The Future of Email: Superhuman CTO on Your Inbox As the Real AI Agent (Not ChatGPT) — Loïc Houssier
Summary
Superhuman’s central bet is that the inbox, not a generic chatbot, becomes the proactive AI agent for knowledge work. Loïc Houssier argues that ChatGPT waits for users to supply context, while Superhuman can act from where work already happens: email, Coda’s ingested knowledge, and Grammarly’s cross-application presence. “We are where you work,” making the strategic contest “reactive versus proactive.”
The product roadmap advances through small, high-confidence automations rather than conspicuous “AI things and the sparkles.” Auto-labeling, thread summaries, two-day follow-up reminders, prepared drafts, availability responses, and newsletter synthesis all serve Superhuman’s defining constraint: never slow demanding users down. Ask AI now combats “agent laziness,” paginates through successive groups of roughly 40 emails, and keeps searching until it finds the needle.
Superhuman wants to become an AI executive assistant, but the last 20% of judgment remains the hard part. Its auto-draft system is being tested to learn routine vendor rejections or route applicants to HR, while a scheduling beta inserts three available slots into a draft. Yet “eighty percent” accuracy may still create too much discard work, and calendar data cannot reliably know when a VIP merits disruption or a relationship changed last week.
A three-engineer AI team supports the product through narrow tools, model routing, and unusually demanding end-to-end evals. Separate tools inspect availability, relationships, search, and writing; Sonnet was especially good at agent handoff and avoiding lazy handoffs, while the tested OpenAI model was weaker there and Gemini had recently become another strong option. Calibration includes Rahul’s demand to recover a contractor email naming a table’s wood from five years earlier: “Until we nailed that query, he was not satisfied.”
Quality comes before inference cost, creating room for premium pricing and predictable-capacity vendors. Superhuman starts with “the highest and more expensive model,” then optimizes after the experience works; Baseten’s box-based capacity helps forecast three- and six-month costs better than volatile per-token usage. Some customers say they would pay $200 monthly for the best model rather than accept “half crap,” because one saved executive hour can be worth ten times the subscription.
The operating evidence suggests AI is raising engineering throughput, although Loïc explicitly refuses to assign it all the credit. The data is self-reported: about 80% of people flag PRs; among flagged PRs, roughly 80%–90% involve AI, and about 90% of those uses are positive. Output rose from about four PRs per engineer per week in Q1 to five in Q2 and six in Q3. Superhuman supports this with one-hour budget approval, 24-hour security review, roughly 50 engineers, approximately 100,000 paying users, and a senior remote team averaging about four years’ tenure.
As implementation gets cheaper, product judgment and engineering fundamentals become more valuable—not less. Loïc compares AI coding with the move from assembly to C: another abstraction layer that still requires understanding memory, servers, and the machinery underneath. His categorical call is that AI “will separate faster the good engineers from the bad engineers,” while startups can increasingly copy technology in two weeks and differentiation shifts toward workflow, latency, interface quality, and user empathy.
Deep dive
1. DocuSign taught Loïc that distribution and compliance can outweigh elegant technology
Loïc began in mathematics and nearly pursued a three-year applied-math and cryptography PhD before moving into offensive security. His year around submarines then stripped away the technical authority he had enjoyed as a security researcher. Confronted by specialists in torpedoes, radar, and nuclear engines, he learned to “put my ego in my pocket” and lead through questions—becoming comfortable helping people who understood their fields far better than he did.
His Paris e-signature company was technically strong and becoming a French leader, but DocuSign spoke a different commercial language. The Europeans sold technical capability to CIOs; DocuSign sold business value to heads of HR and other functions. That contrast made partnership, and eventually acquisition, more compelling than continuing alone.
The deal required a six-month carve-out because France’s Ministry of Finance blocked the sale: the company also handled security and authentication work for France’s Department of Defense. Teams, source code, systems, and even data-center infrastructure had to be split and replicated before DocuSign could acquire the signature business. Loïc’s retrospective recommendation is wonderfully blunt: “Crazy. Don’t do that.”
Swyx’s “What do all the people do at DocuSign?” challenge surfaced the hidden scale economics: local data centers, different European signature standards, FedRAMP, country-specific residency, and on-premises appliances whose keys disappear if stolen. Japan required understanding hanko rather than assuming Western signatures; what looks like staffing bloat from outside can be market-access infrastructure.
His broader France lesson is that strong education and engineering are not enough without grit, an entrepreneurial mindset, product-led growth, and global ambition. European startups should consider targeting English and the world from the outset rather than starting with France or Italy.
2. Superhuman adds AI only when it makes demanding users faster
Loïc’s product rule is not to ship “AI things and the sparkles”; every feature must accelerate people who already have unusually high expectations. Latency and interruption are product failures, so intelligence has to appear at the right moment without disturbing Superhuman’s speed.
Early features were deliberately unglamorous: classify pitches and marketing, summarize long threads, and detect messages that still need an answer. After two unanswered days, Superhuman can resurface the thread; the next iteration prepares the follow-up itself, converting “Damn, shoot, yes, I wanted to remind people” into one press of Send.
The same progression handles availability requests and delegation. A draft can suggest copying the appropriate executive assistant, while the emerging scheduling beta inserts usable times. Each step removes another small decision rather than demanding that the user enter a separate AI workflow.
Loïc now auto-archives his roughly 30–40 newsletter subscriptions, then asks on Friday for a synthesis of everything received and what deserves attention. The behavior change matters as much as the feature: consumers have learned from ChatGPT that querying a corpus can be easier than maintaining “a to-do list of emails.”
3. Ask AI is engineered to finish the task, not return choices
Swyx’s experience moved from accidental activation in Superhuman’s search bar to deliberately asking broader, half-formed questions. He now retrieves contracts; Alessio finds phone numbers. Loïc recovered a PowerPoint link from a conference six months earlier, estimating that the contextual search saved approximately 30 minutes.
Loïc identifies “agent laziness” as a specific quality defect. When he asks for 15 minutes tomorrow to review a document, the system should choose and reserve a slot—not offer several options and hand the work back. A capable human administrator hears “Find it, do it,” and the agent is evaluated against that standard.
Deep retrieval cannot fit an entire mailbox into one context window, so Ask AI uses semantic pagination. It might inspect the first 40 plausible emails, reject them, inspect the next 40, and continue widening proximity until the correct result appears—even when hundreds of messages are semantically close.
The orchestration layer is composed of narrow tools for availability, correspondent relationships, drafting, and other actions. The agent first makes a plan, selects the best tool for each step, then executes; there is no single “magic big tool that can do pretty much everything.”
4. Model selection matters less than evals that expose distinct failure modes
Only three engineers work on this AI stack, according to Loïc, but they have iterated across frameworks, models, and internal proxy layers. Sonnet performed especially well on agent handoff and avoiding lazy handoffs, the tested OpenAI model was weaker there, and Gemini had just entered the comparison as another promising option.
Superhuman’s initial eval was the naive “query, answer, query, answer” pattern. It evolved into canonical query sets organized by dimensions: agent handoff, deep search across a “shit ton of emails,” date interpretation, and other recurring weaknesses. Each end-to-end result is scored along those dimensions, not merely judged as generically good or bad.
Rahul supplies adversarial reality: he receives perhaps 500–1,000 emails daily and demanded retrieval of a five-year-old contractor discussion identifying the wood used for a table. Date questions add another failure class because models struggle with expressions such as “last quarter” relative to today.
Loïc distinguishes ordinary reliability from the “grade” Superhuman users expect. A Toyota can be good quality, but an Audi or Porsche buyer expects something different; Superhuman’s high-expectation user forces disproportionate investment in details that a lower-priced, lower-intensity product might tolerate.
5. The inbox is becoming a computable system of record
Swyx’s provocative query—“How much time do I spend in Waymo last month?”—turns receipts into an accounting database. The agent selects Waymo emails, extracts trip durations, and aggregates them through generated code because “LLMs are not good at math.” Superhuman was discussing broader code-execution components with Anthropic.
Superhuman still relies on Gmail and Outlook for infrastructure such as spam detection rather than becoming another mail server. Loïc sees little user value in rebuilding capabilities already embedded in customers’ company systems: “If we can just plug in and make that better,” the layer above the mailbox is enough.
Speed requires a local copy because every interaction is meant to land under 100 milliseconds. Installation downloads the latest 30 days for offline operation, then retains subsequent history; a two-year customer may therefore carry two years of mail locally. Realm was used on mobile, but it is now somewhat sunsetted, leaving the team considering alternatives such as SQLite.
AI retrieval adds embeddings and hybrid search over roughly five years of history in Turbopuffer. Inference is distributed across open models and providers including OpenAI, Anthropic, Gemini, and Baseten-hosted Llama or BERT classifiers—the chosen route depends on the use case and current model quality.
6. Offline inference is a quality problem before it is a cost problem
On-device startups often pitch lower inference expense, but Loïc says Superhuman users “want quality, and they are okay to pay for that quality.” His interest is offline semantic search: the product already works without a network, while AI capabilities become materially weaker when remote inference disappears.
The constraint is footprint as much as intelligence. Superhuman already consumes storage and memory because mail is cached locally; Swyx’s DeepSeek-V3.1 example—roughly 600 billion parameters—illustrates why simply shipping a frontier-scale model is impossible. Any local capability must fit without worsening existing complaints about application size.
Superhuman would prefer to own the end-to-end experience, just as its mobile applications use Swift and Kotlin rather than React Native to feel native. Yet Loïc would “love the device provider to be better”; iOS work had been underwhelming so far, leaving room for specialized companies to supply the missing layer.
7. Premium AI economics favor proving value before optimizing tokens
Loïc’s sequencing is categorical: begin with the best, most expensive model, establish the required quality, and only then reduce cost. “If it’s successful, it’s great, even if it’s expensive”—success creates the volume that makes optimization worthwhile instead of burdening an unproven feature with premature constraints.
Baseten’s fixed-capacity, box-based model makes expenses easier to control than purely token-metered serverless inference. Under a larger, “pre-IPO-ish” umbrella, Superhuman increasingly needs credible three- and six-month projections; fixed capacity gives the CFO a bounded bracket even as adoption varies.
Token forecasts remain messy because Superhuman processes emails ranging from tiny replies to long threads. Teams must estimate median lengths, average consumption, and adoption, with “always some magic” left in the model. Swyx reframed enterprise buying as price per trillion tokens rather than the solo developer’s price per million.
Alessio asked whether drafting every incoming email would destroy a $40 subscription’s economics. Loïc’s answer was value-based: some customers explicitly request the best model and say they are ready to pay $200 per month, since one hour of a CEO’s or VC’s time may exceed the subscription tenfold.
8. The AI executive assistant is blocked by judgment, not drafting
Internal auto-drafting is still being tested and is learning repetitive behavior: Loïc politely declines the hundreds of AI-tool pitches he receives and routes job applicants toward HR. Swyx argued that combining established snippets with AI personalization may solve the practical “last mile” without anything resembling AGI.
Coverage is the unresolved product threshold. If 80% of drafts are useful but users must discard 20%, the automation may create irritation instead of leverage; Loïc asks whether the acceptable split is 90/10 or 80/20. The relevant metric is saved attention, not simply generated text.
Scheduling reveals the deeper gap. Superhuman can draft three lunch times, but Swyx wants no involvement at all; a human EA knows which VIP justifies moving a busy calendar and may know that last week’s conversation changed a relationship. Interaction counts alone cannot encode that judgment.
Loïc still calls the AI EA “the goal,” with human judgment remaining in the loop, perhaps via the user’s EA. Swyx proposed acquiring a human-assistant business to observe workflows and generate proprietary training data; Loïc agreed it might be the best learning path, while calling the operational commitment “intense.”
9. Email may move from written rows to spoken, ambient communication
Loïc’s three children talk on their phones, use WhatsApp with family, and communicate via Snapchat, TikTok, or Instagram; he says everything is becoming vocal and that they increasingly want a video about an article rather than the article. His larger provocation is that humans wrote because spoken stories lacked storage, but ubiquitous audio and video weaken that historical constraint: “What’s even the need to write?”
Applied to email, the question is whether users will still type in a year. Rahul might announce a feature by speaking, while recipients hear his actual voice during a commute—not merely prose imitating his “voice and tone.” The opportunity is to redesign communication rather than attach a chatbot to today’s table of message rows.
The browser debate followed the same logic. Swyx argued that owning the browser could maximize context; Loïc instead expects browsers to become thinner, perhaps disappearing into the OS. Raycast already replaces his bookmarks and navigation, leaving the browser as little more than a web view plus local storage and extensions.
10. Cross-application context is Superhuman’s answer to ChatGPT
Coda Packs are integrations that ingest data, and Superhuman technically has an ingestion pipeline for aggregating company knowledge. Grammarly can know that a user is in Google Docs, crafting a LinkedIn post, or moving among Jira, Salesforce, and LinkedIn before writing an email—though Loïc explicitly notes that knowing this does not mean Grammarly uses all of that data.
Once those assets converge, the system can augment a message with the context of the work that immediately preceded it. The acquisition had happened only three months earlier, so Loïc described this convergence as a future advantage rather than a completed capability.
ChatGPT, by contrast, does not inherently know which application the user just left; it waits for pasted context. Loïc therefore describes OpenAI and Superhuman plainly as competitors. Superhuman’s intended advantage is ubiquity: “We can be more proactive compared to ChatGPT that is waiting for you.”
Swyx challenged Loïc’s suggestion that knowledge-graph systems were “not there yet,” since Superhuman had not tried them; Loïc immediately conceded the point. His substantive concern is taxonomy: people and companies are obvious, but projects, tasks, initiatives, hierarchies, and domain-specific features vary radically among users.
An Obsidian user may build an entirely subjective graph, making a universal productivity ontology difficult. Loïc expects Superhuman to consume a specialized graph engine rather than build it: continuously remembering everything resembles “Jarvis” and demands heavy recomputation whenever new mail or interaction data arrives.
11. AI raises throughput, but organizational quality determines whether it compounds
In Q1, Loïc removed purchasing friction: experimental tools received budget approval within an hour and security review within 24 hours. Adoption surged, performing especially well on front-end work but less well in the Go backend and Swift at that time; v0 won an internal free market for interactive product prototypes.
Q2 focused on measurement and use-case discovery. Exploring an unfamiliar code area fell from roughly a day to about 30 minutes with Claude Code; in his example, he also used Warp. Engineers labeled each PR for AI use and whether it helped; approximately 80% reported, around 80%–90% of those used AI, and roughly 90% of reported uses were positive.
Throughput rose from about four PRs per engineer per week in Q1 to five in Q2 and six in Q3. Loïc stresses that AI is only one cause alongside technical strategy, organizational design, and clearer priorities—and that PR count is merely a throughput proxy, not a complete productivity measure.
Superhuman remains lean: roughly 50 engineers serving perhaps 100,000 paying users, with average tenure near four years. Its three-person AI team spans Patagonia and Canada; after the acquisition, the mail business operates as a “compound startup” with its own P&L, Rahul still leading, and Shishir’s executive team effectively serving as its board.
Alessio’s pushback—worth keeping—is that faster engineering can overwhelm marketing, support, and product coherence, while more shipping can mean more bugs. Loïc therefore resists a reflexive “we can do more, let’s do more” cycle even as the larger company supplies capacity for expansion.
On labor, Loïc rejects the idea that moving from four to six PRs should automatically mean 50% higher compensation. Fair pay remains necessary, but senior engineers also derive satisfaction from shipping useful work in a low-friction team: “I can basically be the best version of myself.”
His closing call is unequivocal: people should still learn to code. AI is like the transition from assembly to C, not the abolition of underlying systems knowledge; good engineers become “amazing,” while “poor, lazy” engineers can mistake generated output for understanding and become worse.
The commercial moat consequently shifts toward product craft. If a startup can approximate the technology in two weeks, differentiation comes from understanding the user, workflow, visual quality, and latency. Superhuman is hiring remote product-minded engineers—including backend engineers who care about latency because users feel it—across the Americas.