What we look for in AI companies (and why they are different)
This week, we held our first AI Invest event at Hambro Perks. Technology experts, entrepreneurs and AI practitioners discussed the latest trends, opportunities and challenges of these new technologies.
I talked about our AI investment strategy alongside a stellar panel that I’ve known for many years. Charlie Beckett (Professor and director of Polis, the LSE’s journalism think-tank), Alessandro Ialongo (AI researcher, Cambridge University/Max Planck) and Tim Porter (Founder, Kare, an Hambro Perks portfolio company).
You can find the video recording of the conference here.
I’m not a machine learning engineer, but I’ve worked with data and machine learning throughout my career. After studying at the London School of Economics, I founded a machine learning journalism startup, which used NLP (natural language processing) and applied it to news articles. I then used machine learning to solve a completely different problem when I was the head of digital for the Congress Party in the 2019 Indian general elections. We built the party’s predictive voter modelling to target over 700m voters. I now invest in AI companies as a Principal at Hambro Perks.
My talk was about the things we look for when we invest in AI companies.
But firstly, let me tell you how we think about AI itself. And address the big elephant in the room, which is this:
What is AI?
AI doesn’t exist. We are far away from achieving AGI (artificial general intelligence) and for computers to recreate a human level of intelligence.
What we do have is machine learning, which is a subcategory of AI.
There are a lot of different definitions out there, but the one that I like to use internally is:
“Programmes that can perform a defined task based on training data”.
These programmes can do tasks faster, more efficiently and at a scale that was never imagined before.
Machine learning has already completely changed the way many sectors operate from health-tech to fin-tech to media. And they do this in different ways.
There are different types of ML companies…
This is how, internally, we categorise them.
- Horizontal vs Vertical applications
- Horizontal use cases: companies that design products that are applicable to many industries. A good example is a CRM software like Salesforce.
- Vertical use cases: companies which build applications that are focused on one specific industry. An example of vertical use case is software for autonomous vehicles.
2. Enterprise vs Consumer
Applications that target enterprises and those who target consumers.
3. Platform vs tool
Startups that want to control the fundamental layer of machine learning and others which are building a tool on top of those platforms.
Our investment thesis
We are more excited by startups that go vertical. We think ML techniques are built to go vertical. The model needs to recognise a very specific set of signal data in a massive, noisy dataset. To get accuracy, you need the same type of data that is relevant to the problem you are trying to solve.
We also tend to be more excited by enterprise focused solutions. Big companies like Facebook, Amazon and Google are concentrated on solving large consumer problems and have all the data in the world to do that really well, so it’s harder for startups to compete.
When it comes to platform vs tools, we prefer companies that start as a tool and then develop as a platform. These are startups that solve a complete business problem. They figure out how to get the data and then how to build a loop to learn from that data and create a user experience that solves a specific business problem.
ML companies are doing different things
There are many trends that we are seeing lately, but there are four areas that we are especially excited about:
- Companies that are building entirely new products and services.
- Companies that are building applications to perform tasks more efficiently or to complete tasks more rapidly
- Companies that are able to solve problems that we weren’t able to solve before, or at least with the same level of time and resources.
- The one I’m personally most excited about is companies which are extending capabilities to new market participants. AI massively reduces the cost of serving a mass market, allowing everybody to enjoy a premium service. A very good example of this is 24/7 security guards. It would cost you a lot of money to hire a security guard to control your house around the clock, but using computer vision you can drastically reduce the cost of that service.
ML companies have unique challenges
While we believe that ML businesses have great potential, we think they also have unique challenges.
The first thing to realise is that ML businesses just don’t have the same units economics of software businesses. ‘Sometimes, they look more like traditional service companies’.
For example they have lower gross margins. Anecdotally, we have seen gross margins of ML companies at around 50/60% vs 60–80% of comparable SaaS businesses. This is due to two factors. Firstly, the hidden cost of the cloud infrastructure. Training a single ML model costs thousands of dollars in computing resources. Secondly, many applications require ‘humans in the loop”, because training models involves the manual labeling of large datasets. Both of these things increase the cost of delivering your service.
Another issue is low tech defensibility. In the AI world tech defensibility is harder to achieve because new/revolutionary models are being developed mostly in open, academic settings and so are available in open source.
What we look for in ML companies
What we look for in ML companies is different from software companies. Here’s some key things we consider before investing in a startups:
👀 Problem
Are you trying to solve a real business need and does it make sense to use ML to solve this problem? ML is all about optimisation. Theoretically you can use ML to optimise most business processes. This, in itself, doesn’t mean you are an ML business.
☄️Go to market
Selling to enterprises is brutal, especially for ML companies. There is lots of competition and cultural push backs. So, we look for companies whose product is so outstandingly great that they can ask for a high enough price to afford the go to market/sales motion.
🔐 Proprietary models built on proprietary data set
You need a proprietary data set to build a model that generates results that are far better than someone else can generate by throwing commodity data into a model. There is a lot of open source data out there so if you use the same data you’ll get the same result as a competitor.
“The unreasonable effectiveness of data” published by Google, demonstrates that by throwing more data at a particular machine learning problem you could get better outcomes than just tweaking the algorithms. Accuracy would improve.
🔄 Data network effects
In other words: the more users use your product, the more data they contribute; the more data they contribute, the smarter your product becomes.
👩💻Tech is not enough
Tech is hard to make defensible with machine learning. We don’t just want outstanding tech. It is the deployment of the tech — the execution — that really matters.
Due diligence phase: ML companies are different from software companies
What is unique about the due diligence on ML companies is that we focus heavily on the data set. We assess things like:
Accessible — How hard is it to get the data set you have? Can it be substituted by similar data that is easy to get?
Rich — Can you use the data to feed a model to generate real outcomes and do you have enough of it?
Perishable — Is the data fresh enough to be close to reality? For example in financial markets, which are changing every second, you need to work with live data.
I hope this clarifies a little more how we invest in AI companies, what we look for in AI companies and why we think AI companies are different from software companies.
Conclusion
Far from the AI winters, the field of vertical AI seems to be exploding, as witnessed by the increased number of startups, capital investment and public interest. We could even be at a tipping point with regards to the scale AI companies could achieve in the coming years. I hope that this framework can bring a little transparency to the way we currently look at the market.
If you’re building an AI company, or have any feedback on this article, please get in touch, we’d love to chat!
