers [00:57:58].
The Ideal AI-Enabled Workplace
If building a generative AI product from scratch, Superhuman would target products like Slack or Microsoft Teams [00:58:53].
- Critique of Current Chat Tools: Slack lacks structured conversations (no ordered threads by last reply, no subject lines unless manual), done/snooze features, or the ability to assign/share threads [00:59:29]. This leads to overwhelming inboxes and disorganization [01:00:00].
- Vision for the Future: Combine the best of email (structured conversations, subject lines, archiving, inbox zero, assignability) with chat [01:01:18]. The goal is a workplace experience designed for thoughtful asynchronous discussion with choice of modality, allowing users to legitimately run a company without being overwhelmed [01:02:26].
- Future of Work Agents: This future implies an agent that understands user priorities, calendar, and preferences across all communication platforms to organize tasks and reduce anxiety [01:02:21].
Superhuman’s mission is to help professionals feel happier, more productive, and closer to achieving their potential, with email as the vehicle for this broader vision [01:03:43].