Building a More Intelligent IVR Through Machine Learning

Intelligent voice response (IVR) technology has been around for decades, appearing in its earliest forms in the 1960s and 1970s. For many companies, the technology remains stuck in the past. Often perceived as a bit archaic, IVR can spiral into a frustrating web of number punching if it lacks thoughtful design.

Despite this technology's shortcomings, it still deserves a place at the forefront of modern customer service if altered to function as a dynamic infrastructure. As we've adapted our IVR using machine learning at Clearlink, we've seen its value to our business and to our customers expand dramatically.

After establishing and relying on a basic IVR years ago, we built a data warehouse to take a more hands-on role in customers' IVR experiences. Machine learning has facilitated changes that have shaped an entirely new intelligent customer experience that continues to evolve.

An intelligent IVR isn't just a broader, more convoluted web of options for customers to navigate while trying to address their needs. If built correctly, it can be a simplified system that addresses customers' needs efficiently.

In an ideal world, an intuitive IVR would be able to address customer issues as well as a live agent could, decreasing handling time without hurting customer experience. Getting to that point is difficult, but it's possible.

Before we could build a more intelligent IVR, we needed to understand our existing IVR, so we started by mapping every single piece of it. We identified every path that a customer could take from beginning to end and then identified each of those paths to see what people did versus what we hoped they would do. We watched people start on one path and then go back, or start down a path and get confused and drop off. By mapping all of that data, we created a detailed picture of what realistic customer experiences in our IVR look like.

After you've evaluated your current IVR and identified opportunities for improvement, consider the following three points as you move forward:

  1. Use inputs on your web domain to dynamically accelerate a customer through your prompts in real time.
  2. Assess the advantages of using natural language processing for a more natural voice interaction versus button pressing for your business application.
  3. Always be sure you're using your customers' inputs to route them to the person or group that can help them fastest. Make your IVR actually listen and react.

An Ongoing Process

Once we established our baseline, we could target areas for improvement. We put all of the data we had about our customers to use, including information like whether they're an existing customer, if they've visited specific web pages prior to calling us, or if they've expressed an interest in a specific product. When they call in, we use that information to guide IVR routing. The entire premise of this is connecting a customer to where their question will be answered the fastest, but it requires constant monitoring. We continuously track friction levels throughout the IVR, looking for spikes in abandonment percentages that alert us to points in the IVR where we aren't serving customers as well as we could be.

For an organization where sales and marketing work so closely together, these insights are significant. They allow us to reduce customer frustration by answering questions faster, cutting down the a time a customer needs to spend on the phone. We also try to reduce that time by looking at customers in our call queue relative to the amount of active pay-per-click ads we have running at any given point. With machine learning, we've been able to balance the two better than ever.

Historically, we used data to help us make decisions about when to increase or decrease our volume of PPC ads driving customers to our call center. Now, we use machine learning to develop predictive algorithms that forecast our likely call volume for a point in time so we can then optimize ad spending accordingly. We can also predict how many agents will become available in a given time frame and set up PPC ads to start driving customers in again when we'll have the capacity to talk to them without long wait times. Instead of looking back in time and seeing where we overreacted to a short-lived increase in call volume by shutting down PPC ads prematurely, machine learning makes our process much more proactive. Customers have a better experience because they aren't frustrated, and we eliminate spending on ads that aren't going to actually drive profits for us.

We see our IVR as a system that will always have a place in the customer experience. For a lot of interactions, it's always going to be the best option. Right now, we aim to address 50 percent to 60 percent of customer concerns via the IVR and the rest through alternate channels like chat, or buy flows. When we see that we're handling more than that ideal percentage of callers using IVR, we know we haven't found the sweet spot with a given customer path. We work on finding ways to get customers to the path that will create the least amount of friction for them, which we do with the help of an enormous amount of customer intent data as well as machine learning.

Before we can turn to machine learning predictions, we take a very concerted approach to understanding customer intent long before they enter our IVR. Our intelligent customer experience platform consists of a centralized graph database that allows us to capture and connect every customer contact point we have across all of our channels. It allows us to observe all of a customer's initial intent signals in real time, from what they're searching for on Google to the websites they visit, the landing pages they reach, and so on.

This information gives us major insight into what their intent is as we see what they click on the website, where their mouse is hovering, and more. With each incremental input, we can anticipate customer intent. We use unsupervised classifications to construct a journey based on all of our historical interactions with customers that triggered and signaled the exact intent inputs of that customer, and we use A/B testing to then extract the probability that a particular customer journey will be what that customer wants.

After a few inputs, we hit a certain level of critical confidence in outcomes for various customer journeys, and we then roll out the red carpet: a fully optimized, expedited customer experience journey that we know has a high probability of success. We present customers with this journey, which could be the IVR with preset intelligence routing inputs. However, it doesn't necessarily have to be the IVR. For some customers, IVR will always be the best choice, but we've found that this same approach can make other avenues just as effective for the right customers.

Refining Customer Experience

As we gather unsupervised classification algorithms while directing customers, we're able to conduct customer clustering to classify and more quickly direct customers with similar intents, friction points, and preferences to solutions they're going to enjoy the most. An unsupervised clustering algorithm provides us with a definition of a particular customer group that we can quickly access when a new customer visits our site.

Through this, we've seen that for difficult conversations, a very high percentage of both baby boomers and millennials still gravitate toward a phone call. But for other situations, we have groups that will do anything to avoid using the IVR. IVR used to be the only destination for customer journeys, but we now can provide other options, like chat and buy flows.

Ultimately, we don't want to force people to use channels they don't want to use. We should always cater to the solution that a particular group most wants to use. With that goal in mind, we expanded our chat division and now have more than 20 chat agents in our call center who directly integrate sales into our chat environment. We'd like to continue to grow this part of our business as we see our customer base evolve. When we modify this path, we use data from just the group of customers who prefer chat over IVR, so we never get to the point when we're trying to make one channel work for every customer.

As we test and identify different groups and their preferences, they challenge us to develop better ways to meet their needs. For customers who would like to do everything online, we might have a self-service, highly user-friendly application. The customer not only saves time when our agents are busy, they also solve their problems the way they prefer. We believe that alternative ways to support customers better will always exist, and we need to continue to look for them wherever possible.

On the other hand, for customers who still prefer to use IVR, we're providing agents with information to empower them to have more relevant conversations with customers. We've started building the technological infrastructure to pass information to them faster so we set them up for better customer interactions and better chances of success.

Ultimately, there are multiple outcomes that we could deem successful beyond just a sale. When customers have a great experience with us, that's a success. When they're more likely to recommend us to a friend or feel compelled to call again, that's a success. As we refine this process, understanding the breadth and interaction of all of our key performance indicators is extremely important. We ultimately want to attach the most value to each customer experience.

If someone comes into your service center and you don't close a sale, but they had an amazing customer experience, you'll get the chance to sell them something else in the future. Keeping that idea in mind while refining an IVR is crucial. If you're answering customer questions or helping them with problems they're experiencing along the way, you'll positively impact your brand and raise your chances of future success with those customers.

When we maximize the customer experience, we also maximize the more important elements of conversion rate, like Net Promoter Score or overall customer satisfaction. Service centers are commonly blinded by more traditional IVR KPIs, like average handle time Unfortunately, that doesn't mean anything in the long term if a customer ends up having to call back because his problem wasn't resolved or if he hangs up dissatisfied.

We keep all of these factors in mind as we continue to modify our IVR. A truly intelligent IVR and machine learning partnership is one that learns and evolves over time, improving its algorithms as more customer data becomes available and customer journeys change. It also includes finding the data to manage our call volume and PPC spending as best we can. To think of an IVR as finished or set in stone is to lose the opportunity to build the best customer experience possible. Accepting that we can always improve, investing in the tools that will help us do that has been the best way to facilitate ongoing customer experience success.

Pieter van Ispelen is head of decision science at Clearlink.