Email startup Front recently announced a $10M Series A capital raise. As always with company announcements there is intriguing information underneath the corporate talk. Front founder Mathilde Collin highlights how the company started as convenient inboxes to managed “collective” email, and that they are now looking to help users deliver AI-assisted email. This perspective made me think of a famous post by Chris Dixon, but with a twist: come from the tool, stay for the AI.

Dixon uses the example of Instagram to illustrate the tool/network dynamic. It started with cool photo filters to improve your pic quality. Then, as more and more users started using it, the network appeared. You would follow you friends and interact with their pictures. Building the network from scratch would have been way more difficult: the filters served as a standalone tool to build the user base. The first users saw enough value in the filters themselves: they didn’t worry about not having any friends on the network.

Same for AI: difficult to show tremendous value from scratch. How good can AI be if it’s not aware of your personal habits? And how to collect that information about you without harassing you with questions? And what users do might differ from what they say they do. Of course, any software application can use data collected from all the existing users to seem smarter when onboarding new users. The Twitter example might say: it’s far from enough. So when the user opens the app for the first time, the AI value is still limited by the lack of user history. To ensure short time to value, users must find happiness somewhere else. What if the app was providing a standalone tool to perform certain actions? Users see the benefit, start using it. From there, its intelligence grows, from usage and pattern recognition. The product sticks because users don’t want to switch (and restart the learning process with another app). And the company builds a defensible data network effect. The challenge here is the velocity of the learning cycle. Machine learning necessitates as many data points as possible to build strong insights and learnings, so interactions must be very frequent. The tool must be used as many times as possible.

Here’s a map of great tools evolving into AI-powered products:


Netflix and Spotify are well-designed tools to consume content. Then they learn your habits and make personalized recommendations. But content catalogs are not so large: human curation could do the trick. Tesla builds great cars to drive and collects massive amount of data about driving habits for their autonomous driving system. Apple builds the best tools of our era and records so much data about habits, location, lifestyle. I wouldn’t be too quick leaving them out of the AI wars. Since the previous era, Microsoft has been also building market-leading tools. But those don’t record enough interactions to produce any relevant AI. I’ve opened Excel more than 10,000 times in my lifetime, and every time feels like the first time. Ten thousand times is a rounding error when it comes to the amount of data necessary to build AI.