The voice bot that saved MegaFon 75% on communcation costs

Teams of Yandex SpeechKit and Neuro.net helped MegaFon to automate their customer contacts. Voice bots use Yandex SpeechKit to flexibly manage a huge customer base: they can communicate with thousands of customers at the same time and inform them of the company’s daily offers. Now such high-volume campaigns take MegaFon a day instead of than a week.

At the same time, conversion rates for consent to a targeted action range between 5% and 30% case-by-case, largely the same as with human operators. As a result, the company has been able to reduce the cost of effective contact 75%, increasing the number of campaigns along with their market reach without adding new staff.

MegaFon reached out to our partner Neuro.net for help in developing a voice bot for customer contacts. The client wanted to:

• Promptly notify their customers about custom offers from MegaFon and their partners.

• Maintain customer contact quality.

• Relieve human operators of routine tasks that could be effectively automated through voice technologies.

The digital operator Neuro.net had to inform the customers, offer them new pricing plans and services, process customer objections, carry out surveys, and collect feedback. The digital agent’s offers were tailored for the customer’s preferences and behavioral dynamics. A key task in deploying a digital contact center was to promptly inform customers about new offers while they are still relevant and likely to be welcomed by the user.

The Neuro.net team suggested involving the use of Yandex SpeechKit and the server capacities of Yandex Cloud. Such a high-volume, national project demanded cloud technology. The client wanted to launch the pilot as soon as possible and quickly scale it up, but preserve the scale-as-you-grow flexibility. MegaFon expected to be able to increase or decrease the number of calls to customers depending on the market situation.

Another point in favor of the cloud was clients' ability to keep their solutions up-to-date, to leverage state-of-the-art technology, and to relieve their IT team from time-consuming support service and update tasks — these tasks are now delegated to the engineers of Yandex Cloud as the SaaS provider.

The cloud is also more cost-effective: MegaFon could save on their equipment cost of ownership, technical expert payroll, and other related costs.

To help MegaFon reach their goals, Neuro.net used their own platform, enhanced by the Yandex SpeechKit for speech recognition and synthesis. More specifically, the project leveraged the Automatic Speech Recognition (ASR) technology.

Conversion technologies have been used to ensure interaction with the Neuro.net platform. During each call, Yandex SpeechKit recognizes the speech and transcribes it, i.e. converts it to text. This data is then sent in small sets to the Natural Language Understanding Engine (NLU) that implements the speech processing algorithm from Neuro.net. The engine recognizes target objects within the raw data: entities and intents. This information is used to structure the dialogs.

To improve the quality of recognition, Neuro.net and MegaFon selected a custom Yandex SpeechKit model provided as part of the SpeechKit Pro program. They trained this recognition model on MegaFon’s datasets and adapted it to operators' tasks.

To shrink the pauses before responses that often betray voice bots, their project features an intelligent module from Neuro.net to detect voice activity during conversations. It recognizes noises during call and accurately identifying pauses in the customer’s speech. In addition, the smart interruption system allows the customer to interrupt the robot when necessary.

A team of experts from Neuro.net and MegaFon set up the solution over several steps:

• Analyzing the tasks and developing scenario-based scripts.

• Training the model based on digitized records of real call-center contacts.

• Testing and retraining the model. The bot was challenged with real cases of unusual customer calls. Whenever the digital Call Center Representative demonstrated undesirable behavior, the system was retrained and updated with new dialog branches.

• Implementing and re-configuring pilot scenarios.

• Scaling the system.

The contact center’s voice bot can make an unlimited number of calls at the same time, supports multichannel capabilities and dozens of communication scenarios. The contact center was integrated into the MegaFon workflow.

Now the Neuro.net team continues to improve the digital contact center based on feedback from MegaFon experts and their recommendations to develop scripts and technologies. Neuro.net experts also keep track of the speech recognition quality metrics and, together with the SpeechKit specialists, retrain the model on a regular basis.

Using Yandex Cloud virtual servers in the project allowed the team to facilitate the solution’s scalability. Today, the project uses 80 servers.

Neuro.net’s contact center, featuring a voice bot enhanced by the Yandex SpeechKit technology, made it possible for MegaFon to tell their customers about tailored offers with greater agility. At the same time, conversion rates for consent to a targeted action range between 5% and 30% case-by-case, largely the same as with human operators. As a result, the company has been able to reduce the cost of effective contact 75%, increasing the number of campaigns along with their market reach without adding new staff.

The automation project allows MegaFon offer both its own and partner products. Promotion through a voice bot has become one of the competitive edges that MegaFon can offer to its partner program members.