Designing Bots for Faster Customer Service

Here at HYLA Mobile, an Assurant company, we offer a white-labeled customer service function for our trade-in services. In other words, if a customer contacts one of our customers (carrier, retailer, OEM) with a question about the phone they just traded in, that question gets seamlessly shunted in our direction.

Our support is offered via phone, chat, and email, and many if not most of our support requests are received online. In fact, the number of support queries we receive via the internet is very large and continues to grow, which could lead to a problem with scale.

First, although the volume of queries we receive is large, it isn’t always large. From September through January, the number of trade-ins always increases dramatically due to the launch of new flagship phones and the advent of the holiday shopping season. Meanwhile, the number of our customer service representatives (CSRs) remains constant. We would rather not hire and train CSRs just to let them go four months later.

Second, customers love speed. They want answers to their questions quickly, but with more questions coming in online, each CSR must juggle more and more questions at once. This means that it can take up to ten minutes for a customer to get an answer to their question, especially if they can’t give us complete information such as their order number.

Based on these constraints, we began to ask how we could scale to meet customer demands while keeping staffing levels consistent—and without putting increased pressure on our existing staff. As such, we began turning to bots.

Creating Test Cases for Chatbot Support

The first thing we began to notice with regard to customer support was that most requests fell into three categories:

• What is the status of my trade-in?
• You’ve received my device—where is my payment?
• I’ve begun my trade-in, now when will I receive my shipment kit?

What we noticed about these three questions is that they’re all answerable based on the information within our database and that it tends to take a while for human operators to answer them. This is because customers don’t always provide us with the critical information we need. Instead of giving us their order number, they’ll give the email associated with their order, their home address, their phone number, etc. Each of these means forming a different kind of database query.

Forming a database query can take a long time for a human operator, but a relative instant for a machine. Therefore, we decided that this kind of question—a single query that can be answered via a database lookup—was the perfect starting point for our experiment with AI-powered customer support.

Building an AI-Powered Customer Support Module

Based on our research, we began creating a customer support module powered by AI, machine learning, and natural language understanding (NLU). The end-goal was this: as soon as we receive an email or a chat request, the chatbot uses its NLU to understand the intent and the sentiment behind the question. Based on this understanding, it formulates a database query, retrieves the information in a few seconds, sends a composed response, and then logs the interaction in our CRM.

We started small. First, we started by answering easier questions, such as “what’s my order status?” At this part of the test, the bot didn’t even answer any questions directly. Instead, it only performed the sentiment analysis and subsequent lookup. A human technician would receive the answer retrieved by the bot, validate it to ensure that it was the correct answer, and then use that information to respond via email or chat.

Even this experimental validation turned out to be a large win, because it eliminated the time and effort of the research phase. In fact, this step turned out to be so successful that we rapidly moved on to our next step, which was allowing our bots to automatically respond with prewritten emails and dialogue to certain identifiable intents.

The Future of Chatbot Support at HYLA

Although we’re still improving the logic of our customer support chatbot, we are now answering up to 80% of customer support requests without any human interaction—and not a moment too soon. Between the successful launch of the iPhone 12 and the increase in support requests due to COVID-19, the number of requests on our platform have more than quadrupled.

We’ve made some specific improvements over the last nine months. These mostly have to do with the way that the bot escalates requests. For example, we’ll escalate to a human representative if we are unable to understand the context of the email with a high enough certainty or if the customer gets engaged in follow-up discussion after the initial response.

These scenarios can either confuse our bot, or they may be too insensitive for our bot to handle. What this means, however, is that with the bot taking care of simpler interactions, the CSR team can give more attention to these more complex questions, increasing customer satisfaction.

Lastly, we can now collect even more data about our customer service interactions. When we understand which issues are the most prevalent, we can improve our process, helping to tamp down the number of requests we receive overall-and ensuring that our clients’ customers are happy throughout every interaction with our company.

If you’d like to learn more about how HYLA can improve your smartphone trade-in process with AI, machine learning, and other advanced technologies, reach out to us today!

By Chintan Shah