Hi there! My name is Isak and I work for Peltarion in Stockholm, Sweden. In this post we’ll demonstrate how you can act upon negative customer experiences and try to save your NPS score with operational AI and automation, using the Peltarion Platform and Zapier - an automation tool requiring no code at all, that we’ve supplied with a Peltarion connector.
Save your NPS with operational AI
03/ Summary of the project
05/ The project
The good news is that Peltarion provides an easy solution for letting machines (instead of humans) read, understand and decide how to act upon receiving customer feedback. This could make life easier for lots of people. Meanwhile it does of course require some amount of manual setup, which is what we’re going to look at during the next couple of minutes.
I’ll start out introducing what we’ll build in this demo, then we’ll build it, and finally we’ll wrap up and tell you how to get started.
We will build a simple app that, given free-text customer reviews, uses deep learning to predict whether or not the customer would recommend a given article. Based on prediction results, the app will connect with potentially dissatisfied customers and try to win them back as fast as possible, by sending them a message with a promo code. Now oftentimes it is not that simple, this flow is mainly for simplicity and it is a good start. If you’d like to make a more complex automation flow, everything is possible, it will just take a bit longer to build. When we’re done building, everything will be running autonomously using Google Sheets Peltarion and Zapier. Note that we will not actually predict or calculate NPS scores, but act on caught detractors in order to push the NPS score up.
Before we start the project we need to clarify three things. First, the dataset we’re going to use. Then, the problem type in terms of machine learning, and finally based on that, decide what model we could use to attack the problem.
We’ll use a dataset found on Kaggle / the Datafiniti API. The dataset consists of Amazon Kindle Reviews. It has a free text review label, and a DoRecommend-label that tells if the customer would/would not recommend a given product. In other words, it fits our project very well. Here's the dataset.
Regarding the problem type, classifying text reviews and mapping them to a binary value (recommendation or not) is what one would call a binary text classification problem.
We’re going to classify a binary value based on free text. Fortunately there’s a perfectly suitable deep learning model pre-implemented on the Peltarion platform - the English Bert.
Now that we have what we need (dataset, problem type, model), let's go ahead and build our project. First we’ll build, train, evaluate and deploy the model with the Peltarion Platform. Then we’ll connect the model to Zapier and set up our simple automation flow with Google Sheets, Peltarion and Gmail.
Whenever you’re ready - feel free to start the 12 min video. Good luck!