ChatGPT is on everyone's minds. A number of experts consider the dynamic development of the AI machine to be "historic" and a "real Internet revolution”. And indeed, the disruptive potential of the controversial language model seems enormous. For customer service, as well? In the following, we look at the example of Helvetia and the possibilities that ChatGPT already offers in customer service. The bottom line is: the machine and the performance it delivers is only as good as the human who controls it. And that is a good thing.
At the end of March 2023, Helvetia Insurance announced that it would use ChatGPT in the future to answer customer questions about insurance and pension plans. Clara is the name of the chatbot lady who provides information as part of a live experiment. And indeed, at first glance, the digital assistant provides reasonable answers to my question.
It is important that the software uses clearly defined web content from Helvetia Switzerland - for example, product pages or advice articles.
It is interesting that an insurance company, of all companies, is experimenting with ChatGPT. After all, it belongs - like the medical industry, for example - to the critical sectors in which an iron law applies to customer service: the answer or solution must always be reliable and quality-assured. In this case, the system's answer takes the questioner only a small step further. It does not address the question directly but refers to a generally formulated website. It does not address age as a criterion for taking out life insurance. In addition, it does not consider that I am calling from Germany and that the answer refers to Swiss conditions. In sum, the answer to my specific question is unsatisfactory - a service agent would have advised me better and more individually.
Many random checks and tests have shown in recent months that the AI chatbots are not always completely accurate with the factual correctness of their answers. People who happen to have the same last name are mixed up, and in an automotive context the systems mix up the specification of power in horsepower and kilowatts. But: the system learns extremely fast. In version 4 of ChatGPT, which was introduced in March 2023, many shortcomings could be eliminated. To avoid typical errors in the future, the system was intensively trained with human feedback. The fact that the accompanying circumstances of this training are subject to much justified criticism should also be mentioned.
Despite significant improvements in statement quality, the basic problem remains that the system works with the texts that are available and draws its own arithmetic conclusions from them. ChatGPT has no idea about real factual contexts, it assumes its "knowledge". Thus, crucial for the beneficial use of ChatGPT is the understanding of what the system can do, and what it cannot do.
Although chatbots have become an integral part of everyday customer service, classic dialog systems can still only cover limited functions and answer simple questions, e.g., about the ability to deliver products, make appointments, insurance policies, weather data, and so on. In this way, they relieve service teams of routine inquiries. No less, but also no more than that. A new generation of intelligent chatbots can also access the contents of a professional knowledge database, narrow down topics through queries if necessary, and provide quality-assured answers on this basis. Powerful chatbots are also able to actively perform small tasks, such as conducting diagnoses or making bookings. Especially in more standard subject areas such as IT, bots can fix problems themselves, such as clearing the browser cache or restoring network drives.
And where can ChatGPT be used in customer service? When it comes to text classification and generation for narrowly defined, less critical topics and consistently prepared content, ChatGPT can play to its strengths - for example, when quoting opening hours. When asked, "Are you open on Monday?" the system not only delivers the opening hours page, but responds specifically, "Yes, we're open from 9:00 to 20:00 on Monday." Other usage scenarios are also conceivable, such as generating ideas or suggestions for an advertising slogan. The automatic switching of contact persons is also conceivable. ChatGPT can also take over the training of other bots. Last but not least, an international furniture company is currently using ChatGPT to adapt the existing texts in the service to the company-specific "tone of voice". And of course, the language model also acts as a "small talk bot" when it comes to sharing everyday knowledge.
Conversational user interfaces (CUI) have become the standard in human-machine communication for customer-oriented dialog. It is therefore important to orchestrate the conversational services made possible by these interfaces.
A musical orchestra is a good image to describe the technology idea for the collaboration of different chatbots. According to the motto "alone strong, together unbeatable", several different chatbots are interconnected based on a multibot architecture and can thus solve even complex tasks.
The model divides chatbots into two roles, expert and lead bots. The expert bots provide information on specific specialist topics, while the lead bot (= the conductor, to stay in the picture) acts as a moderator and assigns the user the expert bot that is appropriate for his or her request (which is waiting to be used as a soloist).
Communication proceeds as follows: First, the lead bot receives the user's request. In the second step, the lead bot sends the user's request to each expert bot registered in the bot network. Each expert bot then determines whether it can do something with the query. If a queried expert bot recognizes the topic, it reports this back accordingly. Based on the feedback, the lead bot decides how to proceed: If only one expert bot feels responsible, the communication is handed over to this expert bot. If several bots show interest, the lead bot asks the user exactly which topic he is interested in. The user then selects accordingly, and the lead bot establishes the connection to the specific expert bot. If no bot is in charge, the user is asked to specify their query. If a connection to an expert bot has been established, the user now communicates directly with the expert bot.
The described architecture makes it very easy to integrate one or more ChatGPT bots. This can be tested on the USU chatbot page, for example. The content for such a "diverse" bot team is managed via a knowledge database - for top-class customer service.