A case study of C3.ai
I have been waiting for a pure-play artificial intelligence (AI) company to file for an IPO to get an in depth look at the company’s business model and how varied it is to traditional software products. My interest lies in trying to understand the impact of the long tail of AI on the company’s business model and financials. I had written about this topic earlier in a couple of previous posts. The first offered a glimpse on how building AI products is fundamentally different than traditional software. The second focused in part of the complexities of adapting trained AI models and making them generalize on customer data, a phenomena I called hill climbing.
Well, a few days ago, I learned about C3.ai, which recently filed for a ~$100M IPO. C3.ai dubs itself as “an Enterprise AI software company”. The company provides a suite of SaaS based AI applications. Which can be deployed on all major clouds : Azure, AWS, Google Cloud Platform and IBM’s cloud. They offer their products on-premises as well.
C3 offers a core suite of products which they call the C3 AI Applications. These applications are built on top of the core AI Suite, which is “our core technology, is a comprehensive application development and runtime environment that is designed to allow our customers to rapidly design, develop, and deploy Enterprise AI applications of any type.”
In summary C3 provides a platform — C3 AI Suite — that allows their customers to build and deploy AI applications on. They also utilize this platform for a set of products — C3 AI Applications — that they offer as well to their customers. The diagram below depicts how both product suites interact and are part of C3 AI platform.
C3’s revenues are broken down into two components: the sale of software subscriptions and more importantly professional services. Software subscriptions represented 88% of their total revenue in 2020 with the balance derived from their professional services. Furthermore, the company’s revenue, which is shown in the table below, was derived from only 29 customers.
The professional services portion of their revenue is one I was keen to explore. My motivation being a hypothesis I have that AI companies will require to offer products and services. Professional services for an AI company is a necessity and not an option.
“We also generate revenue from professional services, which consist primarily of fees associated with our implementation services for new customer deployments of C3 AI Applications. Professional services revenue represented 15%, 14%, 11%, and 12% of total revenue for the fiscal years ended April 30, 2019 and 2020 and the three months ended July 31, 2019 and 2020, respectively. Our professional services are provided both onsite and remotely, and can include training, application design, project management, system design, data modeling, data integration, application design, development support, data science, and application and C3 AI Suite administration support.” Source: C3.ai S1
The implications of requiring professional services for most it not all sales are obvious. Gross margins will be lower due to the PS revenue and sales cycles will be long. To validate my hypothesis I compared C3’s gross margins to a basket of recently IPOed software companies. I included Palantir, which I suspect is closest to C3 from a business model. Palantir’s Forward Deployed Engineers are analogous to professional services. I also included Accenture to represent a pure-play professional-services (consulting) like company.
One can clearly see that C3’s margins are almost identical to those of Palantir and that both are notably lower than software companies (Snowflake is an interesting exception). More importantly, both Palantir and C3 have margins that are substantially higher than Accenture. That makes sense, they both sell software, in spite of this software being accompanied with professional services engagements, while Accenture practically doesn’t sell any software. This shouldn’t come as a surprise and is aligned with a recent article by Martin Casado & Matt Bornstein both of Andreessen Horowitz. They present the following argument in their article.
My next hypothesis was on the nature of what I call go-to-production (GTP) cycles for AI companies. I suspected that they are longer than traditional software companies. That, admittedly, is harder ascertain since most companies don’t publicly disclose this data. However, an example might illustrate this point.
Supposing that your company wanted to use Zoom, or a similar SaaS like software offering. One way to do so is to head to Zoom’s website and with a few clicks of a button you’d be up and running. There’s very little in terms of customizations or adapting Zoom to your company’s environment. Your go-to-production cycle is measured in hours, perhaps a few days.
That’s not the case with AI, which requires a lot of customization and adaptation to the customer’s environment. This adaptation mostly arises due to machine learning models not being able to fully generalize on a wide distribution of data and therefore these models would have to be trained (calibrated) on the customer’s data. This process allows the model to learn from the customer’s data and thus able to perform (predict) better. I covered this topic in an earlier article here.
C3’s S1 provides several data points on the impact of this customization, which you see reflected in the long time to deploy their products. I suspect that another reason for prolonged deployments is customers using C3 Core AI to build their own ML models, which is time consuming. Below is an excerpt on this from C3’s S1. Bold emphasis mine.
“Our typical sales cycle begins with one or more product and technical presentations about C3.ai, leading to a mapping of our capabilities to customer use cases. This frequently leads to a paid trial that lasts from five to 16 weeks. During that period, we deploy a production-level application that is representative of our customer’s AI and digital transformation requirements. Examples include: Stochastic Optimization of the Supply Chain, Production Optimization, Fraud Detection, and Predictive Maintenance. After completing a successful trial, our customers will frequently license one or more C3 AI Applications. Either concurrent with or subsequent to licensing C3 AI Applications, our customers will often license”
C3’s S1 offers additional proof on the need to retrain their models making them ever so adaptable to new data. Bold emphasis mine.
“Areas of Judgment and Estimates. Determining whether the software subscriptions and the related support are considered distinct performance obligations that should be accounted for separately or as a single performance obligation requires significant judgment. In reaching its conclusion, we considered the nature of our promise to provide the customer real time analytics and machine learning algorithms that require regular re-training to maintain and improve prediction accuracy. “
The net effect of this, along with the need to get access to customers’ data — always a messy business — and helping customers build their own ML models, for those who use C3 Core AI could explain the lengthy deployment cycles, which are illustrated below.
I am keen to continue exploring whether these hypotheses hold true for other AI companies. I suspect they will. In the meantime, it would appear that if you are in the business of building AI products then offering services isn’t an option. AI companies are a blend of traditional software and services companies.
Originally published at https://karimfanous.substack.com.