Latent Space: The AI Engineer Podcast

Episode Summary: The Future of Email — Superhuman CTO on Your Inbox as the Real AI Agent (Not ChatGPT) with Loïc Houssier

Air Date: December 11, 2025
Guest: Loïc Houssier, CTO of Superhuman
Hosts: Alessio (Kernel Labs), Swix (Latent.Space Editor)

Overview

This episode explores the future of productivity and email through AI, with a deep dive into how Superhuman—a product at the leading edge of email and productivity tooling—is leveraging AI agents to reinvent the email experience. Rather than framing ChatGPT or foundational models as the locus of intelligence, Superhuman is betting on the personal inbox as the real, proactive AI agent. Loïc Houssier, Superhuman’s CTO, discusses building agentic frameworks, offline and privacy-first architecture, the evolution of user experience from typing to voice, the technical stack, and AI’s impact on engineering teams and culture.

Key Discussion Points & Insights

1. Loïc’s Backstory: From Nuclear Submarines to AI-Driven Productivity

• Loïc's career includes a mathematics and security background, experience at DocuSign via acquisition, and time in the French defense sector.
• Emphasized value in curiosity, humility, and learning from non-technical domains, including his “crash course” managing process improvement on real submarines.

  “The fun, like, icebreaker that I give to people sometimes... I spent one year walking around submarines.” — Loïc, [02:00]

2. Superhuman: AI Product Philosophy and Roadmap

• AI is implemented for real productivity acceleration—not gratuitous features or surface sparkle.
• Use-cases include:
  • Auto-labelling emails (e.g., pitch, marketing, executive requests)
  • Email thread summarization
  • Draft detection and proactive reminders for required responses
  • Smart assistant features like “Ask AI” that let users query their entire inbox for summaries, documents, or availability
• Key priority: No added latency, high trust and high-speed UX for demanding users.

  “Our people are pretty high expectation oriented... you cannot add latency. Everything is to improve productivity. So AI included.” — Loïc, [10:46]

Recent Notable AI Features

• ‘Ask AI’ feature is gaining more use as querying via chat becomes normalized.

• Most users’ AI usage is shifting from accidental (triggered by searches) to intentional, thanks to improved agent quality and consumer muscle memory from ChatGPT et al.

  “People are more and more used to the muscle of querying things because of ChatGPT.” — Loïc, [13:58]

• Intelligent agentic frameworks (“agent laziness” problem): How do agents move from just retrieving information to acting proactively—e.g., booking calendar slots based on vague user intent, as a human assistant would?

  “Working on this agent laziness because the handoff they were doing to the user is losing time... things happen faster.” — Loïc, [14:50]

3. Superhuman’s Agentic Architecture: Tools, Evaluation & Models

• Uses multiple small, composable tools (availability detection, people graph, email composer, etc.) rather than one giant “do-everything” agent—each tool is optimized for a specific job.

  “It’s a set of small tools that are used within the agentic framework." — Loïc, [18:37]

• Different LLMs/models are chosen for strengths: Sonet for agent handoff, OpenAI for others, with continual evaluation as new entrants (Gemini, Anthropic) improve.

• The team iterates on evals and canonical queries—testing both the agent’s ability to do “deep search” (needle-in-haystack) and handle complex date/logical queries.

  “You have to understand your data: short emails, long emails... you need to have canonical queries that are targeting [each] dimension” — Loïc, [21:01]

• Quality bar is extremely high due to “high expectations” users; Rahul (Superhuman’s CEO) is often the most aggressive user and bar-setter for new features.

4. Inbox as Platform and Limitations vs. ChatGPT

• Superhuman is positioning itself as a proactive, context-rich agent—contrasted with ChatGPT, which is primarily reactive.

  “We are everywhere and everything. We want to be more proactive compared to ChatGPT that is waiting for you…” — Loïc, [57:26]

• The inbox is framed as “the ultimate private data source” and the ideal place for an AI-powered executive assistant (EA).

• Superhuman’s vision: Become an AI EA that can answer, draft, summarize, and eventually act on behalf of busy professionals—with human-like judgment and memory, but anchored in privacy and user control.

  “That's the vision. Like, you have an EA... We want to do most of that job.” — Loïc, [43:10]

• Limits exist: Human EAs have context and judgment (e.g., relational nuance, current mood), which LLMs still struggle to learn.

5. Technical Deep Dive: Stack, Deployment, and Privacy

• Superhuman is deeply local-first for speed, privacy, and offline capability: most email data, embeddings, and search indexes are synced to user devices.

  “Everything needs to be local... every interaction under 100ms.” — Loïc, [27:47]

• Backend leverages scalable vector stores (like TurboBuffer), hybrid local/cloud architectures, and model inference via providers like Base10, GCP (Gemini), OpenAI, Anthropic.
• Mobile stack previously used Realm (now defunct); local model inference is considered but constrained by device storage and performance budgets.
• The company is exploring local-on-device model partnerships (e.g., with YC-backed startups) to enable privacy-preserving, offline-first LLM uses, but premium model cost is not the top concern for their high-value-user base.

  “We want to solve for offline. Our users want quality and are okay to pay for it.” — Loïc, [30:29]

• Cost optimization is secondary to UX: they start with the most expensive models and optimize after hitting quality benchmarks.

6. AI’s Impact on Team Culture & Productivity at Superhuman

• Superhuman operates with a lean, senior engineering team (~50 engineers, 3 in AI), fully remote, with a culture of autonomy and “free for all” tool experimentation.

  “First thing we've done is cut the red tape. Free for all... You can try anything.” — Loïc, [58:31]

• Massive AI adoption: ~80% of engineers report using AI tools regularly in PRs, mostly with positive impact on productivity and code discovery.
• Measurable throughput improvement: PRs per engineer/week have climbed from 4 → 5 → 6 over the past three quarters.

  “PR per engineer per week... increased quite a lot... It's only a piece of it, but AI is helping.” — Loïc, [62:20]

• AI accelerates both code writing and especially codebase navigation (“discovery”), decreasing engineering onboarding time.

7. Organizational & Industry Reflections: The Future of Coding, Product, and Work

• Even with AI, core software skills matter: AI is an abstraction, but deep understanding (algorithms, systems, tradeoffs) will continue to differentiate great engineers from the rest.

  “AI will separate faster the good engineers from the bad engineers. If you're good... you'll be amazing. If you're poor, AI will make you worse.” — Loïc, [68:29]

• Productivity gains do not threaten well-motivated engineers; satisfaction comes from impact and environment, not mashing out more code.
• “Inbox as Agent” points toward a future where the UI could move towards voice, speech, and agent-driven interactions—the classic list/table paradigm is fading.

  “My kids don't type on their phones. They talk. Maybe next year [your CEO] wants to talk to you [in your inbox].” — Loïc, [38:17]

Notable Quotes & Memorable Moments

• On AI’s role in email productivity

  “We don't care about sparkles... Everything we do is to improve the productivity of people.”
  — Loïc, [10:46]

• On querying your inbox with ‘Ask AI’

  “Now I don’t read [my newsletters]. I auto-archive those and every week I ask my email, ‘tell me about the summary’ of all the Substack that I received this week. What should I pay attention to?”
  — Loïc, [12:41]

• On agent laziness and emulating EAs

  “If I say, ‘find me 15 minutes tomorrow,’ I don’t want the agent to give me options. I want it to do it, like an admin would.”
  — Loïc, [14:50]

• On the ongoing war with ChatGPT & foundation models

  “ChatGPT and Superhuman are competitors. We want to be more proactive... That’s what we fight against.”
  — Loïc, [57:26]

• On shipping and AI’s impact on engineering

  “If you ship more, it doesn’t mean you always need more people. But if you do ship 100 features, you might not have support to help, or marketing to communicate.”
  — Alessio, [63:06]

• On the future UI for communication

  “Now, like, all content can be on YouTube or TikTok. What’s even the need to write? Everything can be vocal. I see kids, they want a TikTok of the article, not the article.”
  — Loïc, [38:33]

Important Timestamps

• Loïc’s background/story: [01:54] – [04:27]
• Superhuman’s AI Product Philosophy: [10:25] – [14:46]
• Ask AI and agentic search: [13:35] – [16:45]
• Tooling/agent design and evals: [18:25] – [20:46]
• Proactive agent vs. ChatGPT: [43:10] – [44:46], [57:26]
• Inbox as the real “agentic platform”: [40:22] – [44:46]
• Privacy, local vs. cloud, and cost models: [27:47] – [31:58]
• Engineering culture, AI adoption: [58:31] – [63:06]
• Reflections on the future of coding/software: [68:12] – [69:29]

Conclusion

Loïc Houssier presents a compelling vision: Your inbox—full of data, context, memory, and personalized patterns—is the place where true AI agents can flourish, far outstripping what current cloud-only chatbots offer. Superhuman is building towards an AI-powered executive assistant, integrating high-quality agentic frameworks, voice and context UI, and privacy-centric, local-first technology. The episode provides technical and cultural insights for AI engineers, product builders, and anyone invested in the future of work, productivity, and interfaces.

Call to Action

• Try Superhuman and give feedback on AI features.
• Superhuman is hiring strong product engineers passionate about UX—anywhere in the Americas ([70:49]).
• Stay updated via latent.space for more AI news, interviews, and show notes.