Network Anomaly Detection with AI Customer Service bot - Elisa case

Callbots are the next wave of developments
AI chatbots are becoming more and more common

Annika provides real-time information

Up until now, Annika, the Elisa customer service bot, has been mainly used in the task of serving clients. Now, however, it has acquired new skills that help with network anomaly detection and service error monitoring. The artificial intelligence listens to incoming calls, identifies customers’ problems, classifies the anomalies and informs specialists about likely failures and the probable locations of these on the map of Estonia.

AI has been used in Elisa’s customer service in the form of service bot Annika since 2018. “Annika started as an AI chatbot. A year later, she also became a callbot to work in the call service, directing the client to the service staff member best equipped to solve the client’s issue. This year a third function was added. Now, Annika can provide real-time information about potential failures to Elisa’s monitoring service,” says Mailiis Ploomann, Head of Telecom Services at Elisa Eesti AS.

“The special situation that we were all faced within the second quarter of 2020 proved very well how good Elisa’s previously taken decisions had been. This encouraged us to continue to be bold in expanding Annika’s skills and features. Elisa offers a vital service to its clients and, therefore, needs to monitor the functioning of all services in real-time and at all times. By now we have learned to monitor hundreds of different IT applications with the help of special systems that allow us to detect failures and anomalies as quickly as possible,” explains Ploomann.

However, there are still situations that the used algorithms are unable to spot or that the previously agreed metrics cannot adequately reflect. In such a situation monitoring is best aided by AI that is quickly able to process hundreds or thousands of complaints in order to calculate what clients are unhappy about and where the problems lie.

Because Annika answers all customer calls now, we have established a connection between AI and monitoring. At that, Annika uses various classifications to identify anomalies and to signal if a particular service is causing more issues than before.

Antti Suursalu , Leading Monitoring Specialist at Elisa

Elisa’s AI and service bot transmits monitoring information in three categories:

Location – there’s a map view of incoming calls where each unique caller is depicted as a dot. The more dots in a region, the more precise the fault communication. Also, we are able to avoid treating the customer’s complaint as an individual case.
Problem – Annika transcribes the call and displays the information at the dot of each unique caller together with all other information Elisa knows about that client. Through that, Elisa is able to relay the contract or the call number to the specialist dealing with solving the problem. Additionally, the transcription enables to assess the problem and its impact.
Service and network anomaly detection – for most calls, Annika is able to understand which service-related fault the customer wants to notify Elisa about. This enables to measure the number of incoming calls that concern a particular service and compare it to previous metrics. If the number of complaints differs remarkably from what it was at a compared time period, the bot informs a specialist about the anomaly

AI chatbots are becoming more and more common

Customer service bot Annika is AI that has been developed in cooperation between Elisa and MindTitan. It has been in use since 2018, whereas Annika started to take client’s calls in May 2019. Markus Lippus, the Leading Data Scientist at MindTitan, explains the difficulties of speech recognition in telephone calls, “Firstly, archived phone calls have very low sound quality, and secondly, these are full of various kinds of noise which can be caused by the specifics of the telephone or the way the signal is transmitted or anything that may cause background noise. These reasons make work harder for the acoustic model – a component that separates human vocals from the sound stream. The third problem is that people speak very differently from one another and also imperfectly – some only pronounce half of the word, use synonyms or general terms instead, or even skip words.

All of this complicates the work of the language model that has the task of putting the identified sounds into words and sentences.

“People are able to understand such speech because we spend a large part of our lives distinguishing human speech from other sounds. We have extensive knowledge of the world and how it works. This helps us to assume from context which word is missing from the sentence, what was mispronounced or what was only hinted at.”

Adding complexity to failure detection are two more models: the classification model, which must understand very different problem descriptions to understand exactly where the error might be; and the network anomaly detection model, which assesses the content, volume and area of calls to see if the problem is pointing to a particular device or a whole area.

“These steps are crucial because if calls are transferred incorrectly or a service is misdiagnosed, it is easy for regional failures to go undetected or for some services to be given false alarms which waste time,” Lippus adds.

Service bots are one of the most common types of the practical applications of AI in the world. Only a few years ago, Annika was one of the few of its kind customers came into contact with, but AI chatbots are becoming more and more common in most major service environments.