Quality of bots

Piotr Kempa, creator of Primebot, is interviewed by Karolina Kania, project manager at Voice Contact Center.

Karolina Kania: What determines the quality of a bot?

Piotr Kempa: In my opinion, the quality of a bot should be measured by how quickly and ergonomically it guides the caller through our defined business process. An ideal bot should not make a mistake in interpreting what a human has said. And that is why it is so important to create a good engine and teach it well.

KK: Does every company that creates bots have its own proprietary “engines”?
PK: This is not usually public information, but as far as I know most companies on our market use their own proprietary solutions in this area. However, there are ready-made solutions or solutions that make it easier to create your own engines. Firstly, you can use ready-made programming libraries in popular programming languages. They often help in creating bots, doing some of the work for the programmer and imposing some specific restrictions in return. And secondly, there are ready-made solutions in the form of cloud services, practically ready to use. One of the most popular such solutions is for example Google Dialogflow. These solutions take care of most of the work for us, but they impose more restrictions and, of course, are often paid for. There are also often significant limitations in the supported languages.

KK: Is there anything that bots cannot do?
PK: Bots often fail to evoke empathy. If our process is to elicit empathy from the caller, chances are the bot will perform lower than a human. Similarly, bots lack empathy – in the case of a complaint, they may say that they understand that the caller is upset, but this will often not have the expected effect of calming the caller. Another element a bot may not have is persuasion. If our sales process, for example, involves persuading an interlocutor who does not want to buy, bots will also perform much poorer here compared to humans. It is simply much easier for a bot to refuse. In relation to a bot, people are simply more honest and say exactly what they want. In other processes this can also be an advantage.

KK: What else do bots have problems with?

  • Voicebots will most often have trouble where the speech recognition service has trouble. Therefore, special attention should be paid to such dialogue elements as:
  • Recognition of strings of numbers and letters (for example, registration numbers, VIN numbers)
  • Recognition of rare first names or surnames (with popular ones there is no such problem)
  • Recognition of more complex email addresses
  • Recognition of complex addresses.

KK: What about chatbots?
PK: As far as chatbots are concerned, they will have problems with typos, ungrammatical writing or very long sentences. Emoticons or unusual abbreviations can also cause problems. Generally speaking, the most common problems for bots are not so much in the understanding of the bot as in the wrong input. In the case of chatbots, it will be errors in the text (such as typos), and in the case of voicebots, it will simply be errors in speech recognition.

KK: And what about other languages?
PK: In other languages there are often even more speech synthesis solutions available, for example IBM Watson. Of course, the ones we use have a very wide range of supported languages. The Primebot solution allows for building bots in other languages, the only limitation here is the availability of synthesis and speech recognition in a given language and having employees fluent in a given language. Here, Voice Contact Center can boast a long-established Lingo multilingual service. This allows us to build bots in other languages as well as provide their support with consultants who speak the language. We are currently working on technology to build a multilingual bot that can seamlessly recognize the language of the caller without asking “what language do you want to talk in” and seamlessly switch to the language of the caller. Our solution, which is currently under development, should enable the creation of a bot supporting up to 4 languages in this way. Of course, the set of these 4 languages will be different for each bot.