The Power of Vertical Machine Learning Startups

Aquaculture — a perfect use case for machine learning?

Many of the biggest enterprise software companies in Silicon Valley history followed a similar playbook. That playbook? Leverage a platform shift — a fundamental change in the way technology is built and delivered to users — to build a horizontal business that serves multiple industries and unseats an incumbent built on last generation technology. Salesforce, Workday, and Square are classic examples.

There are many obvious benefits to going horizontal as a startup, but the tech ecosystem today may not be particularly friendly to startups that pursue that strategy. We’re in the second act of cloud and mobile and horizontal incumbents look more dominant than ever. It’s difficult to get excited about the prospects for a startup that is competing with Salesforce in CRM, much less one competing with Amazon, Google, Facebook, or Apple.

The technology we’ve been most excited about for 6 years and counting at Costanoa is machine learning. We continue to see jaw dropping advances in both academic research and real world application. But for all the promise of machine learning, it isn’t really a platform shift. Modern ML techniques are fundamentally built on drawing correlations across large data sets — the advantages generally accrue to those with the best data. Unfortunately, the underlying data which ML models leverage often live in the system of record from last generation. It is thus no surprise Incumbents have moved quickly to integrate ML into their existing products and leverage their existing datasets.

This all paints a bleak picture for startups. But as optimists at Costanoa, we see a path to leveraging ML in a way that incumbents can’t and towards building a large, enduring and defensible businesses.

Go Vertical

Vertical machine learning applications take a business process in a specific industry and automate it in ways never previously thought possible. They leverage technical capabilities including but not limited to computer vision, natural language processing, audio transcription, and pattern recognition across large data sets to fundamentally transform an industry’s value chain.

I see many founders attempting to use ML to rebuild horizontal, industry-agnostic applications like a next-gen CRM or ERP. While these are worthy endeavors, we’ve been more excited about the potential of vertical enterprise ML startups that are purpose-built for underserved industries. Here’s why:

Today’s machine learning techniques are really built to go vertical.

For all the promising advances of the current AI summer, modern ML techniques still have limited ability to translate learnings across domains or solve highly general problems. Neural nets work best in a highly confined domain where they are trained to recognize a specific set of signal data in a massive, noisy data set. They often struggle mightily in more general problems.

I see this time and time again with startups using ML to automate general tasks like note taking without a specific industry focus — because of the nuances around data type, the accuracy is just not good enough to be very useful. ML startups that target a specific business process in a specific industry massively narrow the domain to learn on and reach viable product at much faster rates than competitors that stay broad.

For example, Synapse uses computer vision to detect contraband in x-ray scans of luggage at airports. By focusing all of their effort learning specific contraband in a specific image type (x-ray scans of luggage) they’ve reached detection rates that vastly exceed human agents in <2 years. Vertical startups can also partner with pilot customers to quickly collect data that no other company has.

Vertical startups face less competition and have superior ability to cut through the noise.

Vertically-oriented approaches can also shield startups from getting trampled by the mastodons of the tech world. FANG and its ilk are so big and so horizontal that they’re unlikely to ever go deep into a vertical, particularly one that isn’t a decabillion market. And more importantly, they don’t always have the foresight to see the initial TAM, expansion TAM, and tangential TAM opportunity (to be discussed below in depth).

Startups that carefully select an underserved vertical also have phenomenal ability to cut through the noise and reach customers quickly — while there are thousands of startups vying for the attention of the CMO or CISO, there are relatively few building the best product in the world for an unsexy but big, important industry. We’ve seen this in companies like Cedar.ai (AI for industrial rail) and Aquabyte (ML for aquaculture), who out of the gate have commanded the attention of executives at large companies in their respective verticals. It isn’t everyday they get a pitch from someone with an idea for how to transform their industry.

Vertical ML can scale quickly

This leads into the third point about the power of vertical machine learning applications — when product market fit is achieved, they can scale quickly due to the efficiency of their focused sales efforts. The universe of potential buyers is well-known and one can be highly targeted reaching those buyers. Furthermore, if you can deliver on the promise of automating portions of the value chain, first followers will move quickly to buy your product or risk becoming uncompetitive with rivals that have already adopted.

There is often potential for market expansion and second horizon opportunities

The knock on vertical businesses is constrained addressable market. It is probably the case that a vertical ML app will never have the market cap of Facebook or Google. But we often find these fears to be overestimated and expect many vertical ML apps to ultimately become much bigger than their early TAM might appear. What happens to the size of the security imaging market when you no longer need to staff x-ray scanners with agents? How does the utilization of industrial rail change when rail cars no longer spend 30% of their time idle in rail yards?

We also often find that machine learning applications have second horizon opportunities once their initial product is delivered. For example, the aquaculture industry is plagued by inefficiencies in biomass estimation — it is common for fish farmers to forecast a yield of 1,000 tons of fish and only deliver 800. Aquabyte’s technology can bring improvements in the accuracy of those forecasts (first horizon opportunity) which over time they can use to become a marketplace themselves, helping farmers cut out the inefficient network of brokers to sell product because they have true visibility into supply (second horizon).

So don’t be scared to go vertical with ML. If you spot an opportunity to fundamentally alter the economics of an industry, good things can happen. Big things, too.