| Episode | Status |
|---|---|
| Episode | Status |
|---|---|
From applied cryptography and offensive security in France’s defense industry to optimizing nuclear submarine workflows, then selling his e-signature startup to Docusign (https://www.docusign.com/comp...
Loïc Houssier, CTO of Superhuman Mail, discusses his journey from French defense cryptography to building AI-powered email productivity tools. The conversation covers Superhuman's AI strategy focused on speed and quality over flashy features, their agentic search framework using multiple models, and how they're achieving 6 PRs per engineer per week with AI tooling. Key insights include their approach to agent laziness, eval strategies across multiple dimensions, and the vision for email as a conversational interface rather than a to-do list.
Loïc's unconventional path from applied cryptography in France's defense industry to process improvement on nuclear submarines, then through DocuSign acquisition to Superhuman. He explains how leaving technical work to become a financial controller taught him to drive change through questioning rather than technical authority.
Addressing the common meme about DocuSign's headcount, Loïc breaks down the actual complexity: separate stacks for EU compliance, FedRAMP, data residency requirements across countries, and even physical security appliances that self-destruct keys if tampered with. Different markets like Japan require entirely different products (hanko stamps vs signatures).
Superhuman's AI strategy focuses on productivity acceleration without adding latency. Features include auto-labeling emails, automatic follow-up detection with pre-drafted responses, and detecting when to CC executive admins. The key principle: AI must improve productivity, not just add flashy features.
The Ask AI feature enables semantic search across entire email history with agentic capabilities. Loïc uses it to summarize 30-40 Substack subscriptions weekly and find specific details from years-old conversations. The system handles complex queries like finding a specific wood type discussed with a contractor 5 years ago.
Superhuman spent significant effort combating 'agent laziness' - when AI asks for user confirmation instead of just executing. Example: asking for 15 minutes on calendar should just book it, not present options. Different models excel at different tasks, with Claude being particularly good at reducing handoffs.
For queries that could match hundreds of emails, Superhuman implements pagination search - semantically searching 40 emails at a time, then expanding proximity until finding the answer. The system uses TurboPuffer for vector storage and maintains local cache for offline functionality, with everything under 100ms response time.
Superhuman structures evals around specific problem dimensions rather than generic query-answer pairs. Dimensions include agent handoff, deep search (needle in haystack), date handling, and code execution for aggregations. Each dimension has canonical queries that must pass before shipping.
Superhuman uses multiple inference providers (BaseTen, Fireworks, Together) and models (OpenAI, Claude, Gemini) based on task requirements. BaseTen's fixed-capacity pricing provides better cost predictability than token-based pricing, crucial for financial planning in a pre-IPO company (Grammarly).
Loïc observes his kids communicate entirely through voice on social platforms, never typing. He questions whether writing itself is becoming obsolete now that everything can be stored as video/audio. This suggests email UX may shift from text-based to-do lists to conversational voice interfaces where CEOs record messages instead of typing.
Despite the appeal of knowledge graphs for entity relationships, Loïc argues that meaningful entity taxonomies are highly subjective per user. Even in Obsidian, no two users' graphs are similar. The compute cost of rebuilding graphs with each new email and unclear productivity gains make this a problem better solved by specialized infrastructure companies.
Superhuman went from free-for-all AI tool experimentation in Q1 to measuring impact in Q2-Q3. They track AI usage via PR labels: 80% of engineers use AI, 90% find it productive. Engineer throughput increased from 4 PRs/week (Q1) to 6 PRs/week (Q3), though AI is only one factor alongside better technical strategy and organization.
With technical moats eroding (startups can replicate features in 2 weeks), Superhuman prioritizes engineers who obsess over user experience and product quality. They hire remotely across Americas, looking for engineers who care about latency impact on end users even if they're backend-focused. Loïc believes AI will separate good engineers from bad ones faster.
The Future of Email: Superhuman CTO on Your Inbox As the Real AI Agent (Not ChatGPT) — Loïc Houssier
Ask me anything about this podcast episode...
Try asking: