How to Streamline Customer Service with NLP?

How to Streamline Customer Service with NLP?

streamline and improve customer service with artificial intelligence and nlp

Companies receive support enquiries from various channels. This may include emails, support tickets, tweets, chat conversations with customer support representatives (CSRs), chatbot conversations and more.

This is a lot of data that you are dealing with and it’s mostly unstructured and scattered in nature, making it that much harder to manage.

The question becomes, how can you leverage all this text data to improve speed in responding to customer support enquiries or reducing the number of tickets that are being opened?

This can partly be done with automation using Natural Language Processing (NLP) and Text Mining. This article will give you an overview of 5 areas in your customer support workflow that can benefit from A.I. based automation. This is not an exhaustive list, but a list that applies to most customer support teams at medium to large organizations.

 

What is NLP & Text Mining?

NLP is a form of artificial intelligence that enables computer programs to process and analyze unstructured data such as free-text data. Text mining, its closely related cousin uses principles from NLP and Data Mining  to be able to find patterns in any type of text data, spoken or not.

In combination with Machine Learning, Text Mining and NLP can be powerful tools for transforming messy unstructured data to something more structured (e.g. predict labels for each support enquiry). It can also help you find patterns in your data (e.g. groups of similar enquiries) as well as have a natural language conversation with humans (e.g. a chatbot).

Without these automation approaches, you’ll be left to perform a lot of manual, repetitive work or use a set of complex rules to achieve some level of automation.

 

How NLP helps in Customer Service?

Here are 5 ways in which NLP and Text Mining can be useful in reducing friction in your customer support processes.

#1 Recommend best answers

When your CSR tries to respond to a customer problem, they can be overwhelmed with identifying the best answer from the pool of possible answers. What they need is one answer that will address the customer problem.

Some companies maintain an exhaustive list of problems and corresponding answers which the CSR has to search through, sometimes even manually. This can be painfully slow and draining if you have to perform a search for each and every question.

NLP can be really useful here in recommending the best answers given a support enquiry. It becomes even better when the answers have a “score” which indicate the likelihood of the particular answer solving the customer problem.

Using this approach, instead of explicitly performing a search, the CSR is now having information pushed to them automatically, preventing a break in their workflow.

With this, the response time can be much faster, which also means, your CSRs will be able to handle a larger volume of support issues. What’s even nicer is that they are not overwhelmed by the end of the day due to the number of searches they have to perform.

 

#2 Suggest historical threads

While some support questions can be easily answered with the recommended best answers, others can be complex, requiring extended research. One way for CSRs to solve complex issues is by looking into related historical threads (which have been successfully resolved), and understanding how those issues were resolved. With this, your CSR will be able to better resolve the issue at hand or form a more complete answer to the support question.

With NLP and Text Mining technologies we can automate this process by recommending related historical threads for any given support enquiry. This saves your CSRs from having to conduct various searches, contact peers and their manager for help on an issue.

The benefit of surfacing related historical threads is (a) the potential decrease in response time and (b) reducing follow-up support questions as complex issues are resolved in fewer interactions.

 

#3 Group questions to limit context switching

As we all know, context switching can be hard. Going from resolving issues related to signups to billing and then back to signups can be a productivity killer. According to Bud Roth, author of Be More Productive: Slow Down, distributing your energy over a wide variety of tasks can dilute your effectiveness the same way interruptions do.

By grouping similar support questions, CSRs can address similar problems in chunks, where the knowledge bank that they’ll have to tap into and the pool of potential answers are related.

customer support question grouping with nlp

Figure 3: Grouped customer support questions

With the use of Natural Language processing and Text Mining, we can automatically group similar questions as shown in the example in Figure 3. Notice that the first question is all about adding a profile picture.

The benefit of doing this is that, it maintains the same train of thought in resolving issues. In some cases, the solutions may be identical, while in others your CSRs will know what steps to take to resolve an issue while while everything is still fresh in memory. By limiting context switching you can expect to see a reduction in response times.

 

#4 Auto-route Questions Based on Expertise

Support questions come in all shapes and form and customers may express the very same question, quite differently. For example, the questions in figure 3, are all related to adding a picture to the customer’s profile, with a slight difference in how it’s structured.

 Figure 3: Similar questions, different expression

By classifying each incoming question to a predefined set of categories with NLP (e.g. profile, picture, attachment), you can use these categories to route the questions to agents who are best at handling those topics. Some CSRs may be highly qualified at handling certain topics more than others. By intelligently routing questions to relevant expertise, you’ll be increasing the productivity of your CSRs as they are not spending time learning how to address support issues out of their expertise. 

 

#5 Auto-prioritize support threads

A few companies we spoke with, mentioned addressing support threads in the First in First Out order (FIFO), meaning the oldest support threads get addressed first. Other companies we know, manually assign priority based on the severity level of the issue.

Don’t forget, that not all customers are equal and not all problems deserve the same level of attention. By addressing threads in the FIFO order or assigning priority based on severity alone, you are missing out on the opportunity to retain your highest valued customers. If you are spending your time solving a 100 low priority problems for low value customers before serving your most valuable ones – it’s time to think about making changes.

While you can start serving your VIP customers first, with A.I. based automation, you can combine various factors into prioritization. For example, you can develop an NLP model that combines priority of historical support threads with the value that each customer brings (e.g. customer lifetime value) to auto-assign priority of new support questions. This will ensure that your highest value customers with high priority issues get served promptly and by your best CSRs.

 

How to get started?

We recommend that you start by listing down your most inefficient processes. What’s taking you the most time? Is it the search for answers or question routing gone bad? Once you know what’s hurting, the next step is to determine if the inefficiency can benefit from Machine Learning, Text Mining and NLP automation.

Start small and slowly grow your NLP capabilities to address inefficiencies. Doing it all at once will set you up for failure as there’ll be too many changes in your workflow. In addition, some of the automation can be at odds with each other. For example, by auto-prioritizing threads, grouping questions may or may not be effective. Optimizing all at once also prevents your ability to measure success. You’ll not know if reduction in response times was due to recommendation of best answers or if it was due to auto-prioritization of threads.

For some of the companies we’ve worked with, our audit showed that what they needed was a search engine or small changes in their software—not A.I. based automation. Despite the hype, not all problems benefit from A.I. and can be resolved with other standard solutions. In cases of uncertainty, you can always ask us for our recommendation!


Opinosis Analytics is a Natural Language Processing & Text Analytics company. Part of what we do is help companies improve processes around their unstructured data through custom NLP solutions, technology audits and via implementation guidance. Get in touch, for more questions!

 

Have a thought? Share it below.

Join 100+ other executives. 
Subscribe to keep receiving tips, reports & case studies.