22nd May 2023, Mechelen Belgium
Xdroid’s 2023 edition of Partner Days was built around the theme of “taking the next steps together” with the Partners. Along with their Partners, Xdroid aims to take significant steps towards revolutionizing the Customer Service and CX Industry. As a part of this journey, Xdroid announced that they have undertaken multiple projects to take the collective business forward and facilitate growth for all. One of the key projects in this next phase for Xdroid and its partners is the new Net Promoter Score (NPS) Prediction Model.
While NPS is a good indicator of customer satisfaction levels, the greatest challenge with its implementation is the low response rate from the customers. The average response rate for NPS surveys is roughly 30-40%. This means that out of the total customer interactions, only 30 to 40 per cent of the customers take the time to score the call. This further varies based on the platform used to collect the data. Furthermore, the data is mostly quantitative with little to no helpful feedback.
NPS also doesn’t factor in everything that goes into ensuring a successful call. External factors like the quality of the network are not considered. Often the agents are blamed for matters beyond their control and get stuck with low NPS. The responses, therefore, tend to be skewed, as customers are more likely to go out of their way to let the company know that they were displeased with the service than make the extra effort to compliment good customer service. This leads to a lack of accuracy in NPS.
Xdroid’s NPS Prediction Model
To tackle this inaccuracy, Xdroid has developed a predictive model for NPS that analyses every single aspect of the call and predicts what the NPS could be for that particular call. The model studies various factors including the customer’s sentiment at the beginning of the call and towards the end. The model does so by identifying the use of certain phrases through call tags, and sentiment tags.
One of the interesting features of the model, for instance, is analysing the evolution of customer sentiment from the first half of the call to the second half. So if an agent successfully calms a distressed customer and takes their sentiment on the call from anger (I am very disappointed, my product is still not here!) to gratitude for solving the problem (Thank you so much for addressing my issue!), the predictive model will predict a good score for that call.
The success of the model lies in how the final clients utilise it. The more they analyse, i.e. the more data they put in, the better the model will get. The clients have the freedom to fully customise the model and fine-tune it. They can also build the model to cater to their specific projects. Once the model is calibrated, clients will have their scorecards along with Zero-day KPIs for the model. They have the flexibility to check the call in totality and take both agent and customer experience into account.
The model categorises the customer as a detractor, neutral, or promoter, and scores the call on a scale of 1 to 10. Based on the analysed call, the agent also receives feedback on their performance. Factors that contributed to the success of the call are laid out to the clients for further review. The model analyses agent behaviour and categorises them into Good, bad, or neutral.
Additionally, the model highlights the stress levels of the customer to get their issue solved during the call. This involves checking for customer effort based on the amount of time and patience they had to devote to the call. This effort could be high, medium, or low. The clients, therefore, have access to everything that happened during the customer interaction and how it affected the success/failure of the call. They can then compare the actual NPS of the surveys to the predictions of the model, and find out what factors influence the customer satisfaction levels the most.
Call for Collaboration
Xdroid’s testing of the model with clients has been promising. The model was tested on 80,000 customer-rated calls and showed an accuracy of 82% in predicting the overall NPS, and an accuracy of 97% in predicting Promoters, Detractors, and Neutrals. The next phase of major testing through collaborations with partners and clients will focus on verification in multiple languages. They also have the opportunity to set up and test separate models for different categories.
