24 May Learning to personalise through AI experimentation
The Facebook/Cambridge Analytica scandal is not the easiest backdrop against which to write a blog post espousing the benefits of micro-personalising customer content. It does, however, demonstrate that if companies want to justify the data they collect and hold, they need to use it in a way that tangibly benefits their customers. The best way to achieve this is to improve customer experience. Common approaches to engaging customers in conversations, such as cold-calling or sending a barrage of text messages, both frustrate customers and produce poor results. ContactEngine is using demographic data and machine learning to personalise the approach, so everything we say is relevant and convenient for each customer.
We currently work with several large energy suppliers to approach their customers about installing a smart meter. By learning from previous interactions and using demographic information, we engage more people in conversations with fewer contact attempts, both saving money for our clients and making the process frictionless for customers. Here’s a few things we have learned through these projects.
We began by building a model to predict how a customer will respond to an offer to upgrade to a smart meter. Although we only categorised three responses (Yes, No or Non-response), it was enough to create an interesting level of complexity. Some demographic attributes might be associated with a higher likelihood to say yes, but also a lower response rate; other attributes might be associated with a high level of scepticism about smart meters, but also a willingness to respond. It is difficult for humans to reason about overlapping effects, but machine learning algorithms can help us here. We found that by using principal components methods, a mathematical technique to identify the dimensions along which we see the most variation in our data, we were able to identify the axes along which demographic information most affects response patterns, allowing us to create geometric visualisations which can be understood by non-statisticians. Our statistical models, in this case neural networks, were able to capture the interactions between demographic attributes and learn the features which most influence how people respond.
From prediction to decision…
Our modelling, which has been carefully validated, has demonstrated the ability to predict how people respond. But what can we do with this knowledge? A common idea is customer segmentation: dividing the customer base into clusters which are then assigned different contact strategies. This has a few problems. Firstly, ‘natural’ clusters need not exist; there is no law that says your customers will fall into a few easily definable groups. Here, by natural clusters I mean clusters which stand out within the data, and would be identified by anyone analysing the data, without them having to make the same arbitrary initial decisions or share the same set of biases. Secondly, there is a great deal of guesswork involved in deciding how to approach each customer segment, since they have been formed exclusively by considering demographics with no reference to communication strategy.
Instead, because ContactEngine is configurable, we can quickly set up and monitor experiments in which we adapt our communication strategies; changing the timings, channels and content of messages. By combining this empirical approach with demographic data, we can establish the right way to contact each person, without making assumptions based on discrete clustering or relying on unjustified stereotypes.
The key message is that data is valuable if you want frictionless conversations with your customers, but it works best if its combined with a platform and a mindset that allows you to experiment and learn.