Part 1 — The Secret Master Plan of Enterprise AI/ML

At the time I write this essay, Tesla just reached its 4th consecutive quarter of profitability. With almost 400,000 cars sold in 2019, they have arguably already reached their goal of producing electric vehicles for the masses.

But how is that relevant to AI/ML, you ask? Well, in his 2006 Master Plan, Elon Musk revealed the 3 “simple” steps that would make Tesla a wild success. Tesla would first build a sports car, then use that money to build a more affordable car, and then use that money to build an even more affordable car. I think the enterprise AI/ML industry is following the same model, and we are about to enter step 2 of the master plan.

I keep reading that AI is fundamentally biased, or evil, so we should scrape it from our lives and enterprise processes. Meanwhile — as an operator, founder & investor in 12+ venture-backed AI/ML startups — I keep getting asked by my investor friends if I have any cool lists of AI/ML startups as the sector continues to bring in massive investment dollars. So I’m gonna let you in on a few secrets that our AI overlords don’t want you to know.

Credits to Anonymous person on the internet

There is a clear gap between how AI/ML insiders and outsiders evaluate the sector and what it takes for an AI/ML upstart to win its category. I encourage you to use the framework in these essays when thinking of your own enterprise AI/ML startup and building your product roadmap. This will be a constant work in progress so any feedback is welcome!

Alright, let’s get back to this Tesla thing. Ten years ago, Tesla was undergoing rapid change in its production, supply chain, and more to get to the more affordable Model S. Well, Enterprise AI/ML applications built in the 2010's were essentially in the “Roadster phase.” And we are now entering the “Model S phase.” Or Step 2 of the secret master plan.

To be clear, the general mathematical theories on which most modern AI/ML models are built, like the Convolutional Neural Network, were already there in the late 1970's. Similarly, the first electric cars were conceived long before the modern Tesla, and actually held the land speed record in as early as 1900.

Thomas Edison taking a look at the record-breaking electric car he didn’t invent.

Here’s a (not really) super, mega contrarian idea. Designing, building and delivering a breakthrough product category to the masses takes decades. After all, it took us almost a century to go from a combustion engine in 1820 to the first mass-produced combustion engine car, the Model T, in 1914.

My core thesis is this - enterprise AI/ML today in 2020 is where Tesla was a decade ago. Somewhere between the roadster and the Model S. Here are 5 core reasons why:

1. The enterprise AI/ML applications of the past few years were extremely good at specific tasks like object recognition, transcribing audio, or basic security threat detection. It was awesome for flashy applications like beating the best Chess/Go/Starcraft player. And automatically differentiating a cat from a dog in a picture without asking your 3-year old. It was also pretty good at transcribing your awkward Zoom sales calls to see how many times you said “uh” and “hmm” in under a minute. Besides that, not so awesome for the mass-market. Enterprise AI/ML apps lacked the infrastructure (and product-market fit) to be truly scalable, integrated and explainable. Or, more often than not, to even solve a real business need. That was the “Roadster phase.”
2. The scientific foundation of AI/ML has been around since the 1980’s. But the foundation to get us “Model S” enterprise AI/ML apps is finally here: better data infrastructure, data volume/quality, nascent regulation, and elegant data scientist tools. And most importantly, product-market zeitgeist. We basically had the engine and a crappy battery in 2010. And now we have the radiator, the alternator, front-steering, half-decent brakes and a few other parts you’d probably want to have in your mass-market vehicle. That means you don’t have to raise tens of millions of $ in funding anymore before getting your AI/ML app to “Minimum Viable Happiness”.
3. The market doesn’t want “pure” AI in their enterprise apps. It wants AI/ML that will be a complement to, rather than a replacement for, many of their jobs in the foreseeable future. Just as Tesla’ Autopilot is effectively a human/machine hybrid approach to the task of driving, many AI/ML apps will provide coaching to workers. They will learn from them as they work and share those learnings with all the other workers performing a given task. Some call this the human-in-the-loop approach to AI/ML. Others like my friends at Emergence Capital have a whole, incredibly thoughtful thesis about it. In part 2, we cover why enterprise AI/ML apps with a human-in-the-loop will dominate “pure” AI/ML apps.
4. New platforms like Zoom, Carta and Plaid, with their API’s, provide the infrastructure for data scientists to leverage more complex data (e.g. video, financial information) and work on more interesting problems. This data was previously impossible to gather IRL with the context and scale necessary to build mass-market AI/ML apps. Now, you can get it with 1 software engineer intern. But this new ocean of opportunity, in turn, increases the need for accurate data capture in many new contexts/scenarios. More on how solving for this can get your enterprise AI/ML app to mass-market adoption in part 2/3.
5. Lastly, all industries are not equally susceptible to adopt enterprise AI/ML apps at scale, beyond small pilots. Some, like HR, financial services or healthcare, require much higher standards around explainability, privacy and mitigating bias. Incidentally, AI/ML apps disrupting these industries require additional infrastructure tools to tackle these requirements. And yes, bro, we get it. Your “AI/ML” app will get better with more data. The average Fortune 500 executive gets it, too. At least conceptually. But in 80%+ of use cases, enterprises would rather see your model’s output be a bit less accurate than even their current manual process today. As long as it easily fits into their workflow, it’s transparent in how it learns from their data, and it can explain itself fairly well. Those are key ingredients of “Minimum Viable Happiness” in enterprise AI/ML. And adding a human-in-the-loop can solve for a lot of this. More on how it should inform your product roadmap in Part 2.

Note: A cautionary tale of this is Amazon’s sexist AI recruiter disaster. Similarly, but with no relation whatsoever, if you own a Tesla Model S you are more susceptible to be a 52-year old male homeowner with better credit than a Porsche owner. Nicely done, bro!

To wrap up, the decade of 2020’s will be the one where we see applications of enterprise AI/ML bridge the chasm to the mass-market. Or in other words, go from the Roadster to the Model 3 of enterprise AI/ML. We will see this in almost all verticals, like Education, HR, Manufacturing, Healthcare, IT Security and more. Stay tuned for Part 2, where we will cover how to get your enterprise AI/ML app to tip into mass-market adoption and win its category.

Thank you Kat Orekhova, Jake Saper , Pietro Invernizzi, Cyril Brignone and Gregory Haardt for your invaluable feedback to help this see the light of day! And Sara Tavel for inspiring this format and some of the terminologies used here with her amazing series ‘The hierarchy of Marketplaces.’

Liked it? Read Part 2 here. Follow me on Twitter here, or on Linkedin, to be the first to be notified about future stories.