Companies of all sizes are getting deeper and deeper into deploying virtual agents to support customer service operations. When this was a nascent trend, experts recommended that companies experiment with various technologies and start to develop the expertise around the tools and, importantly, the customer experience being delivered.
Traditionally, developing a virtual agent involved a three-step process: 1) determine which call types, or intents, are best to automate, 2) determine the cost and business case to support the implementation and, 3) manually design, develop, and test the solution.
Today you can evolve those methods with artificial intelligence to deliver faster value by having a robot do these steps for you. Customer support functions are already largely supported by humans performing voice and digital (chat) interactions with customers. What if you could have a robot listen in on a very large number of interactions and determine the main call types and then dig deeper, determining all the various paths these intents take? One of the biggest issues with virtual agents is that they often only handle the happy path (the flow that is most common and issue-free) and as soon as someone asks a follow-up question or takes a left turn, the system has to send the contact to an agent. That significantly reduces the value of the virtual agent and ruins the customer experience.
Historically, virtual agent designers reviewed call routing reports and agent disposition codes and interviewed super users to determine what customers were calling about. Not only is this slow and expensive, it is likely inaccurate. Today, you can have conversational analytics language modeling tools listen in on interactions and produce detailed reports about the main call types and complex network diagrams for the further afield interactions and the relationship between call types and these seemingly random but related inquiries. Assuming you can import hundreds of thousands of these calls and digital interactions, you will get a very good idea of why people are calling and exactly what they are asking, which allows you to move to the second step in the process, the business case.
With the level of detail that comes from these customer support systems, you also get a very good idea of the effort required to automate them. The main insight is around the questions and conversation flow in general, such as scheduling an appointment, making a retail return, or looking up an order. You will also identify which data is needed and the associated interfaces required, all input to the business case. The logic of the business case, of course, is that while you can build a virtual agent for every anticipated customer request, the effort to automate should be less than the gross savings over an investment period, ideally less than two years. For all contact types that pass this test, you can safely develop the virtual agent to automate them.
The final step is the design of the conversation flow and general interaction with the customer. It is important to note here that most companies are convinced they know what the caller needs and, therefore, are their own best design experts. I have multiple stories that would contradict that. Anecdotally, I once asked a team of project managers who designed these solutions how often their design suggestions were rejected by their clients during design, only to be reintroduced, and therefore redeployed, at great expense shortly after go-live when the client discovered that customers did not interact with the virtual agents as anticipated. The answer? Ninety-five percent of the time.
With this caution in mind, a further advantage of these design tools is that the report and associated conversation designs are based on actual exchanges between customers and agents, and, therefore, reflect what is actually happening on the phone and in live agent chats, not on what the super-users conveyed during interviews.
Some of these conversational analytics language modeling tools also take this input and ingest it into their own virtual agent tools, but you can use tools that are virtual agent technology-agnostic and only provide the data, the business case, and the design, allowing you to use the technology you have already.
As companies continue to iterate and expand the use of virtual agents, these robots will become more effective at perfecting the conversation flow, identifying the gaps, and continuing to refine the business case for even further expansion. Maybe someday they will also do all the implementation from end to end for us.
Tom Lewis is managing director of Accenture Applied Intelligence.