What I’ve learned from interacting with more than 100 AI startups

Photo credit: Thomas Kvistholt.

Disclosure: Manish is an investor of Locus.sh.

Artificial intelligence is getting a lot of attention these days from startups and investors alike. Since Pi Ventures is a fund focused on supporting applied AI companies, we meet quite a few entrepreneurs every week who are building companies in the field. Interacting with these startups has been a very enlightening journey for us. Here are some lessons which may be useful for folks wanting to venture out into the space.

Technology or business first?

Typically, startups take two broad paths: either they build an engine which is disruptive then look for a business problem where it can be applied or they have a business problem in mind and then they build an engine to solve it. For example, I may have developed a really good computer vision algorithm which I would then apply in healthcare for diagnostics or for security applications. Alternatively, I already have some insights on the gaps in the logistics industry and then I figure out an AI engine to solve it.

It does not matter which path you take. What matters is that you have some unique insight on the technology or the business problem you are addressing—or better yet, both. Otherwise, it is easy to get lost.

Platform (horizontal) or applied (vertical)

Another important decision for startups to make or discover in the process is whether they can do a better job in enabling other startups to make products/solutions on top of their engine or directly go and address the vertical use case. For example, I recently met a startup that created a B2C chatbot avatar in Hindi. Now, they have two options: they can either use the data they get from the chatbot to launch an NLP API which other products can use (the platform play) or they themselves can launch a chatbot which lets you interact with a bank in Hindi, for example (the applied play).

Both are valid business cases. You can decide which way to go based on your own skills and market scenario. Note that in the horizontal or platform play, you have to be really good at technology and preferably enter the market early. Whereas in the vertical or applied play, you have to understand the domain and, if possible, start in a very narrow application area, as you might be able to bootstrap with a lower amount of data and still deliver business value.

Understanding the domain

Now let us assume you are building an applied startup. One of the questions we investors often have is how deeply you need to understand the domain. As I look at the deals we have done, there is a mix. In some startups, the founding team’s main skill is technology and they either learned the domain or took advisors onboard to bridge the gap. Other startups, on the other hand, have a domain specialist within the founding team.

Again, either way is OK. However, if you are in the applied AI space, you will need to understand the domain to build a relevant solution.

Figure out the data strategy

Data is key to an AI algorithm. Hence, figuring out how much data you need and where it will come from is critical. To build a solution with business value, you need to bootstrap it with enough relevant data. We have seen very innovative solutions on getting data to start building your models.

Figuring out how much data you need and where it will come from is critical.

One cool example is a startup called Vernacular.ai. They built a Facebook chatbot called Ayesha that chats with you endlessly in Hindi, telling jokes, reciting poems, conversing about astrology, and more. The bot went viral and it gave the startup rich Hindi chat data to build their NLP engines on.

In another example, Locus.sh built a smart routing and allocation engine (without AI in the beginning) for hyper-local delivery, which served a need for their customers. As their customers used the product, the startup started receiving the much-needed data to build their machine learning (ML) and deep learning algorithms. Now, their solution uses real-time data to predict outcomes.

The cost of acquiring such data sets is also critical.  The initial cost of data may be much lower than acquiring the data at scale. Hence, it is very helpful if the startup can figure out a non-linear, smart way to get data for the AI engine.

Figure out machine learning vs deep learning