Death of a startup

Yesterday, one of my portfolio companies died.

They haven’t formally dissolved the company yet, but that step is merely a formality. The other events foreshadowing company death are firmly in motion: employees are actively interviewing (many had offers within a week, this being Silicon Valley circa 2019), investors have already told their CPAs to write off the investment, the landlord has been informed, etc.

Two years ago, when I invested in their seed round, the company was full of promise. The two founders were strong technologists with a product bent; both had worked together at their previous employer; they had validated their idea by interviewing decision makers at 100+ companies to determine that yes, there was a clear pain point here and yes, enterprises were willing to pay for a solution; the company’s solution had the two magic letters — A and I — (or the other two equally seductive letters — M and L), and it was real technology, not just made up BS. The round was oversubscribed.

Two years later, I sit here thinking: what happened? Why did this company, so full of promise and so clearly solving a very clear pain point, shut down?

The answer can be neatly expressed in one word: data. Or lack thereof.

In order to bootstrap their machine learning models, the company needed to acquire training data from enterprises. To do this, they built a passive agent that would sit on their customer’s network, ingest the data flowing through said network, and learn.

All well and good. Except no company would allow them to deploy this agent.

It was incredibly frustrating. Nearly every prospect they spoke to kept talking about how serious the pain they were addressing was, and how great it would be to finally have a solution; in the same breath, they absolutely refused to let the company passively collect data from their network to train their models.

Without training data, there was no ML model. There was no AI. Without the model, there was no solution. Without the solution, no company.

What could the company have done differently? In retrospect, they should not have spent so much time trying to jam the agent through. Instead, as soon as they realized the agent was a no-go, they should have pivoted to a different initial wedge that delivered lower (but still non-zero) value to end customers but didn’t need training data. Once their customers got comfortable with the initial solution, it would have been an easier conversation around the agent. Instead, the company went for the “go big or go home” approach, leaving themselves no time or runway to try something different. Turns out we are all going home.

Lessons

• Data is critical to ML but hard to access: The most obvious lesson is that any machine learning company needs access to data, ideally lots of data. An ex-colleague writes: “A lot of my peers in DS/ML have started/are starting enterprise ML platforms, without understanding that you don’t get to do ML unless you have data, and you don’t get data until you build something people want to use. That last part requires a hell of a lot more product intuition that most ML folks don’t have, unfortunately.” Another friend writes: “Getting data for ML in enterprises is a huge hurdle and even big companies are struggling with it.”
• Be careful about proclaiming PMF too early: You don’t have product market fit until you’ve solved for every execution risk. Just because prospects want your product and you’ve built a product that solves their pain doesn’t mean you have PMF. Onboarding customers is a critical part of the puzzle, as is serving them well.
• Listen to the market: Don’t sacrifice your company at the altar of the purity of your vision. Listen to the market. Really listen. Adjust. Adapt. Deliver value versus trying to push through your vision of a perfect world. Live to fight another day.

Postscript

Got some good feedback / thoughts, which I’ve copied and pasted below. They all reinforce the value and defensibility of data.

• Product leader at security company: “A big reason that we have been able to build and train useful ML models is because we were able to get companies to deploy our agent. It was an uphill battle that has taken years, but now, amazingly, the largest companies in the world are letting us access their transaction data. This has been the key to our success, and we regard our position with our customers — and the data it has given us access to — as our greatest competitive advantage. Technological advantages are really just a matter of time for other smart competitors to bridge. It’s the market advantages that are more sustainable.”
• CEO of scientific data company: “The main challenge for both scientific and non-scientific data is the lack of structured and relevant data to train AI models. That is why we focus on making scientific data “AI ready” and getting scientific insight into their data, versus predictive AI. Companies are reluctant to let others apply predictive AI on their data. That’s also the reason why we are seeing strong traction.”
• CEO of stealth AI company: “100% agree with your post on value of data. The other part of data I’ve realized most folks don’t talk about is the value of the data in the format you need it. This is why so many companies are building businesses that look like traditional ones (eg Pilot for accounting, Atrium for law) to collect the data in the format they need and then optimize away parts using AI. But at the core is a fully verticalized business. If things don’t work out, they still have a good low margin (non Silicon Valley type) “business”. But if they do they disrupt something big. A lot of companies I’ve seen struggling even after getting data struggle because the data didn’t have what they need, but they limit themselves to it.”