X.ai

Dennis Mortensen had just sold his predictive analytics company in 2013 and found himself with something rare for an entrepreneur: time. As he sat down one day and opened his calendar out of curiosity, he counted 1,019 meetings from the prior year. What shocked him more was the number: 672 of those meetings—roughly 65%—had to be rescheduled. After two decades of building startups, he wondered if he'd spend another 20 years dealing with this same administrative burden. The idea of an AI assistant that could handle meeting scheduling seemed like it could genuinely solve a widespread problem.

Instead of immediately jumping into engineering, Mortensen took a disciplined approach to validation. He and his VP of Engineering set up a simple test: they created an email address called "Amy" and played each other's personal assistants for a month to experience the joy (and pain) of having someone handle their scheduling. The experiment worked—he loved not scheduling meetings, but hated doing them for someone else, proving there was real value in the service.

Next, he hired a full-time assistant and offered her services free to 50 friends across tech companies, with the constraint that they could only communicate via email. The feedback was overwhelmingly positive. People loved the service, and more importantly, they couldn't tell whether "Amy" was human or machine—a critical insight for proving the concept was viable.

With this validation complete, Mortensen raised a $2 million seed round with a bold premise: no MVP, no customers, no revenue. Instead, the funding would go entirely to answering a binary question: "Can we actually build this?" This was unusual territory—most founders would bootstrap initial customers and then raise. But with his track record of three successful exits, he could pitch a more dramatic narrative.

The seed money was spent on internal infrastructure, not customer-facing products. The team built annotation guidelines and a labeling console to define every possible "intent" in meeting scheduling conversations: setting up new meetings, canceling, rescheduling, making someone optional or mandatory, extending duration, running late, and countless other scenarios. For an entire year, about 70 people in Manila and 50 engineers in New York systematically labeled data—eventually reaching 32-33 million labeled elements. This wasn't glamorous, but it was foundational: you can't train a machine learning model on data you haven't painstakingly defined and labeled.

By the time Mortensen was ready to launch, he'd essentially proven the technical risk was surmountable. The product launched with a freemium model ($8/month for individuals, $24/seat/month for teams), but the freemium approach initially failed because the product required education—people didn't instinctively know how to use an AI scheduling assistant the way they did with, say, Dropbox or Slack.

The freemium model was abandoned because the product was too novel. Users needed onboarding and support that the freemium model couldn't afford to provide. However, what worked brilliantly was the viral loop built into the product itself. When you CC'd Amy or Andrew (the two AI agents) into an email, everyone on that thread saw them in action. Over half of all new signups came from people who had been exposed to the product through colleagues using it—making the product itself the primary acquisition channel.

Moreover, the bottom-up adoption proved powerful. Individuals at massive enterprises like Uber, Microsoft, VMware, and Pepsi signed up independently, created small pockets of usage within their organizations, and eventually those informal groups became the foundation for enterprise relationships—exactly how Slack scaled in its early days.

By the time of this interview, X.ai had raised over $44 million and was operating at a significant scale. The company had hit a major inflection point: on June 1st (implied to be around the time of the interview), Mortensen announced they would shut down the entire 70-person data labeling operation in Manila. Why? Because the product had become robust enough, and they had enough training data, that future improvements would come from actual product usage—the same way Waymo's self-driving cars improve through real-world miles driven.

The company serves individual knowledge workers with multiple external meetings (particularly sales, recruiting, and similar roles) and has expanded to serve entire enterprises while maintaining a land-and-expand motion where individuals adopt it first and teams follow.