Click to learn more about author Jonathan Lee.
In my last column for DATAVERSITY I sought to define the concept of a Customer Data Platform (CDP), noting that a CDP “creates a comprehensive view of each customer by capturing data from multiple systems.” What I didn’t do however, was to delve into a discussion of the data sources a CDP relies on. While a full discussion is outside of the scope of any single article, I’d like to address one of the most important sources of data: customer feedback.
I was recently part of an internal discussion with the goal of quantifying how customers felt about the products and services they were receiving. As we continued to discuss a wide variety of different metrics and the conversation dragged on, one of the analysts began to lose patience and blurted out “why don’t we just ask them?” As it turns out, that’s a great idea! However, the challenge for any organization once they reach a certain size is that reaching out to every customer and processing all of their feedback becomes hard to scale, there are simply too many customers and too few support staff. Enter Natural Language Processing (NLP).
What is NLP?
DATAVERSITY defines NLP as “a branch of Artificial Intelligence (AI) that automates language recognition and generation so that computers and humans can communicate seamlessly.” In practice, this means running algorithms on customer (and prospect) communications, such as email exchanges with sales, forum posts, conversations with chatbots, etc. The goal is to determine how customers are feeling, without asking them directly, so that businesses can react. For example, if customers are unhappy, customer success teams can receive automated alerts to reach out and rectify outstanding issues, thereby potentially saving an account. If customers are happy, sales can be alerted in order to pitch upsell/cross-sell opportunities, thereby generating more revenue.
Even if you plan to use a simple rating-style survey with customers, NLP can be leveraged to more easily identify their exact pain points. Yelp, for example, uses a zero to five-star rating system, along with text. As a user of the platform, understanding why a lower rating was given is often more important than the rating itself. For example, if I do not care about the ambiance or music played at a restaurant, yet many users rated a restaurant poorly due to these issues, I may choose to ignore the negative reviews and still enjoy the restaurant. NLP can be used to scrape the user’s feedback along with the rating to determine the cause of the poor ratings. In this (not-so) hypothetical case study, using this technology to automatically explain the reason behind ratings can provide a much better experience for the customer.
From NLP to CDP
These are not new ideas. Much has already been written about using NLP to understand customer feedback and even examples of use cases within specific industries. The challenge is that for the results of NLP to have any real value to business teams like sales and customer success, they need to be integrated into business-friendly data store, of which a CDP is the best choice. However, according to David Raab, founder of the CDP Institute, “probably the biggest challenge [for CDPs] is dealing with unstructured and semi-structured data.” Unfortunately, little of the data generated by an NLP system is likely to be in structured format. As anyone who’s spent much time reading Yelp reviews can tell you, not all customers write grammatical sentences without typos. A large portion of an analyst/data scientist’s time is therefore spent on data cleansing and preparation. This includes splitting sentences/words/phrases, recognizing named entities and removing tenses so that the data is normalized (Woohoo, grammar!). This must be done in a way that’s accurate for multiple text sources and types, hence the importance of good Data Governance practices to ensure the data that ultimately gets deposited in your CDP is usable for analysis.
Beyond data cleansing, an analyst is going to need to determine which texts and words are valuable for evaluating customer sentiment. From there, analysts and data scientists can use all the algorithms they know to analyze this data and parameterize it as machine learning inputs. Machine learning models are then trained and, after multiple iterations, form a working model. But even after models are trained and fine-tuned, there will always be a need for constant monitoring since models may become inaccurate over time. The pipeline is never-ending.
What’s the Result?
Analysts and data scientists are some of your most limited and expensive staff, so you need to use their time wisely. However, the results of an effective NLP-to-CDP program are well worth it because they allow large businesses to do something they’ve never been able to in the past: scale customer feedback in a cost-effective manner. With such a program in place executives can essentially ask each customer “how are we doing?” If the results are good they can be used to find cross/up-sell opportunities. If they’re not, the business can take concrete steps to address issues before customers churn. Either way, the value is clear.