Three Crucial Lessons For Launching an AI Startup

Because AI startups require subtly different strategies than Software-as-a-Service (SaaS) startups.


Eddie Pease

3 years ago | 8 min read

Let me be upfront: I was the technical co-founder of an AI startup and it failed.

PharmaForesight was an AI startup in the pharmaceutical business intelligence industry. Here was our elevator pitch:

“The rate of return for pharmaceutical companies on their R&D is currently below their cost of capital — therefore it is becoming less profitable for pharmaceutical companies to invest in innovative drugs. To decide what clinical trials to conduct, the likelihood of approval is a crucial metric which is currently being calculated in a very subjective and biased way. Our AI algorithm can estimate this figure much more accurately, saving time, money and ultimately benefits patients.”

PharmaForesight failed despite following much of the best practices associated with startups.

We had a strong team and iterated fast using a lean startup strategy. We conducted just shy of 100 interviews with a variety of different stakeholders to identify the early adopters and validate demand for our product.

After only four months, we partnered with the global portfolio management office of a large pharmaceutical company that paid us to build our model and we retained all of the IP.

But ultimately, things didn’t work out—due to some bad luck but also poor judgments. AI startups require subtly different strategies and approaches to Software-as-a-Service (SaaS) startups — these aren’t widely appreciated. My aim in writing this article is to tell you about our mistakes so you don’t make the same ones.

First of all, what is an AI startup?

There are plenty of startups claiming that they use AI but in reality, they actually use outsourced human labor or basic statistical techniques. A study by the London-based MMC Ventures found that 40% of so-called AI startups were not in fact using AI.

For the purposes of this article, an AI startup is one that could not exist without relatively modern machine learning techniques. As an example, could not exist without its deep learning algorithms.

This is in contrast to a number of companies whose product is merely AI-enabled. Spotify, for example, has invested heavily in machine learning and it is central to its strategy today. But Spotify could and did exist before it started using machine learning in a concerted way— to me, that’s more of a SaaS company.

With that in mind, here are the lessons I learned from launching an AI startup.

1. Proprietary data is key

The best way to think of AI in a business context is as an underlying enabling technology, much like the advent of SQL databases was in the 1980s. SQL enabled billion dollar industries such as Customer Relationship Management. In the same way, AI will create new industries and enable improvements in a large number of business use cases.

Like SQL databases though, AI relies on data. It’s long settled that data is far more important than a better algorithm. Good quality proprietary data is absolutely crucial for AI startups.

In hindsight, our data strategy was wrong. Initially, we chose the quicker and easier option: building the first version of our tool on publicly-available data. It took a huge amount of time to clean and transform the data to be ready for machine learning and we figured there was a certain amount of defensibility in that.

We also thought that once we had gained a certain amount of credibility, it would be easier to access more interesting and defensible proprietary datasets.

Many of these assumptions were proved to be wrong. When we started building our model, we couldn’t find anyone else (publicly, anyway) who was tackling it, but when we had finished there were a number of competitors. Even if our algorithm was more accurate, it was incredibly hard to differentiate ourselves against more established competitors, particularly because it seemed that everyone was using similar data.

Building our prototype did not seem to make it easier to access proprietary industry data either (although we were running out of funding at this point so we could have no doubt tested this more thoroughly).

The takeaway is that access to a proprietary dataset is absolutely key for an AI company.

In general, there are three ways you can get a proprietary dataset and they are not mutually exclusive:

  1. Gather the data yourself by creating an initial product or service which generates data when users interact with it. This data can subsequently be used to improve the product or service. Think Facebook, Google, Spotify, and many others.
  2. Gather the data yourself by manually collecting a small proprietary dataset. This can be used to train an initial machine learning model, which needs to perform well enough to satisfy the needs of at least some early adopters. Subsequent partnerships will enable the scaling of data gathering which can then improve the model etc. Hoxton Analytics is an example of a company that followed this approach.
  3. Do a deal with an existing data holder (typically a large company or public institution). For example, Sensyne Health has done a deal with a couple of NHS trusts in the UK.

Of the three options, I would recommend the third. Here’s why.

If you go down route one, you are not an AI startup. AI is clearly not crucial to what you do as you can provide the service without AI. Sure, AI might massively improve your product/ or service but it needs to be good enough to collect a critical mass of user data.

With the second option, it’s possible to create an AI startup, but to maximize your chance of success, the initial dataset needs to be sufficiently niche or your approach needs be sufficiently innovative compared to existing solutions.

The risk in following this approach is that before you develop partnerships to gather a critical mass of data, your idea and dataset can easily be copied by a competitor, particularly if you are tackling a well-known use case.

That leaves option three as a key way to build an AI startup — doing a deal with a large data holder to access their data. This is the reason that the vast majority of AI startups are B2B. Big institutions and companies are usually quite slow-moving so this will usually take some time.

There might well be ethical or commercial concerns about allowed another company to access their data, which will need to be ironed out. In general, companies are becoming increasingly aware of the value of the data they hold.

Of course, some AI startups have followed none of the above and have done well based on the strength of their algorithms — examples include DeepMind (bought by Google in 2014 for $500m), MagicPony (bought by Twitter in 2016 for $150m). But this path is tricky. It is much harder to maintain a competitive advantage without a proprietary dataset.

2. Raising money for AI startups is hard

Trying to raise funding was one of the hardest parts of growing our startup. It combines so many skills: telling a good story, pitching, commercial nous, legal, and more.

We found there were particular challenges raising money for an AI startup.

We assumed that if the idea and team were strong enough and we had good enough traction, then we would be able to raise money.

How wrong we were.

It is absolutely vital to think about your likely funding right at the beginning of your startup journey. Different funders have different objectives and constraints — it’s important to appreciate these from the start. The two main groups of funders for early-stage startups are:

  • Technology Venture Capital (VC) — institutional investors in early-stage companies. At the early stages, they look for three main things — a strong team, a large market-size, and good initial traction. A large market size is crucial. As VCs typically put large amounts of money into very risky ventures, they expect most of their investments to go bust. So for the investments that go right, they need to a) see a return of more than 10x and b) see that return within ~ a 5 year time period. This means that VC-backed companies today are typically Software-as-a-Service startups (SaaS) which are focused on disrupting large industries. If you accept VC investment, the founding team will usually have less control over the company. Most VCs insist on preference shares (ability to claw back your equity if the business is sold for under the valuation they invested) and ability to get rid of the founding team (although this is rarely exercised).
  • Angel Investors — Angel investors come in all shapes and sizes — some look to invest alongside VCs, others look to provide more patient capital. Angel investment typically means that you retain more control over your business, although in the UK at least, it is hard to raise angel investment in excess of £0.5m unless you are either well connected or had previous entrepreneurial success. Most angel investors will be looking for an exit within 10 years or so.

Based on the above constraints, there are particular challenges in finding funding for an AI startup.

First, AI startups typically take longer to get off the ground than SaaS startups. AI algorithms rely on data and large data holders are typically big companies. As discussed above, getting any sort of access to data held by large companies is time-consuming.

Even when you have access to data, you not only need to focus on business development and your software platform (like in a SaaS startup) but also your AI algorithm.

Given that you need more specialist skills and it takes longer to get an AI startup off the ground, this means that you typically need more money to launch an AI startup and that money needs to be “patient capital.” For most founders, this rules out long-term angel investment (unless you are insanely well-connected) — the amount of capital required is just too great.

When you do pitch to VCs though, you are competing against traditional SaaS companies, which are likely going to offer a quicker return if all goes well. SaaS is an attractive business model due to its regular recurring revenue and tendency by users to forget to cancel their subscription even if they do not use the service frequently.

As the SaaS business model has been so successful for VCs in the past decade, we found that many VCs were stuck in this way of thinking, even when not appropriate for many AI startups.

We heard a lot of “Come back to us when you have some subscription revenue.”

Most AI startups will find it hard to generate subscription revenue for at least the first few years and may well require a different business model altogether. Your proposition needs to be even more compelling to raise money.

3. Depending on your use case, explainability can be key

Even if you have a proprietary dataset and a brilliant product, it doesn’t necessarily mean that your product will be a hit. If you’re launching an AI startup, you are (hopefully) knowledgeable about AI and machine learning.

The average person, however, isn’t familiar with these topics and might be sceptical of their potential.

Simply put: you need convincing proof that your model works well. A live demonstration might work but if that‘s not plausible, try using specific hand-picked examples rather than high-level accuracy figures. This might sound counter-intuitive, particularly if you’re from a mathematical background.

Being able to explain the predictions from your model will increase trust in your model. Depending on your use case, the ability to explain each prediction from your model will often be as important as the accuracy. Explainability is a huge topic but in general, the more “human-like” your explanations, the better.

A good rule of thumb is that the more important each individual prediction, the more important explainability.

Final thoughts

AI is an amazing enabling technology that will no doubt have an incredible impact on our lives in the years to come. But that doesn’t mean that setting up an AI startup is easy—far from it.

Indeed, in many respects, I’ve come to learn that setting up an AI startup has a number of unique difficulties, many of which I don’t think are fully appreciated.

After setting up PharmaForesight, I’m a big believer in the Henry Ford quote:

“The only real mistake is the one from which we learn nothing.”

Hopefully we made the mistakes above so you don’t have to.


Created by

Eddie Pease







Related Articles