Being first to bring a truly intelligent vertical AI assistant to market

In this series, professionals discuss their experiences accomplishing something for the first time. Read their stories here, then write your own using #IWasTheFirst in the body of the post. 

Crowing about being first can be a sign of hubris. (There’s also something called first-mover disadvantage.) However, I feel comfortable in suggesting that x.ai brought the first truly intelligent vertical AI assistant to market. Our AI agent, which operates under the names Amy and Andrew Ingram, schedules meetings for you via email. (To be fair, we do have some company in this space, but none who bet the farm, like us, on the idea of a fully autonomous machine driven meeting scheduling agent.)

By nature I’m competitive, but that doesn’t explain how we’ve achieved this feat. The key experiences that have enabled us to be first (whether or not we win in the end) fall into two categories: the successes and the failures.

On the success side of the ledger, I had built a number of other ventures with exits that made our investors happy. Broadly speaking, this meant I had learned how to identify a genuine pain point to solve and then how to convince very smart people to join me to translate a product vision into reality. And it also meant that I learned how to sell my vision to investors to have the capital to hire those talented folks. Key to all of this has been extreme focus and its corollary, a pretty strong distaste for the pivot.

In fact, on the company building side, I’ve been honing my game for 20 years, starting straight out of college. I’m a big believer that you can learn entrepreneurship, and you can get better at it over time, which is a somewhat contrarian view.

On the product side, all of my ventures have been in the data and analytics spaces. Having built companies around data in the past gave me and the other founders the confidence we could take on something as ambitious as building a fully autonomous AI personal assistant, even though no one had done so before. And it also meant that we understood how to approach the problem, which was a very staged exercise in data capture and analysis, expanding the data set slowly over months and years. Our seed round underlines this staged approach. We raised it solely to answer the question: is meeting scheduling a tractable problem, could it be done by a machine?

But the failures were just as important in setting x.ai up for this success. June 16th 2014 stands out as a complete operational nightmare. We added the first real batch of beta users (who did not really know what an AI agent was); we hired our first full time AI Trainer, Laney (we did not have any written annotations guidelines); we launched the annotation console (which was buggy)—ALL on the same day. And we had not properly calculated the time to annotate data. That was, in short, a data annotation trifecta from hell. It took a week of 16 hour days to catch up. But we did. And in the process we created a set of super happy beta users, good data, and a solid annotation environment. Laney is still with us, by the way! :-)

The most critical decision that has enabled us to launch a vertical AI personal assistant was to invest heavily in both the “read” and “write” dimensions of the system.

The first point is straightforward: our biggest business unit is Data Science, and this has been true from the inception of the company. Given how ambiguous human communication tends to be, understanding scheduling related conversations is, rightly, a massive NLP challenge.

But we’ve invested equally in the write (or NLG) dimension and ended up creating an entirely new role, the AI Interaction Designer. Our team of AI Interaction Designers have worked hard to humanize Amy and Andrew. This human touch has elicited a ton of delight, which we can see in love notes from happy customers.

Those customer reactions are the ones I care most about. Colleagues, friends and family have generally been supportive, but not all appreciate the steep road we’ve climbed to get here. The team has moved mountains, technically speaking, in order to create an AI personal assistant who schedules meetings better than you do and who is routinely mistaken for a human.

While we may be the first AI vertical agent in market, I believe we’re a sign of the near future. I think we’re entering a new software delivery paradigm in which Vertical AI (or bots) will in many cases replace the need for an App. I personally don’t have to do much to bring that vision forth besides continue the work we’re doing to make Amy and Andrew even smarter; entrepreneurs are working on many vertical AI agents, and VCs are funding them.

I do think it helps for us to tell our story and to report back from the trenches about the challenges, occasional set backs, and small victories we have weekly. No one has built a fully autonomous AI personal assistant before. We want to share as much as we can about what we’re learning along the way.