Expert Commentary

Sidestepping Survey Fatigue, Predictive NPS Gets to the Heart of Customers’ Attitudes

Models can infer whether a customer is delighted or annoyed with a company by analyzing everything known about that person.

By Ilker Carikcioglu

First published on enero 28, 2021
min read


Net Promoter Score℠ is a proven measure of customer advocacy and loyalty, used by two-thirds of the Fortune 1000 firms. Its global adoption rests on many factors, including its simple question: “How likely would you be to recommend [company/brand] to a friend or colleague?” This short question is the backbone of NPS^® and an integral part of most customer surveys. Leading organizations elicit valuable insights from these studies and use them to create superior customer experiences. In recent years, however, survey response rates have been declining. Customers are bombarded with requests for feedback as companies vie for attention. Any given survey now might have response rates in the low single digits. That means a company is able to collect explicit feedback from only a small sample of its customer base, leaving it virtually blind as to what the majority of its customers perceive. However, we know that many companies possess vast amounts of customer data, other than explicit feedback. Enter Predictive NPS. The idea is simple: Predictive NPS infers a customer’s NPS status (promoter, passive, or detractor) even in the absence of a survey, by using advanced analytics tools to process and model all structured and unstructured data pertaining to that customer. But it takes a meticulous and organized effort to bring together the required elements: vision, data, technology, talent, and a commitment to act. Can we infer whether a customer is delighted or annoyed with a company by analyzing everything we know about that customer, such as her demographic profile, product ownership, transaction history, interactions with the company, and other behavioral indicators? A variety of models we have built in recent months confirm that Predictive NPS does this quite well. As just one example, we worked with a financial services firm to build a model that predicts the likely outcome of a telephone call placed to the customer service center. We gathered operational metrics surrounding the call (such as hold time, talk time, and number of transfers), the customer’s recent transactions, and her digital footprint. In addition, we deployed advanced natural language processing algorithms to analyze the call transcript to extract call reason, intent, and sentiment. Finally, we fed these sets of structured and unstructured data into a number of machine learning algorithms, using labeled NPS data as the target variable. The results: an overall accuracy above 80% for predicting NPS status, and a fourfold lift above the baseline in identifying detractors. We also identified key drivers for each individual prediction via techniques such as Local Interpretable Model-Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). These local model interpretability methods informed more tailored actions at the individual customer level, including targeted callbacks by frontline managers. Predictive NPS has several benefits that help the process of building better relationships with customers. It gives a company the ability to monitor the pulse of a majority of customers, rather than the sliver who offer explicit feedback. It equips customer experience teams with the information and tools to intervene and improve customer outcomes in near real time. It allows managers to anticipate, rather than react to, customer behaviors. With customer expectations changing rapidly, Predictive NPS could point the way to superior customer experiences, reduced operational cost, and competitive differentiation. Net Promoter Score℠ is a service mark, and NPS^® a registered trademark, of Bain & Company, Inc., Satmetrix Systems, Inc., and Fred Reichheld.

Net Promoter Score℠ is a proven measure of customer advocacy and loyalty, used by two-thirds of the Fortune 1000 firms. Its global adoption rests on many factors, including its simple question: “How likely would you be to recommend [company/brand] to a friend or colleague?” This short question is the backbone of NPS^® and an integral part of most customer surveys. Leading organizations elicit valuable insights from these studies and use them to create superior customer experiences.

In recent years, however, survey response rates have been declining. Customers are bombarded with requests for feedback as companies vie for attention. Any given survey now might have response rates in the low single digits. That means a company is able to collect explicit feedback from only a small sample of its customer base, leaving it virtually blind as to what the majority of its customers perceive. However, we know that many companies possess vast amounts of customer data, other than explicit feedback.

Enter Predictive NPS. The idea is simple: Predictive NPS infers a customer’s NPS status (promoter, passive, or detractor) even in the absence of a survey, by using advanced analytics tools to process and model all structured and unstructured data pertaining to that customer. But it takes a meticulous and organized effort to bring together the required elements: vision, data, technology, talent, and a commitment to act.

Can we infer whether a customer is delighted or annoyed with a company by analyzing everything we know about that customer, such as her demographic profile, product ownership, transaction history, interactions with the company, and other behavioral indicators? A variety of models we have built in recent months confirm that Predictive NPS does this quite well.

As just one example, we worked with a financial services firm to build a model that predicts the likely outcome of a telephone call placed to the customer service center. We gathered operational metrics surrounding the call (such as hold time, talk time, and number of transfers), the customer’s recent transactions, and her digital footprint. In addition, we deployed advanced natural language processing algorithms to analyze the call transcript to extract call reason, intent, and sentiment. Finally, we fed these sets of structured and unstructured data into a number of machine learning algorithms, using labeled NPS data as the target variable.

The results: an overall accuracy above 80% for predicting NPS status, and a fourfold lift above the baseline in identifying detractors. We also identified key drivers for each individual prediction via techniques such as Local Interpretable Model-Agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP). These local model interpretability methods informed more tailored actions at the individual customer level, including targeted callbacks by frontline managers.

Predictive NPS has several benefits that help the process of building better relationships with customers. It gives a company the ability to monitor the pulse of a majority of customers, rather than the sliver who offer explicit feedback. It equips customer experience teams with the information and tools to intervene and improve customer outcomes in near real time. It allows managers to anticipate, rather than react to, customer behaviors. With customer expectations changing rapidly, Predictive NPS could point the way to superior customer experiences, reduced operational cost, and competitive differentiation.

Net Promoter Score℠ is a service mark, and NPS^® a registered trademark, of Bain & Company, Inc., Satmetrix Systems, Inc., and Fred Reichheld.