Many companies are adding some form of predictive analytics to their offers. This is true across all sorts of businesses, from medical device companies, to emergency management systems, to service aggregators and software infrastructure management platforms. All of these have one thing in common: they collect a lot of data in the normal course of operations.
At the same time, the rapid advances in machine learning since 2007 (for which Yoshua Bengio, Geoffrey Hinton and Yann LeCun won the Turing Award in 2018) made it much easier to develop these predictive learning applications. More data to feed better algorithms has opened many new applications and product managers have rushed to take advantage of this. But how do you package and price this new functionality?
We have seen many different approaches.
- Offer predictive analytics as part of professional services
- Create an independent product with its own pricing
- Have predictive analytics as an optional add on to the existing product
- Make predictive analytics part of the standard offer
- Use predictive analytics internally to increase engagement and reduce churn
Machine learning itself is a commodity and will be priced as such
There is a lot of confusion about what approach to take and how to price. Fortunately, a standard value-based pricing approach can clarify this. Before we try this, what is the basic change that machine learning has made possible and what will that mean for how we create value and set pricing strategies?
In their 2018 book Prediction Machines, authors Ajay Agrawal, Joshua Gans and Avi Goldfarb argue that machine learning AIs (there are other important approaches to AI such as the semantic AI) make prediction cheap. The first generation of computing made computing cheap: can you imagine managing even one of your simple spreadsheets by hand, or constructing a multipart query without a database? The deep learning generation makes prediction cheap.
How do computing and prediction differ? Simply put, computing rearranges existing numbers (and things that can be represented as numbers like words and images): prediction forecasts, probabilistically, what future numbers might be. Machine learning became a commodity in record time. Good machine learning platforms are now part of all of the standard cloud platforms such as Microsoft Azure and Amazon Web services, and there are excellent open source code bases. Machine learning is a commodity, like relational databases or electricity. It is widely available and many people have the technical skills to apply it. The best scientists and engineers are mostly working on the big prediction engines that the rest of us use. The question is, use to what purpose?
What are the unique predictions from your proprietary data?
The main driver of differentiation is not machine learning itself, but the data that feeds it. So the first step is to ask what predictions can be made using the proprietary data that your application collects and that others do not have access to. This is especially powerful when data aggregated across customers increases the overall predictive power. When this is the case, you can make better predictions than your client can relying on their own data (but see the caveat below). If the predictions are something important to their business, then that has value and you should develop a solution and figure out how to price it.
Let’s take a simple example. Many companies today, in the midst of the Covid-19 crisis, want to know how well people are going to do working from home. There are many different applications that might give insight into this. Virtual meeting applications (such as Zoom) might be able to track focus while in meetings, which may have predictive value. Communication platforms (such as Slack) could monitor communication patterns for engagement and contributions. Performance management systems (such as Seven Geese) could be used to assess how people are feeling and make predictions based on that data. Project management platforms (such as Asana, Basecamp or Wrike) could track changes in contributions and task completion pre and post Covid-19, or see if task structures have changed. Skill and competency management platforms could see if team skills have changed and link changing skill patterns to outcomes.
Looking through all these different applications that might be able to make predictions about how different people will respond to working from home opens a new question. What if the best prediction engine requires data from multiple applications?
This sets up the new competitive dynamics in predictive analytics. Where will machine learning algorithms with cross customer data outperform those with cross application data and vice versa? Will a third class of applications that combine both value propositions emerge? Will this favor platform vendors over best of breed vendors? What role will the cloud platforms play in this? Will they have the market muscle to force data sharing to feed their own AI engines? There is much to sort out over the coming decade.
You can’t wait on this though. If you are not developing solutions that leverage predictive analytics someone adjacent to you is and they will rapidly sideline you. They may even offer what you regard as standard functionality for free in order to get their hands on the data. So you need to move now, but how will you capture enough value to justify your investments?
A simplified decision tree for pricing predictive analytics
Below is a simplified decision tree to help you decide if and how to price predictive analytics. To use this decision tree, you will need to have a well-defined market segmentation model and to understand your economic, emotional and community value drivers.
There are three basic questions to answer.
- Are the predictive analytics of value to your current customers or to a new set of customers?
This is the market segmentation question. In most cases, you will have to research the value drivers for your predictive analytics and conduct a new market segmentation exercise to answer this question.
- Do the predictive analytics support new use cases or existing use cases?
If new use cases, then you are looking at a new module, that may well have new value metrics and pricing metrics. If existing use cases, then you need to look at the value drivers again.
- Do the predictive analytics create value for your customers by enhancing existing value drivers or by creating new value drivers?
Innovations that enhance existing value drivers for your current customers are sustaining (or incremental) product enhancements no matter how cutting edge the technology used may be. When this is the case, you may want to adjust pricing, but you will not need to change your existing pricing model. If new value drivers come into play, then you should consider a new pricing model. Good value-based pricing requires that value metrics (the unit of consumption by which your customer gets value) and pricing metrics (the unit of consumption by which you price) be connected.
When not to price predictive analytics
There are three cases where you may want to consider not charging for predictive analytics.
One is when the predictive analytics will have a meaningful impact on user engagement and will lead to increased use or reduced churn. For many companies this is a natural place to start. One can even build a system that predicts future engagement from current engagement, Predictive e so to speak. This can be a critical customer success tool.
Predictive analytics can also give you critical insights into how to create additional value for your customers. Creating exceptional value to customer (V2C) is core to building the long-term value of your company. You may be able to create more value by giving away predictive analytics and using the data to accelerate your own value delivery.
Going back up to the decision tree, in some cases the predictions are more valuable to a new set of customers than they are to your current customers. This is how two-sided or platform markets emerge. A company in the precision agriculture space, for example, may find that insurance or trading companies find its predictions on crop yield even more valuable than do farmers. The alternative market can be so valuable that it is worth giving away the service in order to gather the data that is then sold on. This is now most social media platforms work.
A word on data rights
What rights do you have to data collected by your platform? Most companies do not really understand this. Contracts and end user license agreements were developed before data became the core asset of a software platform and rights tend to be poorly defined. Many companies have used privacy to frame this issue when intellectual property rights and data ownership provide a better framework. When rights are too broad, there is a risk that courts will reject them. So the first step in defining your strategy around AI, machine learning and predictive analytics is going to be a visit to your legal department and in all likelihood some hard conversations with your existing customers.