Artificial Intelligence (AI) has gone through a big evolution over the past decade. Through my experience conducting AI research in academia, building an AI development platform startup that got acquired by Salesforce in 2016 (PredictionIO), and launching the Einstein AI Platform there, I have been at the forefront of this evolution and have learned a lot along the way.
As an Investing Executive-in-Residence at Sierra Ventures, I draw from my previous experiences, and I frequently get questions from entrepreneurs about how they can get funding for their AI business. For me, it comes down to the things entrepreneurs do when they’re actually building their business and developing their tech
4 things entrepreneurs need to think about when building an AI business
- Make Sure It’s a Business, Not a Smart Feature
- Do R&D Testing with Real Customers
- Retain Your Customers the AI Way
- Scale with AutoML
Make Sure It’s a Business, Not a Smart Feature
Determining whether a startup idea is just a feature or if it can become a substantial standalone enterprise is sometimes not an easy job for both founders and investors. This question is especially important for AI businesses because incumbents often have the “data volume” advantage by having a large existing customer base. An example is lead scoring in the CRM context.
Infer, Everstring, and other startups pioneered the use of AI techniques to rank the priority of lead contacts. Salesforce later launched Einstein, and lead scoring was one of the bundled features. Startups in the space that cannot broaden or deepen their products fast enough may end up competing with larger platforms that acquire or embed their solution as a product feature, which can bankrupt a startup.
Before you embrace the exciting technical challenges of AI, ask yourself two questions:
- How difficult it is for incumbents to offer what you are building on top of their products? (Hint: “Our AI accuracy will be better because we have the best data scientist team in the world” is often not defensive enough.)
- Are you creating significant value for customers? For example, are you solving a big problem that couldn’t be solved without AI before? Or are you just offering a slight improvement to what customers already have?
Do R&D Testing with Real Customers
Not only does building good AI technology take time, it also takes data. More specifically, it takes real data. I have seen countless pitches that go like this: “We are a technically strong team with PhDs from a prestigious university. We did research with this famous professor in the computer vision domain. You probably have heard of him. We published papers in top journals that prove that our new deep learning technique is more accurate than “X Name” Big Company. We are raising a round to productize this world-best technology for commercial use cases.”
While these are all great achievements, they are not sufficient to show that you own a valuable AI technology. Why? There is often a gap between research results and actual real-world applications.
Questions to follow up would be:
- What dataset did you use in the research? (Often, public datasets curated for research purposes).
- Will your technique work as well for different customers’ operational, thus noisy, datasets universally?
- Can your team actually implement it in production at scale?
- What “X Name” Big Company technique are you referring to, the one that they published openly?
- Also, was this a comparison based on a public dataset again?
From a technical due diligence perspective, all these questions mean uncertainty.
I would much rather know that there are two happy pilot customers today. First, it shows that you have already implemented the AI technique and it is at least working well for some customers. Second, and more importantly, these pilot customers can be strong references to prove that the technology solves a real business problem.
Even if there is no data to train the AI models at the beginning, don’t let it be a reason not to get pilot customers. It may be possible to temporarily simulate the desired AI solution with techniques like human-in-the-loop or rule-based algorithms. From there, you are collecting customer data for real AI.
Retain Your Customers the AI Way
When the sales pipeline is discussed in your potential investor meeting, an obvious red flag is that pilot customers’ POC contracts are dragging along for months without being converted to full-scale contracts. Or even worse, they have churned.
It is crucial to show pilot customers the value of your AI solution early. A simple dashboard with ROI calculations may help. Once the value is visually proven, you are not too far from converting them into full paying customers. When it is difficult to quantify the benefits of some AI features, this dashboard is especially important to help your internal champions show off the positive results in front of management. In addition, they may then be able to concretely explain the business values you have generated during investor due diligence calls.
Making new customers happy about their purchase decisions is a good initial step, but your chance to keep customers will be much higher if there is some sort of “lock-in” mechanism in your business. AI startups often include a slide about the data network effect in their pitch deck. Theoretically, the more customers you have, the more data you can get to train a better AI model, and then the more customers you can acquire again. That’s the beauty of AI business after all.
In my experience, however, the data volume itself is not defensive enough. It is not uncommon that performance gain decreases dramatically after your dataset goes beyond a certain size. Recent technological advancements like Generative Adversarial Networks and other synthetic data techniques enable data scientists to build powerful AI models with much less data than before. To be truly defensive with data, think about a product design or business model that you will be able to:
- guarantee the quality and freshness of data;
- own proprietary data;
- access a larger variety of data sources.
Scale with AutoML
Startups often do things that do not scale at the beginning, which is not a bad thing. It is perfectly fine to manually tune the AI models for the first few customers to gain initial traction quickly and prove the product demands. AI model tuning work is non-trivial though. Otherwise, why would the salary demands by data scientists and data engineers be so high? Imagine the cost required as you grow the customer base if you need to dedicate a team of PhDs to tune a model for every customer for weeks.
During a tech due diligence, showing that your technology can scale with the business is important. One may argue that this problem can be solved easily by throwing in the AutoML technique when needed. If you try this you may realize that the talent that is good at automating AI learning processes is different from the one that is good at fine-tuning a model to the best shape by hand.
You may have built a team of great data scientists who ignore automation. You may also have spent valuable time building an AI workflow that does not work for automation. So if you are building an enterprise SaaS product that targets the small and mid-size market with a large customer set and your product needs a different AI model for each customer, plan for the implementation of AutoML architecture and related team dynamics early.
AI technology creates new challenges and opportunities for startups. Ultimately, the principles of investing do not change simply because it is an AI startup. We always love startups that:
- Are on the path to becoming real big businesses,
- Onboard customers fast,
- Create and retain happy customers, and
- Power their products with scalable technology.