Predictive Applications are everywhere (and they are hardly being noticed). In one of our previous blog posts, I showed that more than half of a typical early adopter's home screen is already covered by predictive applications. These predictive applications collect data, make predictions (hence the name) and ultimately drive decisions, for instance by recommending content, or shaping the customer experience. Yet, in the examples seen so far, we have only looked at consumer predictive applications.
This is no surprise, as consumer applications are far more visible, and given the hidden nature of predictive applications (most are hard to spot, even if they are right in the palm of your hand) it takes a lot of experience to uncover predictive applications in the wild.
Yet the potential for predictive applications is actually much greater in the enterprise software space than in the consumer software space, for one thing the number of consequential decisions is greater when it comes to handling processes like supply chain management, human resources, pricing, inventory maintenance and so on. Many of these decisions today are based either on ad-hoc human decision making (with all its negative consequences) or on rigid rule systems that fail to adapt to the market.
And, despite the many technological and conceptual similarities between consumer and enterprise predictive applications, like being based on data, making predictions and driving decisions, there are a couple of key differences that show those for enterprises need special treatment:
- Enterprise predictive applications arise from existing business processes: Many consumer predictive applications are built from scratch to deliver an entirely new consumer experience. This is great, as it allows you to start from a blank slate, but requires that all data needs to be collected within the application – and it increases the hurdle to making decisions with substantial impact. Enterprise predictive applications on the other hand are introduced to increase the efficiency (usually in the form of more, cheaper and faster decisions) and the effectiveness (in the form of better decisions with fewer mistakes) of an existing business process. When predictive applications are used to automate replenishment of perishable food, we don't re-invent the replenishment process, we integrate it into an existing process with existing systems, making it faster, cheaper and better.
- Enterprise predictive applications deal with higher domain complexity: Because they are based on an existing business process, enterprise predictive applications have to be able to understand the full domain model that is the basis of the final decision. This not only includes existing business rules (many of which will be made obsolete by the predictive application) but also the entities involved and their relationships, as well as the goals and constraints of the organization as a whole (usually described in a cost function). By being able to understand, express and optimize the complexity of the domain, decisions can be made within the context of the domain instead of a vacuum or tabula rasa. Many predictive applications for consumers, on the other hand, rely on a simplified domain model of consumer and behavioral data.
- Enterprise predictive applications require fewer PII and external data: When implementing enterprise predictive applications, two questions need to be answered: “What is the most substantial decision that needs to be made?” and “What data is needed to make these decisions on a statistically sound basis”. The first question can be answered by looking at the problem domain alone, but the second question requires analysis and experimentation with customer data. In our work with customers, we found the surprising result that the two types of data most frequently associated with Big Data and concerns about privacy for Big Data applications are often the least valuable. Instead of using customer demographic, behavioral and publicly accessible or third-party-owned social data, which is available in huge volumes, albeit with low predictive power, the most impactful predictive applications in the enterprise use the customer's own data.
- Enterprise predictive applications require higher model complexity: Due to the nature of enterprise predictive applications, with higher domain and data complexity another factor comes into play: enterprise predictive applications have a more complex underlying machine learning model. Instead of just being able to use the largely domain-independent disciplines of machine learning such as sentiment analysis, named entity recognition, face detection or image recognition, enterprise predictive applications work with models that are fine-tuned to the decision at hand and require techniques such as probability distribution function, custom aggregations and optimization of probability-adjusted cost functions.
- Enterprise predictive applications have more challenging data availability needs: As we have seen before, consumer predictive applications typically start with relatively simple domain and data models, build their own data collections from scratch and gather vast amounts of new data to run their models on. For a predictive application in the enterprise this approach isn't feasible. Instead, existing data needs to be found, extracted, transformed and integrated into the predictive application. This is a process that is being aided by data integration software such as ETL tools and standardized APIs for predictive applications, but it is still an important and necessary part of every predictive application rollout.
Each of these aspects has been informing the design choices in our Blue Yonder Platform when it comes to storing and linking data, building and running predictive applications and finally when it comes to monitoring and operating them.
If you want to learn more about how predictive applications can help your business, get in touch with Blue Yonder to talk about predictive use cases in your industry. And follow us on Twitter, so you don't miss any updates.