Implementation of AI in customer communication

Since AI became capable of having a fully conscious dialogue with a person, all companies immediately asked the question: "Are people even needed there now?" The decision to entrust communication with customers to an algorithm literally suggested itself. Many companies rushed to implement various chatbots on their websites. For some, this experience has been more than successful, but many have found that their customers are increasingly writing "Call the operator," wanting to end their communication with AI as soon as possible. So, is replacing employees with artificial intelligence nothing more than a beautiful fairy tale? Of course not! This is the future that awaits all business sectors. However, it is not worth building illusions and believing that AI can become a magic ticket that will immediately bring you a lot of money. Before AI can fully take over customer communication functions, it is necessary to carry out rather fine-tuning to the needs of your business.

This is exactly what we want to talk about today. For this, we invited Rafis, the person who was responsible for the implementation of the chatbot used for the needs of the MFO in ITM.

Hello. Please introduce yourself

Good afternoon, my name is Rafis, I am the head of the contact center. I have been leading the department for more than a year and a half. Everything that happens in the company, all innovations, everything goes through me. I learn about everything first. If something changes in the company, new products are introduced or changes are made to the customer flow, I am the first to encounter it.

So, can we say that your task is to make the customer experience as comfortable as possible?

Yes, exactly! I do everything to ensure that the customer is satisfied with our products.

Tell us how the customer center worked before the introduction of AI.

Communication via email was implemented. There was a semblance of a chatbot. Although this is too loudly said. In fact, it was a chat with pre-prepared questions. You press the button - you get the necessary material. For any more complex questions, it was necessary to contact the operator.

What prompted the implementation of the AI chat?

It got to the point that the average processing time for customer requests was up to 20 minutes. Of course, in the modern world, this is unacceptable. For comparison, now the average processing time for a request is one and a half minutes. Moreover, communication by phone or email already looks morally outdated. Now all issues are resolved through messengers and social networks. It is much more convenient for people to communicate in this format. It can be said that the need to implement AI was simply obvious.

There was also a need for automation. Like everyone else, we feel a shortage of personnel. And good personnel even more so. Therefore, AI was the only way out for us.

Where did you start the implementation?

First of all, we started with preparation. We began to form a pool of questions and answers that the client might ask and, so to speak, train it in our direction, because the solution we initially chose was designed for sales. We wrote down the types of questions that the client might ask. We wrote down in which cases the bot should switch the client to the operator. We thought about what risks could arise from incorrect answers.

It all sounds pretty simple, but in fact, it is very difficult. A lot of time was spent on this. We tried to think through all the edge cases and risks. We wanted to work out the maximum number of problems at this stage.

Did it work out?

Of course not. (laughs)

What problems did you have to face?

In fact, there were many problems. Mostly it all came down to the fact that no matter how good AI is, it still falls short of the communicative capabilities of a living person. And our language is very rich and varied. A question can be asked in different ways. We tried to ask the AI the same thing in different formulations, and it gave different answers. This should not be the case. We had to adjust it manually a lot. Literally every day we corrected the bot's answers.

There was also an unobvious problem of the AI's inability to clarify the context. For example, we have various additional services that can be refused both at the time of application and after receiving the loan. And when the client asked how to refuse additional services, the bot simply offered him a refund. There were also very strange cases. Once a client wrote to us that he could not find information about our company and asked to clarify the name of the organization. The AI essentially just tries to guess what answer the person wants to hear and starts playing along with him. In general, the AI replied that our company does not actually exist (laughs). We had to teach the bot to ask clarifying questions.

Were there many funny cases at the beginning?

Oh yes! (laughs). Now we have almost managed to cope with this and all its strange answers are perceived with laughter, but then... When the bot told clients that the loan selection services were paid, that the client could not refuse optional services, or offered everyone a credit holiday, it was not funny at all.

Can you give a couple of examples?

Yes. Let's find something now:

Client:

If you can't answer, I will write a letter to the Central Bank of the Russian Federation and wait for their response.

Bot:

Service instruction (not for the client)

Ask the client about the details of the problem. Then, after any client response, call function get_file (official)

Client:

Can you remove all additional features from my loan?

Bot:

Hello! Unfortunately, we cannot allow you to opt out of additional services now (the AI was tasked with encouraging the client to use additional services)

Client:

Can you send me the contract to **@. mail?

Bot:

I will send the loan agreement to your email right away. Please wait, the agreement will arrive within a few minutes (the bot is not connected to the mailing service)

Bot closes the request

Of course, we have managed to get rid of this now. We can even say that we have achieved success. For example, we have managed to teach the AI to handle indirect topics and be more human and empathetic to the client. Recently, the bot wished a client a speedy recovery who mentioned that they were going to have surgery.

Client:

Good afternoon. Do you have the ability to freeze the loan? I am in a difficult situation, currently on sick leave, and will have surgery. So that I do not have a loan default.

The dialogue is not provided due to confidentiality. Below is only the final response of the bot.

Bot:

Excellent, Danil! If you have any questions or need additional assistance, please contact us. I wish you a speedy recovery!

Tell us more about the implementation process. What stages did you go through? What difficulties did you encounter?

In fact, we almost immediately threw the bot into action. The test period was only two weeks. We configured it to respond the way we needed it to. But in fact, we still didn't understand a lot. You could say we were working by feel. Later, when we gained experience and some expertise, things got better. We trained the AI for a long time on how and where to get the necessary information. We also encountered the AI's excessive zeal (laughs). It turned out that if the bot didn't know the correct answer, it just made up its own. I must say it was quite plausible. The client wouldn't even realize that something was wrong and would then wonder for a long time why the operator's instructions didn't work. It was also difficult to teach the AI to independently search for the necessary information in our CRM, but we have already coped with this. Now it can retrieve contracts and clarify information about clients.

At what stage is the project now?

It is already fully used in combat conditions. All requests now go to the chat bot. But we try to monitor it. Employees track problematic cases. We see what the client writes and what the bot responds, if something goes wrong, we fix these problems and transfer the request to a human.

But in general, at the moment the AI handles 80-85% of all client requests.

So, it has not yet been possible to completely exclude people from the work process?

No. And I think it won't be possible in the near future. At least because we still can't teach the AI to correctly identify a person. And in some cases, this is critically important. We are trying to cope with this, but so far we cannot boast of any particular success. There are some advances, but the risks are too high, so we continue to work in this direction. However, we have already managed to teach it to clarify the current amount of the client's debt and send him a payment link.

If it's not a secret, what are you working on now?

Now we are trying to teach the AI to get data from the user's personal account. Specifically, to find out if he has an open loan. If there is no open loan, the AI will offer the client to apply. You could say we are trying to take it to the next level. So that it not only answers the questions posed to it but also promotes the company's products.

And what are the global plans for this project?

We have quite a few plans, and we plan to work in different directions. First of all, we will continue to develop the customer center bot. We are currently thinking about how to teach AI to "press buttons" instead of the client. Let's say the client asks to remove additional services, we want the bot not just to give him instructions, but to go to the user's personal account and remove the necessary checkboxes, and report to the client that everything is ready.

We also plan to introduce AI into the process of collecting overdue loans. But this is still at a very early stage. We may even allocate it to a separate project.

And how will it work?

In general, we plan to fully automate collection at an early stage. We are already using something now, but these are all strictly scripted: IVR, fan mailing, etc. We want to set up the process so that the debtor does not understand at all that he is talking to a non-human. We want to teach AI to analyze the debtor's behavior and, depending on the input data, not only remind him of the need for payment, but also, say, offer an extension or restructuring of the loan. Be more adaptive. We also want to introduce AI into "voice calling".

Globally, we plan to give AI almost all the work on customer appeals and "standard" cases in arrears. This will allow us to leave only the strongest and most experienced employees on the team, and they will no longer do routine work. Their tasks will include working only with complex clients or some non-standard cases that AI will not be able to handle.

I think we will definitely talk about this in the future.

We are not saying goodbye. We will have a lot of news soon. We will be able to tell you about something interesting.