How Superhuman built its AI email assistant in 3 months

Every software company rushed to bolt generative AI onto their products this year, but few actually admitted how messy the process was behind closed doors.

Superhuman founder Rahul Vohra just pulled back the curtain on how his team shipped Superhuman AI in under three months, starting from zero in-house LLM expertise to beating massive competitors to market.

Going rogue with Theta mode

Standard startup roadmaps broke down the moment ChatGPT took off. Superhuman typically prioritized features by calculating user value against business revenue. With LLMs, neither metric had usable data.

Instead of acting as a fast follower, the leadership chose to plunge ahead. CEO Rahul Vohra and incoming president Paul Teyssier launched what they call Theta mode, a drastic shift from their normal operations.

In standard mode, autonomous squads work with minimal executive involvement. In Theta mode, an executive embeds directly into the core team as an individual contributor. Vohra embedded himself alongside a lone designer, a marketer, and a handful of engineers, bypassing almost every company process.

To keep momentum from stalling, they used external psychology. Vohra took a scheduled keynote at Web Summit Rio on May 3rd and asked organizers to turn it into a live press conference for the AI tool, giving the team a hard, unmovable deadline.

Killing preemptive AI for on-demand prompts

When deciding what to build, the team consciously rejected automated, background features. They skipped automatic inbox summarization and auto-drafted emails.

Automated features carry massive API costs and zero margin for error. If an automatic summary hallucinates or a ghost draft misses context, users immediately turn the toggle off forever.

Instead, Superhuman focused strictly on on-demand actions. The AI only works when explicitly summoned to reply, summarize a specific thread, edit tone, or translate a draft in the user voice.

This approach capped token costs strictly to user intent and allowed the product team to iterate safely in real time. They even decoupled marketing from engineering, animating Figma mockups in After Effects to have launch videos ready without waiting for final code.

Why it matters

For knowledge workers drowning in unread threads, on-demand AI drafting tackles the friction of the blank page. Being able to spit out a tailored reply from a few bullet points turns a ten-minute chore into a ten-second review.

Across the industry, this playbook offers a masterclass for incumbent apps racing against platform giants. When you lack deep machine learning research teams, execution speed and UX integration are your only real defense.

We think Superhuman made the right call by resisting the temptation to build fully autonomous inbox agents. By restricting the AI to on-demand triggers, they protected their unit economics while avoiding the embarrassing hallucinations that kill user trust.

However, the risk is clear: prompt engineering and polished interfaces are easy for big tech to copy. Unless Superhuman can leverage deep personal inbox context that general models cannot reach, staying ahead on speed alone will get harder with every update from Google and Microsoft.