During the lockdown, many call centers registered a major increase in call traffic, including one of the largest telecom operators in Eastern Europe. Under the new circumstances, the management decided to reinforce technical support on the second line and allocate additional resources to solving non-standard customer issues. At the same time, the quality of customer service on the first line shouldn’t be affected, so they decided to use the Neuro.net virtual agent to take on two important tasks: offering subscribers a new option plan and collecting customer feedback.
From IVR Bots to Neural Network
Interactive Voice Response (IVR) bots are nothing new. They have been used everywhere for many years. This is a rather simple technology to play recorded voice messages to customers when they pick up the phone. An ordinary IVR bot is incapable of answering questions or “reacting” to words in any way. It has a voice menu that can be navigated through by pressing the phone buttons. For example, when pressing “1” you can check the balance, “2” – your current option plan, and “3” – you can talk to a human agent.
Several years ago Neuro.net started to create a new, smarter generation of voice technology – a virtual agent that could act as a full-featured contact center, alone. This technology is based on a neural network, machine learning, and Big Data. To work efficiently, it should be trained using audio recordings of real phone conversations between humans – contact center agents and customers. Call records are decoded, analyzed, marked up for a neural network, and then used for training. The most difficult thing here is not the technological part, but the legal one. Personal data is protected by law, therefore, call records can be obtained only after the completion of a lengthy bureaucratic procedure.
Neural networks need to be trained from scratch for each new usage scenario. For instance, the neural network trained to collect customer feedback cannot be used for offering upgrades to option plans.
Why Neural Network
A virtual agent based on a specially trained neural network understands complex phrases and non-standard user responses. So, if an ordinary bot can distinguish only “yes” and “no”, an AI-powered virtual agent understands “I can’t hear you”, “okay”, “here we go” and even more ambiguous expressions.
For a virtual agent, there can be hundreds of various communication scenarios, even in simple tasks. When the duration of a conversation between a virtual agent and a customer reaches 30 seconds, the number of possible conversation threads exceeds one thousand. With an IVR system, this is simply impossible.
Underlying an AI-powered virtual agent are speech processing algorithms of Natural Language Understanding (NLU) Engine developed by Neuro.net, which are capable of working with relatively small data sets. The quality of conversation with a virtual agent is almost indistinguishable from a conversation with a human agent. It can manipulate with intonation, make pauses, and form logical replies. Moreover, a virtual agent gets better with every new conversation thanks to machine learning.
During the training process and test launch surprises can happen – for example, when a virtual agent learns to respond according to the prescribed scenario (script) but suddenly one of the users reacts unexpectedly, asks a question that was not provided in the training. In this case, some additional training is required to make sure that no surprises are going to happen in actual practice.
Back to the Telecom Company Case
As we mentioned in the beginning, the telecom company required AI-driven virtual agents to perform two tasks. The first task was to offer customers who were ready to turn to competitors a new plan option with a discount. The discount was offered only if the user rated one of the company’s services below 8 out of 10 points.
Its performance results exceeded expectations. The system made more than 10 thousand calls with a conversion rate of 37 percent instead of the planned 35 percent. This case revealed that an AI-based communication solution can work not just similar but even better than a contact center with human agents.
As for the second task of collecting feedback on the company’s services, it was the Net Promoter Score (NPS) survey with five questions for customers. A call was counted if the agent received answers to all five questions. After analyzing the results, it turned out that the virtual agent managed perfectly 98 percent of the dialogues, answering questions of the respondents almost without errors.
Productivity and capacity of the virtual agent were determined according to the client’s requirements and included 200-500 thousand calls per day. But if a million calls were needed, there would be a million.
Neuro.net virtual agents make thousands of calls simultaneously. If a customer is not available, the next number is dialed from the queue. Approximately 30% percent of customers answer those calls at the first attempt.
By the way, recently the Neuro.net virtual agent was tested by another, also a large telecom company (which name we can’t mention as well). That use case included more than 5000 variables: details of plan options, customer behavior patterns, etc. During the tests, AI-powered virtual agents not even once repeated the same dialogue, and conversion rates were comparable to the results of a contact center, with human employees.
Analysing the Case Results
Neuro.net virtual agents exceeded expectations and KPIs of the telecom company by 2-4 percent. Upon completion of the project, the following conclusions were made:
- Virtual agent is quite capable of working like a human agent even in situations when it is necessary to maintain rather detailed dialogues with customers.
- Quality of customer service during the lockdown was the same as under usual circumstances.
- Contact Center AI can take on the most boring simplest tasks so that human employees could focus on solving more complex and challenging problems and working on cases where a more personal touch is needed.
- AI-powered virtual agents are able to cover the peak load when human resources are not enough.
- Operational cost of a virtual agent is half the cost of a conventional contact center with the same, or even higher, efficiency level.
In conclusion, we can say that digital technologies, including the one described in this case, can free people from purely mechanical tasks. This is not about “taking over jobs by machines” but, more likely, about the fact that companies will be able to use the most valuable resource, time of their employees, to solve more creative and complex problems that no machine can really handle.