How to Run a Successful AI Startup

Credits: Thank you to our sponsors, Madrona Venture Group and Madrona Venture Labs, the AWIP Team, Emily Kruger, Bridget Hughes, Sadaf Hasan, and Brian Xue, and our speaker, Jay Bartot, CTO of Madrona Venture Labs (MVL).

In the past 5 years, artificial intelligence and machine learning have evolved into some of the most-watched technologies, with AI and ML-based startups like UiPath, OpenAI, and OctoML gaining significant traction. Jay Bartot, the Chief Technology Officer of Seattle-based startup accelerator Madrona Venture Labs, says, “We are in a golden era of machine learning development due to massive amounts of accumulated data, the increasing availability of compute power, and the reemergence of Deep Learning.”

On Thursday, January 7th, 2020, Advancing Women in Product (AWIP), Madrona Venture Group, and Madrona Venture Labs hosted a workshop on “How to Run a Successful AI Startup.” This workshop is the first of a series focused on successful AI/ML startups. During our workshop, Jay focused primarily on how to identify, operationalize, and productize an AI startup’s most important asset: its data. Jay, who has co-founded a number of successful venture-funded data startups (including the Hulu-acquired Vhoto, Alliance Health Networks-acquired Medify, and Microsoft-acquired Farecast), presented a wealth of real-world examples and anecdotes from both his own startup experience and his role as CTO of MVL. The event was moderated by Meg Quintero, Senior Product Lead at FLEXE, and included an interactive Q&A session with the audience.

Jay Bartot (left) led the AWIP, Madrona Venture Group, and Madrona Venture Labs workshop on How to Run a Successful AI Startup on Thursday, Jan. 7, 2021. Photo credits: Jay Bartot, AWIP.

Jay’s workshop is the first of a series for AWIP members interested in starting or joining a technology startup, featuring best practices from industry veterans and successful entrepreneurs. The series and programming has been inspired by the Future Founders community founded by Sudip Chakrabarti, an Investing Partner at Madrona Ventures, to facilitate the exchange of ideas across Seattle-based tech and product leaders. As a new AWIP team member and member of Madrona’s Future Founders community, I organized the launch of this series and look forward to informing and empowering your careers with this exciting speaker series.

Jay opened with an overview of the evolution of AI and ML technologies, their current applications in household products, and the types of machine learning. Jay also presented a way to categorize ML companies into horizontal companies, which provide ML/DS tooling and platforms to a technical audience, and vertical companies, which offer an end-to-end product. He observed the wealth of open source models have commoditized machine learning techniques, making it tough for horizontal companies to create a competitive moat. For this reason, Madrona focuses on vertical AI/ML companies who have created and leveraged a valuable, proprietary data set to deliver value to their customers.

Jay spent the bulk of the presentation on the moat of every AI/ML startup: its data. What I found most surprising is his conclusion that the most successful companies win because they have created a proprietary, organized, and comprehensive data set, not because they have the most innovative or cutting edge machine learning technology.

Data takes on many forms and not all types of data are equal in terms of value and ease of use. Jay identified 5 primary categories of data and their benefits and pitfalls:

• Customer data is data you collect from customers using your product. It is inherently proprietary and can lead to a virtuous cycle of data, however, comes with the classic “chicken or the egg problem.” When you are starting out, you have no data to leverage to provide a compelling experience for customers. To collect data from customers, you must provide a compelling experience.

• Enterprise data: operational data that belongs to your customer, such as log files, corporate documents, or your customers’ customers’ data. This data is highly sensitive and protected and comes with security scrutiny and expectations (e.g. SOC 2 compliance). This data helps you optimize your business.

• Public data: “open source” data that public and private institutions release regularly. Public data is plentiful and free. However, everyone else has access to this data, including your competitors.

• Industry data: created and curated by a 3rd party, this data can be scraped or purchased. This is somebody’s else’s data, and they may not want you to have it, making you dependent on a 3rd party.

• Personal Health Information (PHI): this includes all healthcare data, ranging from doctors visits to medical tests. This data can be used to solve powerful missions in the healthcare field, but comes with strict privacy laws, is held by a small number of parties, and is very difficult to obtain.

In the closing, Jay left the audience with the following calls to action:

• Take advantage of online resources: When Jay started his ML journey, his primary resources were academic whitepapers or research journals, which can be unfriendly and intimidating. Today, you can find a wealth of publicly available resources online for both education and tooling. To get started, Jay recommends watching YouTube explanations and downloading open source Python libraries.

• Prioritize data collection and management: Your product will have the strongest competitive moat if you build it on a foundation of heterogeneous sources. Jay recommends bringing together multiple sources of siloed data in a unique and novel way and then adding machine learning. In addition to collection, companies must also be disciplined in how they maintain and protect this data. Jay recommends investing in security, storage, and operational processes, such as labeling up front, to maximize your data’s potential.

• Reach out and get started: Madrona Venture Labs loves helping budding entrepreneurs with their startup ideas. Jay encouraged audience members to reach out to discuss their ideas and to learn more about the resources for founders at MVL.

The full panel discussion is available on AWIP’s YouTube Channel. If you’d like to dig deeper into the topic, check out Jay’s blogs on the Anatomy of an ML startup: Part I and Part II.