Maximize sales efforts by using AI for identifying Upsell opportunities

Upselling is a sales technique used to get customers to upgrade to a higher tier or a better product. By using modern AI techniques, sales departments can be helped on where to best put their resources to get as much out of their efforts as possible. In this article, we will go into detail on how to build your own upselling model on the Peltarion platform.

... where do we place our bets?

02/ The problem

Businesses have limited resources and knowing where and to whom the upselling should target is key to using resources efficiently. Not only will it make the sales department more efficient, it will also lead to happier customers, where each customer gets the product that is best suited to their needs. Current manual upselling techniques are not data-driven and as more and more companies start to use these AI tools for sales, the manual techniques will be even more obsolete.

03/ The opportunity for deep learning

As sales today happen totally digital or at least involving digital interactions, many data points are collected in an automated way. And many different variables and data points are key to a problem where AI could be the solution. Not only can AI be used to automate tasks that a human could do, it can also help to find patterns that would not be visible for the human eye.

04/ AI model to use

For this use case, our model of choice is: Tabular data classification

05/ How does the model work?

From our tabular data, we need to pick out some columns that we know about the customer, the product or the context. These columns will be used as input and are commonly called input features. We then need to pick one column that is the one the model is going to predict, that’s the label feature in the training data set.

What columns the tabular data should contain is depending on what is accessible and a trial-and-error of what is working by trying different variations on the Peltarion platform. It is also useful to be creative when thinking about what features can be used.

An example would be that the tabular data consists of one row per customer. The features could be divided into the ones about the customer, the ones related to how the customer uses the product, and the miscellaneous ones. Customer features could be what kind of user the customer is (novice or advanced), how experienced the person is in another domain, or where the person is located. The customer with product features could be how many days since the customer signed up, how often the customer has interacted with the product for the last month, how many interactions the last week, what parts of the product has the user used, have the company had any other interactions with the customer and more. The miscellaneous features are the hardest to find, but could be things as date, is it during the summer or is it a Friday evening.

One column in the tabular data should be the one to predict. A good label for this case would be a column that says yes or no depending if the user said yes to an upgrade within one month. This column, based on the other features, would be the one the model could then predict.

06/ Data requirements

To succeed we need historical data about our customers in a tabular format like CSV. How many data points or rows that are necessary is hard to know without trying, because it depends on the task and how diverse the data points are. A good starting point is to have a table with at least 100 data points.

07/ Model results and success

You will see the model performance in the evaluation view on the platform after you have run your first experiment. In these kinds of experiments, it is important to look at the model accuracy as well as the amount of false positives and false negatives. You will learn more about this in the tutorial under useful resources below and if you encounter any obstacles along the way, we at Peltarion are always here to support you.

08/ Where to learn more?

If you want to build this yourself with the Peltarion platform. A suggestion would be to follow this tutorial on how to predict with tabular data. And then use your own dataset in a similar way as described. The tutorial also goes into the details on how to measure model performance and what to do to improve it even further. Good luck!