This is part 2 of the blog series, "Data Science Use Cases in Insurance."

The insurance industry is one of the most dynamic and growing industries of the 21st century. To stay competitive and relevant, insurers are investing in technology to enhance customer retention and engagement.

In part one of this blog series, we focused on "Why Data Science is a Game Changer for the Insurance Industry Today and Tomorrow.” This article will focus on how AI tools such as NLP (natural language processing) and text mining are used to enhance customer engagement.

How Insurers Interact with Customers in 2021

There are two ways in which an insurer interacts with customers in today's world of technology:

• Leveraging connected homes to offer additional policy discounts and proactive risk management services. This is done by collecting data on their behavior and then attempt to reward those behaviors that support risk reduction and prevention
• Collecting the customer's feedback through explicit channels or by the customer's comments on various social media platforms or media reports about the insurance company or insurance industry in general

This data helps the insurance company to understand their customers' lifestyles and create customized coverage options to suit their needs.

Insurance companies collect text-based data in many languages and dialects from an array of sources such as:

• Applications
• Claims
• Email
• Social media
• Adjuster notes
• Medical records
• Police statements
• Surveys
• Research studies
• Underwriter notes
• Competitive intelligence

This vast quantity of unstructured and often conversational data piles up and sometimes gets forgotten. While individual employees deal with the data that crosses their desks, until recently, very few companies have been able to look at the data as a whole. We can observe trends, spot new topics, identify potential problems, including individual and group fraud, or flag new business indicators when reviewed holistically.

Below are four use cases that utilize AI tools such as NLP (natural language processing) and text mining for enhancing customer retention and engagement.

#1 CUSTOMER DEDUPLICATION

The first step for using policy information to understand the customer experience, insurers need to match an individual customer to the correct policy or insured person. An insurer's database contains information about policyholders, insured individuals, and nominees. To keep track of individual customers, the insurer needs to create a customer ID and attach it with every role in the policy. A strong customer ID is also important for fraud prevention and risk management purposes – understanding the insurer's exposure to a single individual. While every insurer has a mechanism for customer deduplication, these mechanisms find it difficult to address some data issues. Here are some issues faced by insurers when creating customer IDs:

• One of the key data fields for identifying an individual person is their name. And names may not always be written with exactly the same spelling across different policies. There may be phonetic spelling variations, abbreviations, order changes (“first name, last name” vs. “last name, first name”), missing middle names, and spelling mistakes. These data issues make it difficult for the insurer's customer ID creation system to match the same individual's records. While it may not be complicated to identify partial name matches, the main thing that an insurer wants to avoid is matching policy records of different individuals under the same customer ID, so often times it is safer to consider only perfect name matches.
• Contact information like phone numbers and emails can change from one policy to another, given that people often have multiple phone numbers or emails. The contact information can also sometimes belong to the agent.
• Even date of birth may not always be properly captured in different policies, given that non-standard proofs of age are allowed.
• Permanent IDs like “PAN” and “Aadhar” have recently started being captured, but historically other proofs were also allowed.

Typical customer ID creation logic used by an insurer relies on exact matches. Hence, it creates different IDs for the same individual when their records across different policies do not match exactly.

POSSIBLE SOLUTION: An algorithm that uses a multi-parameter matching logic and weighs all possible evidence related to an individual when checking whether two policy records belong to the same person.

An example is Aureus's proprietary customer ID creation logic. This logic can understand all typical variations in a person's written names and is configurable for the other parameters it can use for matching. It is also able to weigh matching evidence across parameters as per their weightage.

For example, if names appear to be written in different styles (refer to the name variants mentioned earlier), but if the date of birth and the phone number matches exactly, then there is a high likelihood of the two records belonging to the same person.

Additionally, a high confidence parameter like “PAN” matches would make up for mismatched contact information. This algorithm can also parse large volumes of person records when doing the customer ID match historically.

When applying this algorithm, we have observed up to 10% additional matching compared to customer ID created by insurer systems.

#2 IDENTIFYING FAMILIES OF CUSTOMERS

After identifying a customer against a policyholder, the next step is to identify a connected group of customers – such as families. While doing this is not necessary for an insurer, if they can identify families or households in their customer database, it gives them a significant advantage in improving their marketing offers and risk management.

Understanding the overall insurance cover and premium paying capability for the entire household allows the insurer to tailor offers for the entire household. They can also understand the entire household's value and corresponding risk exposure as against dealing with individual customers.

Identifying households, however, is a more complicated task than identifying unique customers because of the reduced number of parameters that can be matched. With more people preferring to rely on mobile phones rather than landlines, the only pieces connecting households are home addresses and last names.

Address matching is also anything but easy in a geography like India, where there is no standard format for writing addresses. Simple string matching or partial matching logic can capture only a fraction of the matches. Partial matching introduces the danger of incorrectly matching different addresses.

POSSIBLE SOLUTION: Aureus's proprietary householding algorithm uses a strong address matching logic customized for matching Indian addresses. It is also capable of using other configurable matching parameters to improve the quality of household matches. Individual matches are then connected using graph networking concepts to create good quality households.

We have observed up to 40% additional matching compared to customer ID grouping when using household matching.

#3 NPS ANALYTICS / Getting Actionable Insights From Text

Understanding written feedback without human intervention is important as it allows processing large volumes of feedback and identifying actionable insights without the need for a human being reading the feedback.

NPS (net promoter score) survey data consists of structured fields (product names, channel, etc.) and unstructured fields like emails and social media posts. The objective here is to analyze these fields together and in conjunction with one another. This will help for automated analysis of freeform text and identification of underlying trends in the data.

The most efficient way to foster loyalty or retention is to meet or exceed customer expectations. To determine whether you're fostering customer loyalty, find out how your customers talk about you. Look at your customer base and determine how they feel:

• Passive customers are unenthusiastic and vulnerable
• Detractors are unhappy and potentially brand-damaging
• Promoters are happy and loyal customers who continue to purchase and help grow your business with referrals

NPS analysis goes beyond number crunching; the main purpose is to get the explicit feedback. With NPS analysis, insurers can benchmark SLAs in the process. It also helps them to identify black sheep in their supply chain, agent, and distribution network.

A low NPS should be viewed as an opportunity to improve customer service to make the customer happier. Reasons for a low score usually fall into the following categories:

• High premiums
• Low claim approval
• Poor communication

POSSIBLE SOLUTION: Listen to your customer! Listen to their feedback, and actually do something to improve customer support and make them happier. It should go without saying, but always strive to strengthen your weaknesses and leverage your strengths. Here are some ways to address the most common low scores:

• High premiums – review your pricing and product structures. Be sure to make the appropriate adjustments.
• Low claim approval – this one can be tricky, but if a review of your claim denials shows nothing amiss, then maybe the wording in the policy needs to be rewritten to better explain the reasons for claim denials.
• Poor communication – this should be the easiest complaint to address. Review your procedures; find the gaps, and make sure the customer is getting the information they require or call-back they request within an acceptable amount of time.

#4 UNDERSTANDING CUSTOMER SENTIMENT

What is Customer Sentiment?

Knowing your customer's sentiment means you understand the different emotions they experience while engaging with your products or services. These emotions vary between positive and negative.

Customer sentiment analysis automatically gathers this information by detecting their emotions during customer interactions using NPL. The information is collected through text or voice analytics and can determine if the customer's emotions are positive, neutral, or negative.

To analyze customer sentiment and determine if it is negative or positive, two parameters are primarily used in the algorithms:

• Polarity: tells whether the sentiment is positive or negative
• Magnitude: indicates the degree of sentiment; how strong the emotion is (very angry / extremely happy)

For an insurer, understanding a customer's true sentiment about them is crucial. Customers who are not happy may terminate their business with the insurer in the long run and will be highly unlikely to give them additional business. Understanding customer sentiment is, however, a somewhat tricky business.

Insurers often rely on customer satisfaction surveys and NPS surveys to understand how their customers feel about them. The primary difficulty with surveys is the low response rate that is typically observed with them. Often insurers can get survey responses from a small fraction of their customers (less than 10%). Even if we include feedback collected from phone calls and other customer interactions, the insurer typically does not have any explicit feedback for 80% or more of their customers.

POSSIBLE SOLUTION: One way to circumvent the low survey response rate is to capture and interpret implicit feedback offered by customers. Implicit feedback is information gathered from the customer's journey with the insurer.

• A customer who always pays on time is more likely to be satisfied with the insurer than one who habitually misses payments or lapses.
• A customer who buys more policies with the insurer is more likely to be satisfied than one who doesn't.
• A customer whose complaint is resolved quickly by the insurer is more likely to be satisfied than one whose complaint resolution takes a long time.

If the insurer can collect such feedback from all customer events, they can get a very good insight into the sentiment of each customer. Aureus's proprietary SentiMeter® algorithm uses this approach to capture and quantify customer sentiment.

SentiMeter® provides the insurer with insight into each customer's sentiment and gives the rationale for the sentiment calculation. This enables the insurer to identify groups of low sentiment customers, examine the causes, and improve the overall sentiment. SentiMeter® is also useful in selecting customers for internal campaigns such as sales and retention.

Conclusion

There are many ways insurers can leverage AI tools such as NLP (natural language processing) and text mining to enhance customer retention and engagement.

The key is to understand your customer and make each interaction with them a positive one. Keep the customer data clean – avoid duplicate information that will make dealing with the customer difficult and time-consuming.

In addition to identifying individual customers for a given policy, identifying connected groups of customers, such as families or households, in an insurers’ customer database provides a significant advantage in improving their marketing offers and risk management.

Know your customer by listening to their feedback and strive to make their experience with you a positive one. Gather any information available to understand your customer's sentiment. There can be varying degrees of positive and negative, and anywhere in between. Understanding this feedback without human intervention is important as it allows processing large volumes of feedback and identifying actionable insights without the need for a human being reading the feedback.

Having the ability to understand a customer's true sentiment about them is crucial. Customers who are not happy may terminate their business with the insurer in the long run and will be highly unlikely to give them additional business.

Interested in learning more about understanding your customer's true sentiment? Click here to download our whitepaper, "Understanding SentiMeter®."