Hello, it's me-a robot
On the eve of the last “black Friday”, the Russian online store of elite clothing and shoes KupiVIP ordered a call to its customers using a robot. The success exceeded expectations: in just two days the robot made 5 thousand calls, and only two clients realized that he was not talking to a living person.
KupiVIP offered a robotic solution for calling customers russian startup Neuro.net. It was created by Russian entrepreneurs who moved to California to search for promising niches for high-tech business. Soon after moving, they realized that in America, robots that call people do not know how to do so much: for example, to clarify the delivery address of a product or accept a simple review about the work of a company. In addition, they endure long pauses and cannot respond to complex questions. And then the specialists Neuro.net we set ourselves the task of making a voice robot that looks as much like a living person as possible, so that it can understand the client’s questions and explain the points of interest. To do this Neuro.net it uses learning technologies using a neural network.
“Technology Neuro.net it is based on the open-source TensorFlow framework, which allows you to create your own projects in the field of Machine Learning, – tells Alexander Kuznetsov, co-founder Neuro.net. — The TensorFlow tool was originally used by the artificial intelligence division of Google AI, and is now available for developers in the field of machine learning and is used by companies such as AMD, NVIDIA, Uber, Intel, eBay and SAP. And the basis of voice recognition of the interlocutor is Neuro.net Google Speech Recognition technologies are used. Google’s tools today provide the highest accuracy of recognizing the words of the interlocutor, which for the English language is more than 95% of correctly recognized words.”
Train the robot
The KupiVIP store became interested in the idea of using a robot in a call center last summer. To train a neural network, specialists Neuro.net they were asked to provide them with audio recordings of calls made by the former call center. Then these records were decoded, analyzed and marked up accordingly for integration into the neural network and linking semantics with the IT component of the solution. The software was also modified accordingly, phrases for the robot were recorded in the studio according to the created conversation structure. As a result, according to representatives Neuro.net they managed to create a very competitive solution.
“The key advantage of our project is an almost indistinguishable level of conversation from the real one, — says Alexander Kuznetsov. — The robot observes intonation, withstands pauses and the logic of constructing answers, and most importantly, it gets better with each new conversation thanks to machine learning technology. Another important feature is that the robot gives the user an answer on average only 700 milliseconds after the end of the interlocutor’s phrase, compared to about 2 seconds for competitors. Another significant advantage is taking into account the history of the dialogue. The robot remembers the user’s previous answers and takes them into account when forming the logic of future answers. As a result, the robot reacts to the words of the interlocutor in accordance with the history of communication.”
According to experts Neuro.net, their software solution is successfully trained. “Instead of typical patterns (templates), we set a vector, in fact, we worked with the creation of deep logic at the intersection of linguistics and programming, — continues Alexander Kuznetsov. — During the analysis of real calls, the system received a huge array of data on a variety of dialog scenarios. And now the robot already knows the possible options, how the client can answer him, and even if he says something unusual, the system understands the meaning of the answer.”
A robot is more profitable than a call center
Representatives Neuro.net they claim that the cost of their solution is at least twice as low as the services of a traditional call center. At the same time, the tasks that can be solved with the help of a robotic call center are almost the same as those solved when calling real people: collecting the so-called NPS (customer loyalty index), working with customer outflow, first-line technical support, etc.
As for the payment model for using the robot, when calculating the cost of their services Neuro.net it is based on the duration of communication between the robot and the client. “We expect to work with customers for a long time, so we do not charge for the development and installation of our solution, — explains Alexander Kuznetsov. — All work on data analysis, script development, robot training and script launch, as well as further support for this scenario (changes, analysis and additional training) are not paid separately. Only the actual cost of working on the line is paid, depending on the length of the conversation. For the customer, this looks the same as working with a regular call center on outsourcing. As for the cost of a minute of a robot conversation, it depends on many factors: the complexity of the scenario, integration with internal systems, as well as the volume of calls. On average, if we compare it with the cost of a minute of a real call center, the cost of the robot will be at least twice lower. Moreover, provided that the robot will show performance metrics no worse than a living person.”
In Neuro.net they say that, in addition to KupiVIP, a number of large companies intend to use the services of robotic calling in the near future, and the greatest potential for the successful use of such a solution is expected primarily in retail, banks and the telecom sector. “Many companies are either already working on the introduction of robots, or are planning to do it, — notes Alexander Kuznetsov. — The only thing that can stop is the volume of such calls within companies. For technical implementation and economic sense, while the integration of such a robot is justified on a large volume of calls, for understanding-these are teams with call centers of 50 people”.