Call Kuzya: How We Recorded an FAQ for Engineers

[Tomcat]

Every month we receive 20-50 thousand calls with questions about ATM maintenance. Most often, engineers call: to find out the status of the application, get access, check software versions, etc. Or collectors - to understand if there are any malfunctions on the ATM they are dissecting. The questions are the same in 90% of cases.

We took engines for voice automation and speech technologies, combined them and got a robot that helps people, connected it and put it on the line.

The functionality was the same as that of the operator, but the engineers fundamentally did not want to communicate with the robot. Even if it was a typical question "is everything okay with the ATM?" Then we changed the voice to a pleasant female one, tested it in AB with a male one - and the number of switches to the operator from the female robot fell: 24% of processing with Denis and 65% with Julia.

When we were thinking about how to automate the support service and process thousands of typical requests, we considered the option of creating a mobile application for these purposes. But we quickly dismissed this story for two reasons. The first is that not everywhere there is Internet from the mobile operator that the employee servicing the ATM uses - an engineer, a collector or another representative of the enterprise as part of the automation of access lists. Some ATMs are located in places where the only means of communication is voice. For example, the Zima station. There is an ATM there, but no access to the network. And at some factories, the Internet can be jammed due to the specifics of production and in the name of security. Accordingly, even if we had at least a dozen special mobile applications, they will not save in such conditions.

The second reason is that not all heroes wear capes, and not all service employees use smartphones. We cannot (and should not) force them to purchase devices that would support a mobile application.

In turn, we ourselves do not see the expediency of creating an application for outdated versions of different operating systems. What if a person has an old push-button phone? The same applies to options with different chatbots and support via messengers.

Accordingly, unlike a mobile application, a phone call that can be made from anywhere and from any device allows the support service to truly become more accessible.

When an operator is needed​

But there are cases when you can't do without the operator's help. These are moments related to organizational issues, and we have not automated them yet.

For example, an ATM needs to be left in some unusual state that is not described in any of the regulations or is rare. For example, the upper cabinet does not close completely. Or there was an identification error of the collector: the employee number, the group of ATMs that he services, and some other data did not match according to some feature. The share of such calls in the total number is small: with unsuccessful identification - 15.93%, with incorrect - 4.82%, with successful - 79.25%, according to data for August 2023. The statistics for other months are approximately the same.

In such a situation, the operator is involved. He figures out what happened. In case of identification error, we cannot trust the work to be carried out via a voice assistant, because we risk the security of the ATM, and security is the main principle of service. At the same time, we cannot lose a call either. Let it not be a bank client, who would be angered by a dropped call - this would damage the bank's reputation. But, leaving the collector without feedback, we risk leaving the ATM in a non-working state, that is, inaccessible to the client. The collector will simply leave, because his time is regulated.

The era before voice assistants​

Previously, before voice assistants, you could wait up to half an hour for a response from support services, since we are not a contact center, but a company engaged in centralized management in the service. During peak times, the number of parallel calls could be several times greater than the number of calls during quiet hours. At the same time, the calls could be of a formal nature.

For example, collectors called to report that they had finished work and to find out whether they could leave. They stood with money at ATMs and were nervous, waiting for an answer. It was necessary that all these calls, which required a formal answer without consultation, were not answered by operators.

What other problems did we want to solve by implementing voice assistants:

• increase network availability, reduce response time;
• save money that can be used to develop and improve the service.

Robot Julia and robot Sigma​

We created voice assistants and divided the responsibilities between them. We have an assistant called Sigma. She automates access lists (remember the story about the engineer who needs to get into the plant where the ATM is located). Everyone who needs to check the access list communicates with her.

The robot Julia talks to the engineers and collectors. She consults them, and this robot can answer the question of whether the collector can leave, whether he leaves the ATM in working condition. She will check the condition of all the equipment on the servers, whether everything is reflected as working. If something goes wrong, for example, there is a jam in the receipt printer, then she will offer help and tell you how to solve the problem.

Before the introduction of our voice assistants, there were often situations when collectors formally carried out their procedures, changed cassettes and left without making sure that the ATM was working. It was necessary to call the collection service again, pay for this visit. Cash cassettes are moving back and forth, recalculations are taking place at the cash desk — the bank is suffering losses. The appearance of Julia helped to get rid of this problem, because you can quickly call her and get all the necessary information. Voice assistants are always in touch and work 24/7.

Entering the service department​

The integration of voice assistants took about six months. Initially, the team included two developers and two business analysts. Now the core team is as follows: lead, IT lead, two business analysts, two business process automation specialists, four developers and a tester.

The work was built based on the Agile approach using the Scrum methodology. Depending on the backlogs, we combined two or three to 12 specialists into a team. We have key deliveries for the year, they are divided into quarterly releases with deliveries every two weeks and feedback. The work was carried out in stages. We needed:

• transform a business idea into technological requirements;
• implement automation at the contact center and integration with the speech technology center;
• provide initial scripting of the trees that the voice assistant follows, depending on what it hears from a person.

And we always had parallel teams for developing our incident management, a team for developing processing. When it was necessary to refine the product in related systems or to obtain information from them, the system analyst described the technical specifications for integration and data, after which he submitted requests for refinements to neighboring teams.

Today, we have moved further: most of the development has moved to streams that unite teams in a certain direction. For example, I lead the product cluster of the stream that develops all service products. There is a stream that develops processing technologies. There is a stream that develops software on ATMs. Within the stream and between them, we set tasks to create a single product, synchronize with the delivery cycle once every two weeks. Thus, the entire story with voice assistants is integrated into the framework of the general spectrum of the platform, the products that we support.

Tasks get into the backlog in three ways - these can be requests for automation, support, or feedback from our business customers. For example, they can calculate what a certain function is missing. For example, ask voice assistants to ask collectors: "How are you today?" For us, this is a meaningless task, but if the customer finances the creation of the product, we can implement such functionality. But often we get more useful feedback and gradually correct and improve something.

The speech is lively and categorical​

We dealt with all the problems associated with parsing through the speech technology center. For example, dividing calls into categories. Who called, why they contacted, how they fixed it - all this needed to be tracked in order to determine the segments where the most calls were made and where it was possible to resolve issues without human participation.

We made a simple IVR, where by listening for a long time and moving through a complex menu, calls were categorized by basic directions. Then the operators supplemented this information with algorithms for solving basic queries. We were primarily interested in the intersection points of call volume and the possibility of delegating them to robots as much as possible.

The center's technologies recognize Russian speech quite well, but not all callers have perfect diction. Some nuances of speech can "break" the voice assistant, and it will not answer. These are speech defects, unclear diction, a strong accent. We solved this problem by entering data through the keyboard. For example, if the robot couldn't recognize the ATM number because the caller doesn't pronounce numbers well, it will ask you to enter the number manually on the phone. We didn't invent this. This is a classic telephony feature. We simply integrated it into our service.

Engineers don't want to talk to a robot​

It would seem that there are no problems. We have set up the work of our voice assistants, provided for everything and did everything technically so that the issue could be resolved without the participation of an operator. But we did not take into account the human factor. Our collectors and engineers admitted that they liked to talk to a girl from the contact center - lively, sweet with a pleasant voice.

We retained the ability to connect an operator, and the specialists waited for his answer, ignoring the robot. The solution was found experimentally. At the very beginning, the voice assistant spoke with a male voice. It was the robot Denis. We did not attach any importance to what gender our robot would be. And the engineers were very indignant when the female voice contact center was replaced by Denis. I admit, we thought for a long time what to do. And then, as an experiment, we replaced the male voice with a female one. This is how our Julia appeared. By the way, her name is associated with our business analyst and key leader Yulia.

As a result, the number of calls that were transferred further to the operator dropped sharply. According to the latest data for September 2023, 65% of requests are resolved through a voice assistant and 35% of calls are received by operators. A year ago, in September 2022, robots processed only 24% of calls. And in 2021, the figure for September was several times less - only 2%.

Results of the work of JSC SSEO of a large bank

In general, all our experience with the implementation of voice assistants is the integration of one working technology into another. If you believe the history, the first voice assistants appeared in the last century, in the late 30s. For our part, we suggest how to expand their functionality.

Source