Applied AI: Beyond The Algorithms

This article was originally published by Chip Hazard on HazardLights

One of the primary areas of focus for Flybridge over the years has been to be the first institutional investor behind companies looking to transform the enterprise technology landscape with modern software. Given the explosion in the volume of data being generated globally, this theme has led to investments in companies such as MongoDB (databases) and Nasuni (storage) that operate at the data infrastructure layer of the enterprise tech stack.

More recently, we have been investing in further advances in data management, analytics, machine learning, and artificial intelligence. While the potential for artificial intelligence has been written about extensively, what is less well understood is that the algorithms and underlying tools are only a fraction of the value and are unlikely to be a source of long-term differentiation. Fully realizing the power of AI requires a deep understanding of the domain and the specific workflows that AI will seek to improve and optimize. In other words, the application layer of AI ultimately drives the business value. And we believe the window of opportunity for the AI application layer is now.

This evolution from platform to applications is not an uncommon one: when a new technology platform is immature and not well understood, there is a lot of room for innovation at the underlying technology layer, but as the platforms mature the application layer is where value accrues. For example, in the PC era, once Windows and its associated tools were well developed, apps accumulated a huge amount of value; and similarly in the early Internet era, once the browser, server, and app server infrastructure were well established, apps became super valuable.

This “Applied AI” investment thesis resulted in four new investments by Flybridge in 2018: Aiera, which is using AI to drive fundamental equity analysis; Kebotix, which is using AI to discover and create advanced chemicals and materials; Looka, which is an AI-powered graphic design platform; and Proscia, an AI-powered digital pathology solution. We also made follow-on investments in our existing, scaling, Applied AI portfolio companies Bitsight, Bowery Farming, Datalogue, and DataXu.

Given the breadth of this emerging portfolio, we thought it would be helpful to expand on what we look for in Applied AI companies and some of the keys to success in building a company in this exciting field. We believe that it is important to:

1. Go Old School. Opportunities for Applied AI companies lie outside of the markets targeted by traditional web-scale companies. As Andrew Ng recently observed, “a lot of the stories to be told next year [2019] will be in AI applications outside the software industry”. Many of these markets, such as Real Estate, Finance, Healthcare, Oil & Gas, Agriculture, Manufacturing, and Logistics, have the advantages of being A) extremely large, B) where innovative AI driven approaches can drive massive levels of improvement versus the status quo, C) where potential customers may not be able to access AI talent on their own, such that build versus buy is less attractive, and D) not being an area of focus of the Googles and Baidus of the world where their massive troves of data can be a source of significant competitive advantage.

2. Combine Talents. The most successful teams in Applied AI will have a unique combination of an understanding of the domain and the technical capabilities to realize the vision. Even more specifically, it is doubly helpful if one of the founders was formerly a practitioner in the field. For example, Dawson Whitfield, the founder of LogoJoy, was previously a top-notch graphic designer himself; Ken Sena, was a top-ranked equity analyst before founding Aiera, and Kebotix co-founder Professor Alan Aspuru-Guzik holds a Ph.D. in Chemistry and is a leader in the field of computational chemistry. In other words, AI experts will do better when teamed up with someone that comes from the field in which they are seeking to operate.

3. Drive continued technical innovation. Given how quickly the field is advancing, a deeply technical co-founder who is up to speed on, and willing to continually learn about the latest advances in AI, and see the application of new approaches to the problems their company is seeking to solve is essential. Whether it is “few-shot” learning approaches, ensemble models, GANs, CNNs, transfer learning, explainability, and a myriad of other developing techniques, knowing and understanding the strengths, weaknesses, and applicability of different approaches is critical. We often see the domain expert mentioned in point 2) renting or borrowing their AI expertise in the form of advisors and part-time experts, but this approach is not good enough given the need to have a tight feedback loop between market-driven customer needs and the AI-driven technology insights and art of the possible.

4. Create Data Network Effects. The most successful companies will have a clear understanding and angle on how to start and continue to spin the data network effects flywheel. Generally, this requires having access to initial datasets that can begin the model building process, and a well thought out and focused strategy on how to increase the quantities of data available for analysis. The initial data sets, which Proscia refers to as “inorganic data”, might be acquired, and are used to overcome the cold start problem of training a new model from scratch. In contrast, “organic data” that comes from the ongoing use of the platform can help hone and refine the algorithms over time. Taken together, this means the cost of data should decline over time because organic data is typically free (or even negative if you can get users to pay you for the service delivered while the data is collected). In the pursuit of data, it is important to remember that the sheer volume is not always inherently better. Yes, size matters, but quality matters more. The data should reflect the real-world variance and dimensionality of the problem space and a data strategy should incorporate input from domain experts to identify the logical gaps that exist and how to fill those gaps. Further, when assessing a technical team (per point 3), we believe it is important that they know how to build an AI infrastructure that can be monitored for changing performance and updated accordingly as the scale and scope of the datasets increase. For example, when Aiera first started making buy-sell calls on stocks, they only did so on 16 companies based on a model that analyzed 10,000 documents a week from 300 data sources. Today, they cover nearly 2,000 securities with a model that analyzes 500,000 documents a week from 22,000 data sources. Perhaps not surprisingly, the accuracy, breadth, and duration of their buy-sell calls increased significantly over this time.

5. Absorb The Algorithm. The specific algorithms and AI techniques themselves are not the sources of defensible value so the most successful AI companies will instantiate the models into a broader application that fits into customer’s broadly defined workflows. We call this “Absorbable AI”, which means customers can incorporate the AI into their business and realize the benefits of the operational insights. Successful AI Applications need to not only explain why the model is generating certain results (and, importantly, explainability also helps understand and highlight bias such as gender-based or racial bias), but also integrate into the customer’s business in a logical and systematic way. It’s also important to manage expectations along the way, as many Applied AI use cases falter based on overselling the potential and customers expecting too much when in reality improvements come incrementally over time. These application and process skills are often found coming from the more traditional application development space in fields such as UX, visualization, workflow/BPM, and integrations.

6. Craft the Business Model. Thinking through the business model of a company is critically important. Depending on the domain, the openness of customers to new approaches, their willingness to pay for innovation, and the scale of the level of AI being incorporated, the best way to realize the value of an Applied AI company could be by selling an application that makes human work more efficient and accurate–an end-to-end automation stack that replaces humans–or it could be by selling a complete product. For example, our indoor farming portfolio company, Bowery, leverages a significant amount of AI to drive efficiency and quality in their operation, but they decided the best way to realize the value of that AI was to sell incredibly tasty, locally grown, pesticide-free green vegetables versus selling an AI-powered Farm Operating System to other growers. A similar example would be Tesla, where the vision is the sell a complete, AI-powered, autonomous vehicle, as opposed to say Cruise, which chose to realize the value in selling the application (and the company) to other automobile producers.

Points 5 and 6 can be better visualized in the following matrix:

With the continued explosion of data availability, advances in AI techniques, and the accessibility and performance of computing (GPU) cycles, we believe the trend of AI as the next great application enabler will continue for some time, and we look forward to finding more Applied AI companies with passionate domain experts and technical founders to invest behind in the coming year.

Thanks to my partner Jeff Bussgang, our advisor Harini Suresh, David West of Proscia, and Bryan Healey of Aiera for their input and insights in developing these thoughts.