Blog: Choosing The Right AI Business Model
After several AI projects with startups, I realized that one aspect of the AI disruption remains relatively unexamined: the right business model for AI companies.
Will the product be sold as a service or as a new feature category for users? Mentioning the type of revenue is an important part of any AI project.
Due to the many technical elements related to AI, I have noticed that existing traditional business models don’t always apply.
Business model: a plan for the successful operation of a business, identifying sources of revenue, the intended customer base, products, and details of financing.
One might assume that AI startups, like cloud / SaaS companies before them, share a common business model. However, I noticed that it is difficult to apply the SaaS model to AI startups.
Indeed, depending on the nature of your AI solution, you’ll always need data, vast amounts of raw computing power and algorithms. It’s inherently more complex for customers to digest than cloud startups and it also requires different things, so the technology has to be sold in different ways, too.
Before getting into AI business models, I believe it is important to present the AI landscape based on my experience.
The AI Landscape can be divided into two segments:
- Infrastructure: These companies run in the back-end and provide computational services to others. The business model they follow is generally based on API calls. A great example here is IBM Watson which provides Sentiment Analysis, NLP and Entity Recognition through its Bluemix platform. IBM charges the user for API calls.
- Application: These can be in the B2B and B2C space. Significant activity, however, is seen in the B2B space where companies offer SaaS-based subscription services. These companies generally develop applications for specific use-cases defined by their clients. I can talk a bit more about this one since I am very familiar with it. I realized that some of them end up being purchased by important firms after several proofs-of-concept. Basically, if what those AI companies do is considered as strategic by their clients, some large client will want to acquire the startup rather than rent the technology.
Let’s focus now on AI startups:
There are basically three categories of AI startups:
Back to business models, I have observed two of them that are starting to work quite well, others exist but I haven’t been introduced to them yet.
In this model, a new AI solution will improve the effectiveness of current workflows.
Because of intensive deployments, the sales cycle is long. As a consequence, each deal must be big to keep the startup alive. On top of the important development costs, significant running costs are to be expected. Usually, companies will charge you for the development of a tailored-made solution and then make you pay monthly running costs as well as operational support/training.
The business model is more or less similar to the SaaS model. It concerns AI solutions that can interact on top of other systems, like a CRM/ERP system. AI accesses data flowing through these systems, fueling business improvements over time. In this business model, the company will charge you a monthly fee. Based on my experience, it is easier to make it work for NLP projects (chatbots, etc.)
Usually, such solutions are fast to deploy, so the sales cycle is quick, with a proven ROI. However, this business model is also fragile. If an AI solution doesn’t prove itself “essential” it will be vulnerable to budget cuts.
It depends on Data
Your next business model will highly depend on the data you can use.
Obviously, your ability to leverage data will impact your business model. Given that data can be copied (many projects use fake data at the beginning in PoC stage), and therefore isn’t intrinsically scarce, the value of a data ends up being pretty low overall in most industries, and is trending lower.
As a startup, you don’t need all the data in the world, just the data necessary to solve the specific problem you’re going after (as long as you’ve defined it precisely).
It also depends on the nature of your project, let’s imagine you want to build AI-powered drones. You will need to integrate hardware costs and other products related costs into your business model.
It is still much harder and time-consuming to build an AI-first startup than a regular SaaS startup. Indeed, data acquisition and AI training take a lot of time. Deploying TensorFlow still requires rare expertise. For all these reasons, the SaaS model can be complicated for AI startups.
I worked with several companies that couldn’t afford to build an AI solution from scratch. As a result, they have decided to partner with AI development startups specialized in tailored-made solutions. Through a revenue-sharing business model, both companies managed to find an interest. They both agree to build a PoC and if it works, they share the benefits. The only tricky question is who will provide the data necessary to build the PoC. Based on my experience, both companies are quite reluctant about spending too much time gathering data.
It works well for software companies trying to improve their existing solutions without spending too much on development costs. I can envision a future in which AI development teams could secure several revenue-sharing contracts and let other companies do all the commercial/marketing job.
The only bad aspect of this is that the other partner would want to add some kind of non-competitive clause. From a contract perspective, it becomes a bit complicated for both parties. Indeed, both have something to lose. The first one becomes heavily dependent on the AI company while the AI company cannot sell this same solution to another competitor.
I noticed that most AI development startups have a few paid pilots running and some early indications that customers will be ready to pay a significant amount of money once these pilots will be running in production.
As a result, they are always interested in securing new partnerships that could create new revenues. Especially, when they have already developed similar algorithms to answer your business issue.
SaaS vs AI
It might be tempting for AI startups to choose the SaaS model but for a number of reasons, this choice can be risky…
The very nature of AI solutions creates a situation in which usually specially annotated training data are mandatory as well as many different sources of data. As a result, you can’t limit an AI test to just a few users. Indeed, it will slow down the solution’s ability to adapt to the needs of your client. The more people using AI, the faster it learns.
For this reason, I have seen several startups using project managers to help the company understand the AI process and train them. Obviously, this additional resource has a cost that needs to be anticipated in the business model.
The value proposition of a traditional SaaS solution is often obvious within days of deploying the solution.
However, until an AI system goes through enough training data and is exposed to several use cases, it likely won’t perform any better than conventional software. Depending on your business issue, it is mostly once the solution is fully operational and after the improvement of its learning curve that the added value of this AI solution will seem obvious.
Because of this crucial element, it is hard for AI startups to use freemium models. AI requires more time than traditional solutions.
My latest AI project did earn revenue, but perhaps only 9% of the amount we spent on it. The numbers are a little better now as we cut spending while subscription revenue continued, but based on my experience, there’s no way an AI product is profitable at first launch.
AI as a Service
I noticed that most companies are using at least one type of “as a service”. Indeed, it enables them to focus on their core business and spend less money on an important service. It is obvious that there has been a paradigm shift in the way that businesses build their technology stacks in recent years driven also by a major move into digital platforms and microservices.
“as a service”: refers to any software that can be called upon across a network because it uses cloud computing.
The success of this solution can be explained by how easy it is to purchase such solutions. Indeed, in most cases, you can buy it from a third-party vendor, make a few modifications, and start using it nearly immediately.
For companies that can’t for budget reasons or are unwilling to build their AI solutions, AI-as-a-service is the perfect answer to not waste a possible opportunity. Like other “as a service” options, the same approach is applied to artificial intelligence.
Data is the driver behind machine learning.
In the coming year, businesses will rapidly begin to adopt machine learning as a service (AIaaS) into their technology stacks for several reasons.
From what I have seen (other business models do exist!), if we look at the bigger picture, global corporations can produce and have access to so much data but not necessarily those necessary to answer a specific business issue.
They are easily able to build and train their own machine learning models. This allows them to offer it to outside companies as MLaaS, the same way that since they have more data center space than smaller companies they can provide IaaS (Infrastructure).
In addition, you have AI development teams specialized in building tailored-made solutions for clients. These companies use the tools created by these global corporations in order to build these AI solutions for smaller companies. An entirely new ecosystem and business model is emerging.
Generally, smaller companies do not have access to as much data to create powerful AI models; however, they do have valuable and precise data (and excellent business knowledge) to start building a good dataset that can be used in an AI project.
Obviously, new AI business models might appear. I believe it will take time before AI startups find the “right” formula for AI success. It is important to choose a business model that enables your business to grow effectively.