How to cause customer`s desire to communicate with a bot by phone. Practices of "Tinkoff"

Speaking at the TAdviser SummIT online conference on November 24, 2020, the head of a product of voice Tinkoff robots Leonid Kolybin shared experience as in bank created the voice chat-bot able to keep up so the conversation with the person to cause in it desire to continue communication.

Leonid Kolybin

From the pragmatical point of view, "pleasant in every respect" the chat-bot was developed to maximize quantity of successful calls to clients with a request to estimate quality of the rendered service, Kolybin explained. "Tinkoff" well knows the cost of outgoing calls for a conversation "properly" of the operator with the client – at office of bank 2.5 thousand operators and 10 thousand more in a cloud work. After implementation the voice chat-bot performs every month 6.7 million calls, including negotiations of similar robots which are implemented at corporate clients of bank. Now it closes the different business directions: client service, recruiting, service of collectings, etc.

By Leonid Kolybin's estimates, the chat-bot was three times cheaper than own call center operator: 4 rub/minute against 12 rub/minute. However, to reach such ratio cost/quality, the Tinkoff command had to work at a NLP system, having integrated it with the mechanism of machine learning.

The normal chat-bot from the homepage of the website categorically did not suit developers.

In general it is very difficult to measure client experience in money. But, provide that the client called bank to open the card, and he was met by the "idol" who is badly understanding words and inadequately reacting to them. What will be with your conversion? - Leonid Kolybin shared the reasons.

In bank developed own system of speech diagnostics Tinkoff.Voicekit which as the specialist claims, on open text data sets shows quality of recognition, is almost twice higher, than at ready-made products. He explains it with the fact that commercial engines of speech recognition most often implement work with so-called regular expressions, i.e. with the combinations of words which are often found in the speech.

Slide from Leonid Kolybin's presentation

However in a situation of an incoming call when the client is taken unawares by the offer to talk about the received service, the analysis based on regular expressions works badly: there are interjections, pauses, "low", etc. The computer system will not be able to understand what signs contain in the speech of the client: consent or disagreement. The combination of own algorithm of NLP to machine learning gave falling of a share of errors at speech recognition by 5 times, the bank representative of "Tinkoff" provided data: 6% against 31% when using traditional methods based on regular expressions.

But high quality of recognition in itself does not guarantee yet desire of the person to continue a conversation: answer questions, listen to the offer. Leonid Kolybin's team began to experiment, modeling different "natures" of a bot. A part of hypotheses failed as, for example, the assumption that the male client will be located to talk to a languid female voice on other end of a wire. Attempt to press on pity – "Please, answer what costs to you? And very much it is necessary to me!" - yielded, in general, good results, but created an image of the "begging" bank and therefore it was rejected.

The image of "polite and persistent" bank was recognized the best, though it fell short of the conversion level (successfully perfect calls) of the person operator in poll on subject of client loyalty. The idea to check a hypothesis of influence on an outcome of a conversation of degree of a zamotivirovannost of the operator and a bot became inspiration. The analysis of negotiations of operators showed that in some cases they swindled, increasing quantity of successful calls.

Slide from Leonid Kolybin's presentation

After recalculation of conversion taking into account correctly passable NPS polls the robot overtook the operator, having shown 24% of conversion against 23% at the operator.

Leonid Kolybin opened some acceptances which allowed to create the chat-bot able is interested to keep up the conversation and to persistently finish the put scenario. First, this ability to distinguish pauses in words, distinguishing them from end of the speech. Technically it is a difficult task for a NLP system. Secondly, filling of pauses with the interested poddakivaniye: "aha", "so", etc. Thirdly, support of open questions like "And why such low mark?". The used psychological acceptances are rather simple, but are uncommon in technical implementation.

As a result we received the robot rather human and very effective, and besides much cheaper, than the call center operator, - Leonid Kolybin summarized. - The same technologies are used on incoming calls of contact center: a system precisely defines a problem of the client and transfers a call to the necessary employee, minimizing the number of excess switchings.

There is one more direction of use of own NLP technologies of bank – support of work with objections. It is difficult section for the operator working, for example, with collectings.

Automate such negotiations very difficult, but the intelligent system can seriously help the operator, creating hints for transfer of a conversation to the necessary bed, - Leonid Kolybin explains.