What I’ve learned from building an AI/ML accelerator from the ground-up
I’m experienced with accelerator programs. Before Zeroth, I was at Techstars, first in New York and then starting the first international Techstars program in London alongside Jon Bradford and Jess Williamson. Through that experience, I worked with over 50 startups, managed over 120 mentors, and hired over 40 Associates and Hackstars. I also helped out and knew other accelerators well, including Spark Labs, Startup Wise Guys, 500, Seedcamp, Eleven, ERA, How to Web, ChinaAccelerator, Oxygen and others.
I went through an accelerator withdrawal phase. After Techstars I was a Venture Partner at Mind Fund, where I focused on working with later stage companies. I increasingly became interested in Artificial Intelligence and Machine Learning through my companies Aire, Lingvist, and Weave. But I couldn’t shake off my interest in working with early stage companies, especially those at the very infancy. And that’s when I married those two interests to form a new vision: to deconstruct, fund, and hyper-accelerate the building of AI/ML startups.
Every accelerator has to solve for the following areas. These areas are pertinent to any investment firm, whether it be tech or not, but for simplicity I’ll only address accelerators:
- Brand: Brand is the precursor to deal flow; if you have a great brand, there are a few areas where deal sourcing is improved: more inbound deals come your way and a higher conversion of those inbounds to completed deals.
Coming from Techstars, the brand issue was very much solved. When you’re building an accelerator from scratch, a brand has to be formed.
- Deal Flow Process: The process to actually find companies
- Operations Process: The team also had to run operations optimally.
- Due Diligence Process: How to perform due diligence on startups.
An accelerator brand is comprised of its community of portfolio companies. But without a portfolio to start from, an accelerator has to build its brand on the strength of its team and the companies that the team has been involved with.
That’s why we assembled one of the best lineups of AI/ML entrepreneurs and venture capitalists, including Jaan Tallinn cofounder of Skype, Nathan Benaich from Playfair Capital, Frank Meehan from Spark Labs, Jung Hee Ryu of FuturePlay, Daniel Chu from Microsoft Cortana, Thomas Stone from Prediction.io, Azeem Azhar from Exponential View, and others. In aggregate, this team has worked with companies like Google Deepmind, Siri, Mapillary, Vicarious, Seldon, Weave, and other important AI companies.
Deal Sourcing Operations
The success of an accelerator relies on its ability to attract companies, or its deal sourcing. Every accelerator has its own sourcing tactics depending on the team’s value proposition, personalities, geographic focus, market, thesis, and/or stage. As time passes, an accelerator’s sourcing tactics may change as well to adjust to the market.
The framework that’s worked so far is a two-pronged approach:
- Scaled Sourcing: Sourcing through actions or events that influence a large number of startups at a single point in time
- Referral Sourcing: Sourcing through direct referrals from other investors or community leaders
The accelerator operations has to run efficiently to optimize for the greatest amount of value creation with every unit of effort. Over time, this ratio should increase with tweaks to the model, but some accelerators also increase value creation via adding more programs.
Some of the things we do:
- Rely on the aptitude of 1: In the beginning, we double-teamed trips and interviews as a process of getting to know each other. Now that everyone is competent, we rely on the power of each individual person to make impact. Emphasizing individual impact forces responsibility, decision-making, and action.
- Batching: We batch process almost everything we do, from recruitment, interviews, value-creation, and fundraising. We offer a fair investment deal and state our terms up front so that we waste no time in negotiating terms, and spend more time on value-creating actions.
- Layering: We layer on top of existing investment, social, or physical, or other networks. No need to waste time recreating networks if they are already there.
Due Diligence Process
How an accelerator selects the best startups for its cohort is a combination of rationality and gut-feel. The rational part is gathering and processing the numerous data points and feedback associated with a startup. The gut-feel is that decision-making point, after processing all the data points, on which startup to pick.
For us, the rational part is:
- Documenting the candidate startups: We use Streak to annotate startups that we talk to, but we’ve heard good feedback about ProsperWorks and Hubspot.
- Sustaining the conversation with candidate startups over a period of time: we constantly ask for micro-updates, whether it’s through email, WhatsApp, Messenger, or any other mediums. It gives us a picture of progression over time, not just at one single point in time.
- Tapping the collective due diligence intelligence: we ask the extended Zeroth team to interview and give feedback on candidates.
The gut feel decision-making part we have yet to do, but some of the things I’ve learned is:
- Decide based on being able to make an impact and add value
- Decide based on non-obvious competitive advantages
- Decide based on good teams
Jon Bradford — maybe you actually did know what you were doing after all.
Thanks to Jon Bradford, Paul Smith, Mike Reiner, Vincent Jacobs, Hussein Kanji, Nathan Benaich, Bill Earner, Rui Ma, and Tytus Michalski for ideas and thoughts.