Difficulties of the project
We knew nothing about the work of the contact center.
There was no data to train the chatbot. We did not transcribe the recorded conversations not to waste time and money.
While we were working on the project, some of the topics of appeals became irrelevant.
The bot made speech recognition errors in 46% of cases.
We have updated the bot knowledge base.
We replaced the voice recognition platform and settled on Yandex SpeechKit.
We used typical phrases that customers say. They were written by the operators of the contact center.
To reduce the risk of error, we asked «Dodo» as much information as possible.