The Power of Customer Service Analytics AI 🤖

By now, almost everybody has heard of artificial intelligence (AI). You’ve probably also heard or experienced that it’s often not that good yet—many companies have developed very basic AIs that are only marginally better than a keyword search, to the detriment of everybody’s trust in AI.

However, when it comes to categorizing unstructured text and speech in a customer service environment, the latest innovations in machine learning and natural language processing are well suited to the job.

In this article, we’ll walk you through the essentials of AI in customer service analytics, how it works, and the many benefits the insights can have.

Contents:

• What Is A Customer Service Analytics AI?
• How Does a Customer Service Analytics AI Work?
• Three Types of Natural Language Processing AI for Customer Service
• Why does AI work particularly well in customer service?
• How To Use Customer Service Analytics AI In Your Business

‍

What Is A Customer Service Analytics AI?

Customer service analytics is the part of your role as a customer service leader where you start digging around in your customer conversation logs for insights. 

That could be done in several ways:

• Manually, by listening to a handful of phone calls or putting a sample of conversations in Excel and trying to draw conclusions.
• More systematically, by having agents categorize every conversation as they happen with tags.
• Or, finally, automatically by leveraging a customer service analytics AI to understand, analyze and display insights in real-time with AI-based tagging.

Each of these methods should help you understand more about what’s causing customer contact as well as customer sentiment—but, the results of each are not equal.

It’s important to note that AI enables your analytics to be more accurate and granular, cut through the subjectivity of human opinion, and do so in real-time.

What the AI does is read or listen to each customer support conversation, much like a human would, and then label that conversation with a predetermined tagging taxonomy (which could include sentiment, happy, sad, angry, or a topic you want to track, or even intent or priority).

AI systematically and scalably draws insights from your support conversations, making it the go to choice for teams with high volumes and frequencies customer contact.

I asked Kirsty Pinner, Chief Product Officer at SentiSum, why AI is such a powerful tool in customer service. Here’s Kirsty’s answer:

“AI can cut through the subjectivity of human opinion, and no matter how something is said, it can report on the customer issue in a simple way. The latest developments in AI analytics can handle complexity extremely well.”

“No other method gives a representation of customer conversations this accurately. Manual tagging is too subjective and trying keyword analysis is too blunt a tool.”

How Does a Customer Service Analytics AI Work?

So, how does it do it? And can it really be as accurate as a human? Haven’t we all heard that AI is still a bit rubbish really?

The main goal of AI is to simulate human intelligence, enhanced with the capabilities of a machine—infinitely scalable, never tiring, and doing it the same every time.

Put to use in a business environment, AI is particularly good at automating repetitive processes related to understanding, problem-solving and learning.

Two advanced subfields of AI are machine learning (ML) and natural language processing (NLP). These are built around algorithms, which take large datasets (e.g. the last 12 months of your support conversation logs) as their ‘training data’ and some initial human direction and together they understand the logic between words, sentences and phrases in the human language, learn from it, and can apply the logic consistently into the future like a human would.

In our article, ‘Why is machine learning NLP so good at Zendesk ticket categorization?’ we discuss the different types of Natural Language Processing AI in depth.

There are three key types of NLP, each of which produce a varied results. Let's look at all three in a customer service analytics environment.

Three Types of Natural Language Processing AI for Customer Service

If you have 10,000 monthly support calls, emails, chats, Messengers, WhatsApps and so on…the last thing you want to do is spend a week going through them manually. In fact, that’d be very boring and the results would probably be inaccurate.

A computer, on the other hand, doesn’t get bored and has no limits to its reading speed. It can process and categorize millions of conversations in near real-time with NLP.

Of the three types of NLP below, we recommend avoiding 1 and 2. Customer service leaders come to us all the time after trying them and feeling like their time is wasted.

Let’s look at why:

1. Keyword Extraction

You may have come across automated tagging in a helpdesk tool like Zendesk or Intercom. They rely on keyword extraction to tell the automation which tags to apply.

How does it work? Simple, if a keyword is present the follow a certain rule: 

“If the word ‘refund’ is found anywhere in the conversation → Apply the tag ‘refund_request’ to the conversation.”

There are obvious flaws in this approach. The main one being that there are hundreds of ways to ask for a refund without using the word ‘refund’, meaning that unless you can think of them all and tell the keyword extraction tool, your analytics will not have accurate results.

This approach fails to provide granular results, too, because it’s simplistic. Unlike other approaches, keyword extraction isn’t an intelligent form of AI, so it can’t tell you additional information like the why behind the request. You may know you had 1,000 mentions of ‘payment issue’ but not that 500 payment issues were caused by Klarna, 300 by Payal, and 200 due to a website bug.

2. Rule-based NLP

Rule-based NLP algorithms leverage libraries of ‘rules’ to help them understand human language.

Instead of you directly telling the automation what to do, rule-based NLP understands things like ‘liberty’ and ‘freedom’ are synonyms.

This has much less complexity and much more accuracy compared to keyword extraction. It has two downsides:

• If the rule doesn’t exist then it won’t understand the meaning.
• It’s still looking for words or sentences, rather than understanding the language like a human would. If a customer said “I’m not happy because my payment keeps getting rejected, I think it’s Paypal’s fault but it could be Klarna”, and then the conversation between the agent and customer continued to determine it was neither Paypal or Klarna, the rule-based NLP would still tag that conversation with “payment_issue”, “klarna”, and “paypal”.

In short, it still isn’t intelligent enough to provide meaningful insights.

3. Machine-learning based NLP

Machine-learning based NLP understands speech and text in a similar way to humans. After digesting a dataset being trained in it, it uses statistical inference to carry that knowledge into new environments it’s never seen before.

For example, it can identify and infer the meaning of misspellings, omitted words, and new words like slang by itself.

Thanks to the ML, it learns the patterns between phrases and sentences and constantly optimizes itself to improve accuracy over time.

We highly recommend you ask your provider the type of NLP they’re using and whether they will build you a customized model or not—if they don’t plan to, it’s unlikely to provide the accuracy or granularity you need to achieve your goals.

‍Deep Dive: Machine learning vs keyword tagging: Why ML is best