Technology Evaluation of Artificial Intelligence (AI) Startups

Our coffee intake has jumped as more and more of our day is spent sitting in coffee-shops listening to passionate founders pitching AI-based solutions. A critical part of investing in AI startups is understanding the technology. But before going deeper, we first try to understand whether the problem the startup is solving requires AI.

The key is not just to use AI, but to use it intelligently for problems that require it. Many tasks that one may feel like applying deep learning to can actually be carried out fairly effectively using rules and basic automation. For a problem like helping a company close its books faster, automation is sufficient, because the process involves a fixed, narrow range of inputs and outputs. Applying AI here would not only lead to a drain of resources but unnecessary complications in the implementation of an otherwise relatively simple program.

On the other hand, a 10,000 people organization is going to create millions of vulnerability points on a regular basis. Not only are these points unpredictable, but even the range of these points is hard to pin down and put into a rules-based setup. Hence a Cybersecurity automation system designed to cater to only limited scenarios will be highly ineffective. Neural networks and deep learning will do a much better job. Here we see an example of an opportunity where AI can add real value.

We like to invest in spaces like these, where an AI-driven solution is the most optimal approach to solving the problem. Once the need for AI is clear, we do a deep dive into the startup’s technology. To help us structure this assessment, we apply SenseAI’s own VDAT framework, that uses 4 core factors required for success as an AI-centric startup:

1. Variety in data
2. Data acquisition
3. Architecture
4. Talent acquisition

1. Variety in Data: Of the 3 Vs of Big Data (Volume, Velocity, and Variety), Variety plays the biggest role in AI.

Variety of Data refers to the different kinds of possible inputs and outputs for a given system. When there are few inputs and outputs and they are easily quantifiable, simple automation suffices. A good example of this is making ledger entries for a firm according to local accounting procedures and rules. While the input is limited in its format to numbers, the output is determined by simple yet static rules.

The game changes when we talk about a dataset like the visual inputs received by a self-driving car. Here, the variety of inputs is virtually infinite. It is important that the system constantly detects and processes input which is chaotic and varied. The same applies to situations like an unauthorized intrusion in the network of a submarine or a corporate environment. Voice-driven assistants, especially in the Vernacular domain are another example of effective use of AI.

2. Data Acquisition: Effective Machine Learning or Deep Learning requires vast quantities of data. Hence, a firm’s strategy to acquire data is crucial to its overall strategy. At a high level, the primary source of data is a combination of publicly available data, a firm's own unique data, and customer/partner provided data. Any successful startup must have a strategy to acquire and integrate data from these sources in a cost-effective manner.

With regulations like GDPR, there are regulatory restrictions around the capture, storage, and analysis of data. We assess startups based on their ability to recognize and work around these issues effectively.

If a startup is able to build large and high-quality datasets it can give them an edge over the competition. We ask questions like: Is it possible to create a Network effect around the data? Can we build a moat around the firm’s own data?

In a product experiencing the Data Network effect, as the volume of data provided to a learning system increases, it generates more value to any individual user even though they only interact personally with a small sliver of the data ‘network’. This effect arises due to relationships that emerge between different segments of data.

A Data Moat, on the other hand, describes a competitive advantage a business holds because of its proprietary data set. It’s a modern version of traditional “moats” such as branding, trade secrets, and regulatory capture.

We like to invest in startups that strategically exploit the above data-driven advantages, and have a comprehensive strategy for data acquisition and processing.

3. Data Architecture: The architecture into which the data is fed needs to be scalable in line with the long-term vision of the company. We like to see if the architecture can be scaled to enterprise-grade or higher. An organization can only fully utilize huge amounts of data from different sources if a sound Data Architecture is in place across the organization.

An important factor the architecture determines is the analysis time-frame. For some kinds of problems - say, obstacle detection in a self-driving vehicle - real-time analysis is essential. The processing power required is then proportionally higher. Other kinds of machine learning applications - say, assigning credit scores to borrowers - can function even with a relatively delayed response. The availability of adequate processing power and building of a data architecture fitting to the solution is important.

Decisions also need to be made about whether the data architecture is built on the premise, on public cloud, or a hybrid combination of the two. These decisions must be made in accordance with the requirements of the product, and in a B2B model, in line with the client’s data policies.

We like to invest in companies that proactively understand the importance of these decisions.

4. Talent Acquisition: We honestly don't expect entrepreneurs to have answers and solutions to all the issues raised above. However, we do prefer those that have at least thought through these issues and have a rough plan.

Recognizing the importance of these issues at a high level goes hand in hand with being able to hire the people that can explore the answers to these questions in detail. It is only in collaboration with domain-experts that AI-based products can be built effectively. An entrepreneur’s ability to understand the talent requirement and bring on those that are, in a specific field or technology, more knowledgeable than them is crucial.

Having a simple framework like VDAT, which distills years of understanding, helps us develop our own hypotheses around the team and product capabilities. Like any framework, we do understand that it may not be all comprehensive but it has allowed us to make some tangible decisions. We combine it with the frameworks we use for Business Evaluation to make the call regarding whether to invest or not.

Written by

Vinish Kathuria

Managing Partner at SenseAI Ventures

Click on this link for more blogs by LetsVenture!