Põhisisu algus

How to make your customers - talk to a robot

18.12.2019

case-study by Elisa Estonia in 2019

Although Elisa has defined itself as an entertainment (and telecom) company in Estonia for quite some time, there is no doubt – we are still by far one of the biggest customer service providers as well. And since our product portfolio is constantly growing, the complexity of our customer service grows correlatively.

For instance, back in 2015 there were a bit less than 400 different topics our customer service agents were supposed to handle, at the end of 2019 the same number is close to 700. Since we are a quadruple play telco, the number of different technologies in every domain is constantly growing as well, as year by year more generations are added, which combined with our new services adds up to a cocktail of services no human agent can digest without smart support.

At the same time, it is obvious that the customers’ expectations for our SLA and quality are exponentially growing as well. The world gets more global and more digitalized every single day and we, as a service provider, are compared to Google, Amazon and Facebook more often than to our local commercial competitors. Once you add the local naturally growing wages and the lack of tech-friendly staff who could speak at least 2 (if not 3) languages, you’ll find that it’s a quite tight place we must operate daily.

But the good news is, there is a little ray of sunshine peeking through the curtains already – new emerging technologies that will help us simplify and speed up the customer service processes for both our front-line agents and customers as well. Our customer service bot Annika is here to prove that point.

Annika is a result of an excellent cooperation between Elisa and MindTitan  (an Estonian tech company that is focused on machine learning and applied AI solutions) and started its customer service career as a chatbot already in the spring of 2018. Though today there are a lot of chatbots for Estonian companies all around, Annika is once again making history as the first callbot able to work in customer service side-by-side with human agents – proving that the work of our language- and data scientists is world class. We are proud to admit that Annika is a solution, envied by many colleagues, both near and far – and since by the end of 2019 there are more than 100 000 people (close to 10% of the population) Annika has dealt with, we can already say, that talking to a robot is the new normal!

So, what does Annika do in our call center?

As was already established before, there are a lot of different topics customers are approaching us with. And it would be simply unhuman to expect any one person to be equally qualified to handle them all. A simple solution is of course separating agents and letting them specialize on certain domains, which also means – we need a way to put together the customer with the agent that is qualified to answer that specific question.
For more than 10 years, that solution has been IVR (Interactive Voice Response), a robot that is commanded by the customer by pushing number keys on the phone (press 1 if your inquiry is regarding invoices, 2 if you need assistance with a technical issue… etc.). In our experience, once you have more than 2 layers of such choices, constantly waiving your phone is starting to annoy your customers and at a certain point they just press something, anything just to get to an agent. Especially in the era of smartphones (with the screen locks turning on during listening the IVR prompts) the error rate of such choices varies from 25 – 30%. And these are all contacts that need to be retransferred by the receiving agent – leading into longer waiting times and therefore longer handling times for both sides. Therefore, it is obvious our motivation to bring Annika to our contact center was high, to be modest.

Today it is Annika’s job to pick up the incoming call, ask what the reason is for contacting us and then routing the customer to the correct agent. Sounds simple? The cool part is that for the customer – that’s exactly how simple it is in real life as well. Behind the scenes it’s of course a bit different story J

Our first criterion was that adding Annika had to be easy and quick for our contact center software (Avaya in our case). It was also important that we had to be able to implement it gradually, with the old logic staying put (just in case we needed to rollback).

Since we want to learn one language at a time, our customers still must make one IVR choice (between Estonian and Russian, with Russian Annika being left to wait until we have gone through our learning curve with Estonian). After Estonian is chosen, Annika takes over, introduces herself and asks (with a recorder prompt) what was the reason for contacting us. That simple prompt was actually our first lesson learnt at the very beginning. During the first days, our prompt said – please tell us why you called, so I can transfer you the correct agent (with the call to action being in the beginning of the sentence) and that led to a situation where customers just stayed silent and waited for the transfer to happen.
We turned the prompt around (in order for me to transfer you, please explain briefly why you called us today) and the explanations started flowing in.


Though the language technology and science are of superior level for Estonian, our project faced several difficulties. The first challenge was Elisa-specific vocabulary, which simply didn’t exist in our national language corpus (words like smartphone packages, 4g dongle, Elisa, Telia or RCU, STB etc). Also, it turns out that people are really sloppy talkers when it comes to language and grammar. While chatting, we take the effort of completing the words and forming meaningful sentences, but this all goes out of the window when we call someone for some reason. In order to help our data scientists, Elisa’s specialists provided hours of domain-specific vocabulary by transcribing customer calls to be used as training data.

The second obstacle was the environment Annika had to start working in – calls. Most transcription modes excel in “lab conditions” (like transcribing speeches or radio programs) but phone calls are nothing like that. You get samples of completely different volume levels; a lot of background noises and people like to speak on top of and into each other (rather than politely taking turns J). These are all conditions that your bot never has to deal with while just chatting (though many of the language models are actually the same or quite similar). It was the hard work and years of development by MindTitan’s data scientists that led to the unique transcription model we are using today. Especially designed for phone calls and for such exotic smaller languages like Estonian, Latvian, Lithuanian or Finnish and others.

And thirdly we needed a model, a classifier, that would connect the customers question and the best suited agent, which again was a totally new task for us in Elisa. To simplify the solution, Annika “listens” to the client and transcribes (live) the “heard” voice into text (ASR or S2T, whichever acronym you prefer) and the rest of the witty part is calculated based on that transcribed text.

Since our clients had never had to explain anything to a robot before, we had no real data (examples) to work with. Therefore, we divided the launch into three stages.

Thanks to MindTitan’s previous work and know-how, we managed to get the technology ready for our first tests within 4 months in the beginning of 2019.

By May 2019 we were ready for our initial beta-testing (in live environment and with real customers) and started to let Annika answer its first calls – without transferring the calls, just directing them to specifically chosen super-agents who could handle most of the topics themselves.
This first phase was probably one of the most critical and with the highest value of importance to our technical team. This way we could safely test the technical solution while gathering valuable training data for the classifier (as customers knowingly talked to the robot).

Such approach proved to be correct choice many times as we discovered several issues that could only appear in live environment and had to take a few steps back in order to move forward.

Our beta testing phase was considered successfully completed by the end of August and we moved toward real live call-transferring Annika in September 2019.
Gradually upscaling, we have already reached the 100 000-customer milestone within the first 3 months and are soon in the phase where Annika has no limiting restrictions (working full-scale).

Besides technical issues we have constantly monitored what the customers felt and thought of such solution and to be honest – we were quite surprised of how positive the feedback was.
Our team has personally contacted those few customers who had been critical and their feedback has given us valuable insight into their thoughts and feelings.

Though the majority of our customers realize that they have been communicating with a robot for a really long time (since the old-school IVR is also a robot) they are pleasantly surprised that they no longer need to push the buttons to give the necessary commands, there is a small group of people that are not so enthusiastic at first.
Digging deeper into their feelings of discomfort taught us that it has nothing to do with robots per se, it’s just that they freeze for a moment because they don’t know how to talk to a robot.
Believing that a robot needs just straightforward commands, keywords – and they don’t know what the correct keywords should be! Therefore, feeling silly and confused.

We soon discovered that in those cases it was extremely important for our human agents to explain to those customers, that Annika is a state-of-the-art AI customer service bot, that doesn’t require you to speak through keywords. Quite the opposite – she is taught to process natural human language and detect the contact reason from a simple human explanation. Just like you would explain it to the person right next to you.

For both Annika and us as human beings, it would be quite hard to transfer a call if you just said “TV”. There are so many different options that might be suitable for you, that you really need to elaborate a bit more for the robot to do its job. (You might want to buy new TV set or have technical question about an existing one. Or maybe you need some help with the TV service or want to sign another contract with us…?). But should you explain that “I would like some help with setting up my new set top box” – you’ll be transferred within seconds!

We have encountered just a few customers who don’t like robots in principle; most of the customers who feel a bit awkward the first time, usually feel at ease after hearing the explanation and speak freely the next time. And taking into consideration that Annika’s accuracy is ~90% (compared to the old IVR’s 70 – 75%) it leads to 3 times less re-transfers, which also means shorter handling times and waiting lines, thus saving time for both sides.

In our case, Annika doesn’t talk back (yet), she just transfers the call, but we already have several new projects in the pipeline that will make the customer contact even more simple for our human agents to handle.
We are constantly moving in the direction of developing Annika into a cross-channel knowledgebase, which has proved to be useful for the business side, but also for the data scientists – rewarding us with record breaking short launch times for new solutions.

So, in conclusion – a successful AI application requires:

  • Several complicated machine learning models to be compliant with each other (MindTitan’s task)
  • Thousands of labeled calls and transcriptions for those models to work in your company (Elisa’s task)
  • The possibility and technological readiness to launch gradually – meaning a lot of open cooperation between the call center, data scientists and business side (including customer service team)
  • Really patient and calm customer service agents who are ready and willing to take on what ever might go wrong in the first days of implementing the new solution.

I am proud to state that Annika had all four criteria met and a strong team working in the sidelines till this day as we are still in the process of constantly improving and implementing new features.

Annika has opened amazing new opportunities for both customer service and product management departments - so stay tuned. This story will definitely be continued :)!

Seotud märksõnad
Uus Elisa Elamus ja Huub Mine vaatama
Vanemad TV-teenused
Liitumised enne 10.08.2021
Mine vaatama