“Artificial intelligence is that activity devoted to making machines intelligent, and intelligence is that quality that enables an entity to function appropriately and with foresight in its environment.”
Without wanting in any way to dismiss AI, which can be incredibly useful in many different arenas like driverless cars, we believe that its role for marketers has been exaggerated. Indeed most of what we need is barely within the spectrum of technology normally defined as AI.
For most cases where marketers want to execute one to one personalisation, the area where AI could most appropriately be applied, the conventional propensity model is all that is required.
What is most often meant by personalisation is the means to carry out selections of customers for communications based on their expected response or their particular needs.
Here are some examples where personalisation is often used:
– Targeting apparently dormant customers (e.g. those who in fact have a high probability of being reactivated) with offers to reactivate them
– Making a relevant offer (e.g. based on customer characteristics that imply a higher than average probability of purchasing in a particular product category) of a specific item
– Responding to risk (e.g. predicting which customers are likely to cancel policies or stop ordering) so that they can be presented with good reasons not to abandon their policy or purchase
In each case a conventional predictive model can be built, using an historic set of customer data, where a target customer population can be distinguished from the remainder who have not evidenced reactivation, response, or reduced risk of lapsing.
The key point is that we are not asking for this kind of model to be adaptive to rapidly changing circumstances; instead it relies on past customer behaviour to inform what is likely to happen in the present or near future. And this is because human behaviour in most situations where we are reacting to propositions put to us by marketers tends to remain reasonably constant.
We have even tested propensity models on historic data going back four years and found them to work well.
However, to build and apply these conventional propensity models there are some essential requirements:
– a single customer view to provide the greatest possible depth of customer data
– the ability to update model scores each time new data about an individual arrives
– the availability of data scientists armed with tools like R, SPSS, or SAS
A typical predictive model will take the form of an algorithm which will attribute a probability score to each member of a customer base; we judge the success of these models by the extent to which these scores are differentiated from random in the way they can be used to predict customers’ behaviour.
Looking at a recent model we built for the reactivation of dormant customers, the top customer decile had an index of 330,compared to the bottom decile’s 17.
In another case, a model for product category preference had a top decile index of 601 and bottom decile index of 11.
For most of us marketers these results will be seen as providing a huge improvement on random and quite fit for purpose. However, the methodology used does not in our opinion qualify the models to be correctly described as AI.
If you would like to talk to us about developing propensity models for you, or providing the technology for a single customer view, then please do email us back
UniFida is the trading name of Marketing Planning Services Ltd, a London based technology and data science company set up in 2014. Our overall aim is to help organisations build more customer value at less marketing cost.
Our technology focus has been to develop UniFida. Our data science business comes both from existing users of UniFida, and from clients looking to us to solve their more complex data related marketing questions.
Marketing is changing at an explosive speed, and our ambition is to help our clients stay empowered and ahead in this challenging environment.