It is companies that are guilty of things driving users mad when dealing with chatbots, according to George Fomichev, the founder and CEO of Endurance Robots engaged in the custom development of chatbots. It's all about the desire to save money. AI Conference interviewed the speaker in order to discover the concept of training virtual assistants, what chatbot is the most difficult to develop, and how much it will cost.
Interviewer: AI Conference (AI)
Respondent: George Fomichev (G. F.)
AI: Is a modern chatbot a code set or artificial intelligence?
G. F.: I would say it is rather a code set but already with an AI potential. I would even notice that it is grounds of weak artificial intelligence.
A lot of chatbots use the so-called heuristic search through a knowledge base. It is not the algorithm applied by search engines when seeking for key words. The heuristic search allows bots to determine synonyms, closeness of relations between words etc. In other words, it is a more complicated and unpredictable process: it is undefined what answer a chatbot with the huge knowledge base will provide. It selects an answer all over again each time.
AI: How chatbots are trained? Simply put, what is required for a chatbot to understand human speech?
G. F.: Indeed, chatbots understand human speech. To achieve this purpose, one often applies Google API called Voice to text: Google records the speech and convert into the text. Therefore, a chatbot works with the text rather than the voice.
Chatbot training is a complicated process. To establish the knowledge base, one should start the text processing operations: the search for synonyms, antonyms, relations between words, as well as the definition of word and phrase vectors… It is a kind of science (smiling - Ed.).
AI: Which chatbot is the most difficult to develop: for an online shop or, for example, for a bank?
G. F.: I would raise a question in a different way. It is difficult to design a seamless, efficient, and multipurpose bot. Developing a bot able to say “please call the hotline” is easy but to make it answer complex and comprehensive customers’ questions is more complicated… Moreover, to make a bot accept orders, record them in CRM, and provide a delivery status is even more difficult. The same can be referred to a bank. If the task is to create active transactions (transfers) using a chatbot rather than just to find out the exchange rate, it will require the integration into the automated banking system. This process needs strong programming skills.
AI: Some chatbots already help customers to conduct a payment, replenish an account, or transfer money. How safe are financial transactions using bots?
G. F.: One has not yet faced any obvious fraud accidents and losses on the part of customers. It is difficult to judge, as it is a new channel for communication between clients and companies. Meanwhile, Telegram offers quite a safe way of conducting payments as well as Facebook does (Messenger is meant - Ed.). In such a case, the point is whether messengers are safe. It is rather safe than not. Just a matter of habit: people should get used to the new method.
AI: Users frequently come across virtual assistants that fail to understand their requests and their conversation is limited to the phrases: “I have found nothing for you. Contact an operator”. Such a situation causes negative emotions for many people. Which tools are used to track users’ response when dealing with bots?
G. F.: Sure, I know this perfectly well. It infuriates, but this is the cost companies pay when asking incompetent developers for bots or desiring to save a lot on the development. In my view, chatbots should answer all the questions they are trained. If they answer wrong, they should be retrained. If they do not answer at all, one should download the answers, supplementing the knowledge base. As we can see, a lot of companies do not follow these steps. And it is wrong.
AI: Integrating bots into business processes requires some expenses: for developers, equipment, maintenance. What is the approximate payback period?
G. F.: Sure enough, many companies realize that bots are a new technology and they need to reorganize and modernize business processes so that a bot will be able to function efficiently. The average chatbot payback period ranges from 3 to 12 months depending on its features.
Equipment usually does not require any additional options. It is a cloud service able to exist perfectly on the server of service provider. In general, the implementation of this idea primarily depends on the customer’s will and desire to make a product. Besides, the expenditure part is not prohibitively high, although it can reach millions of rubles (for instance, in case of a bank). Thus, we recommend everyone to start from a simple MVP (pilot) in order to appreciate the possibilities of this technology.
At the conference on April 19, George Fomichev will talk about chatbots in businesses even more detailed. Do you want to discover more?