There was an interesting article posted a while back on the Andreessen Horowitz website discussing the idea that AI is inherently different from traditional software, and sharing the opinion of the writers that AI businesses do not resemble SaaS companies.

In their words: most AI systems today aren’t quite software, in the traditional sense. And, AI businesses, as a result, don’t look exactly like software businesses. They involve ongoing human support and material variable costs.

They often don’t scale quite as easily as we’d like. And strong defensibility — critical to the “build once / sell many times” software model — doesn’t seem to come for free. Let’s tackle these arguments one-by-one.

This article was co-written with superstar investor and active chairman Gur Shomron

“Deep learning costs a lot in compute resources, for marginal payoffs.”

It’s important to note that the AI business model differs from a typical SaaS business model in the sense that there are higher cloud/compute resources costs associated with building models and execution. The AI business model is costly to maintain. Cloud costs can reach 25% of revenues, which is quite a lot.

Another aspect of the AI model that differs from a traditional SaaS business model is that the “human involvement” component is greater. As a result, moving between “domains” can prove costly.

Improving the AI business model generates marginal benefits in the sense that you can spend a lot and experience relatively small improvements.

While the Andreessen and Horowitz article emphasizes these facts, there are additional facts that are equally important to consider.

For starters, the objective of an AI model is to develop a single model that can be easily adapted to more models or fields. My team developed our AI infrastructure so that it can serve many clients from totally different domains.

This way, moving between domains is relatively easy, and, in the worst case, only requires minor adjustments from our end. We serve clients like Uber with the same AI architecture that serves games like Candy Crush.

Regarding high server costs, we’ve found that these can be dramatically reduced, for example, by using the services of tech companies that aim to help you with exactly that. A great example of this is Granulate (which cut our CPU costs by almost 60%!). Check out the case study we did with them.

Furthermore, there are some ML techniques that can help reduce cost and effort. One is using transfer learning, which involves storing knowledge from one problem and applying it to a different problem. Transfer learning can be used to build models on the basis of models that have already been trained.

Another option is making use of some amazing open source pre-trained models like BERT by Google. These are a great way to reuse powerful models that have been trained by others — and they don’t cost you anything since they are open source (thanks, Google!).

Another way to reduce costs is knowing how to apply the principles of agile development to data science and research. This can significantly reduce labor costs. And it’s not only about talented people; it’s also about competent managers who guide these people to delivering results. More on that here.

A final note: In the near future, we expect special AI chips to be developed that will dramatically reduce the costs of training deep neural networks.

In addition to the anticipated rapid development and improvement of frameworks in the deep learning domain, we also predict that training a deep neural network will, in the near future, be much easier and thus relatively inexpensive.

“Machine learning startups generally have no economic moat or meaningful special sauce.”

Defensibility is difficult to achieve. In my company’s case, our economic moat comes from the unique data we have on our audiences, our unique ability to capture real user interest through our deep categories, and the fact that our customers trust us with their CRM data. Our clients trust us and share their data with us.

Their data is used to train models and deliver the best results back to the client.

In the AI world, your economic moat is your ability to access data others don’t have access to, and generate models that outperform those of your competition.

In some domains, off-the-shelf deep learning (DL) architectures can get you close to state-of-the-art performance, and black box models are kings.

The moat in those fields is primarily the data sets your models are trained on and, to a lesser degree, the DL architecture you are using. In other areas, hand-crafted features and domain knowledge, as well as the unique data sets you have access to, will be the game-changers.

Sometimes, feeding your model the right signals is all you need to create the state-of-the-art. This is about knowing your domain and asking the right questions, as much as it’s about having the right model or the latest DL architecture in place.

“Machine learning startups are mostly services businesses, not software businesses.”

Ouch. This one hurts!

It needs to be clarified that there are opportunities in providing services alongside products. Customers need high-quality consulting surrounding their data, and these services can accelerate the adoption rate of AI technology while yielding better results.

At my company, we’ve seen that we can provide added value by enriching customers’ data with our models’ predictions and precomputed features, and by bringing new insights as a function of “big data fusion.”

The more you use virtually the same model for many companies, your service component is smaller and your business model is closer to the lucrative software business (SaaS).

“Machine learning will be most productive inside large organizations that have data and process inefficiencies.”

The way we see it, organizations of all sizes can benefit from the power of machine learning. While it may seem like larger companies have more data to work with — and that’s indeed often the case — smaller businesses can be much more focused. For example, smaller businesses can gather data in a specific domain and become the market leaders in that domain.

At my company we used this approach and became experts at post-install optimization for game developers. Large companies are less likely to focus on and gather data in only one field.

Another point to consider: when solving business problems using machine learning, moving fast and trying new things is the way to go. This is the way smaller companies or startups work inherently. It’s more difficult to move fast and try new things at larger companies that don’t have data science built into their DNA.

Conclusion

There are a few key points to remember when considering the AI business model. Costs can be better managed by partnering with companies like Granulate that are designed to help save money on servers, and by using transfer learning and open source precomputed models.

Better team management, which increases efficiency and drives teams to develop models that can be transferred between domains, is another important tool for managing costs.

AI businesses must obtain an “economic moat” by standing apart from competition. They can do this by accessing and using domain knowledge that others haven’t been able to obtain, and using that knowledge to craft high-quality features that help models achieve the state-of-the-art.

Whether it’s products or services or a mix of the two, there is much that AI businesses have to offer to companies of any size. AI is here to stay and will mature with time. Companies that adopt AI techniques will gain a competitive edge over companies that continue relying on traditional approaches.