As AI advances, customer contact professionals cling to 'empathy' – after all, that's what differentiates humans and software? Or is that more nuanced and the number of contact center employees will be decimated in the coming years? To answer that question, we'll dive deeper into how AI like ChatGPT works. Part 4 of a series about the significance of AI for customer contact.

Can a chatbot come across as empathetic as a doctor? “Being empathetic doesn't come naturally or easily to all people. Recently, more and more research has been done on 'artificial empathy',” says Art Ligthart, Chief Digital Transformation and partner at Y.digital.

We talk to Ligthart (originally IT architect) about artificial intelligence or AI. Goal of the conversation: a clear answer to the question of whether AI is (or will be) able to respond empathetically to customers. Ligthart starts by stating that AI has become a confusing term: “It is quite a catch-all term. There is no such thing as 'the AI ​​application'.”

AI is a black box

“AI is a black box for many people,” says Ligthart. “You have to open it to understand something about it. Then you will see that it contains all kinds of different software components. It is therefore not a machine that performs one operation, but a multitude of modules that all have their own task. Depending on the command you give, those parts may or may not go to work. Sometimes, for example, three algorithms are used at the same time because that is convenient. Some of them are fully explainable and transparent, others are not. One model is trained on a closed knowledge domain, the other model works on the basis of a database and machine learning. Sometimes data is retrieved and enriched in the black box or the dataset is adjusted in the meantime.” Ligthart therefore advocates that every organization that uses AI opens the black box and describes the underlying architecture.

Workflows under the hood

“Also for AI in customer contact, immediately after an application has received an initial question from a customer, for example, dozens of sub-assignments can be carried out immediately. The first module looks for emotion, the second for intent recognition, a third already retrieves products from a database, a fourth looks at the customer's financial value, a fifth looks at the contact history in the CRM system, a sixth interprets the customer's input and, based on an LLM, looks at what the next sentence might be. You may want to be able to carry out interim checks for some of those partial results. In short, there are large numbers of workflows under the hood of AI.” This also applies to ChatGPT, where determining word order based on probability is combined with a number of other models. [link to Knowledge article – Empathy in customer contact: employee versus AI]

Convergence: combinations of 'smart' technology

That complexity in AI is only increasing, says Ligthart. This is partly because new possibilities arise that can be combined with the existing ones. A good and current example is ChatGPT, where all kinds of new functionalities are now 'put on top'. Think of the multimodal AutoGPT, which can perform complex tasks without human intervention using ChatGPT. Popular example: programming an app based on a prompt of two words: 'weather' and 'app'.

Stacking and combining is therefore already happening. “If you were to do that with sub-applications such as emotion recognition, intent recognition, analysis of contact history and customer value and determining a good output based on an LLM, you could develop an AI solution for customer contact. realize that comes across as very empathetic,” says Ligthart.

There are still a few bumps

Not all data used by AI is neatly stored in databases. “Structured data has long been the starting point in information management. But more and more unstructured data has been added. Think of all the conversations we have with customer service or social media and all the content on the internet,” explains Ligthart. “Moreover, that data – with differences in age, reliability, meaning and formats – is stored in all kinds of different systems. For a long time, information specialists thought that an exhaustive and uniform definition of all these concepts was the solution to properly interpret both data and content. But that is not at all in line with the way people communicate: when we use words, they sometimes seem the same, but the meaning can be slightly different. The concept of 'entrepreneur' means something slightly different within the Chamber of Commerce than within the Tax and Customs Administration. The Tax and Customs Administration alone has seventy different definitions of the term wage. You have to devise AI solutions that can deal with this.”

LLM alone is not enough

LLMs are not good at dealing with different word meanings – they rely on immense data and raw (and expensive) computing power, but lack the ability to make good use of the language, words and domain knowledge that already exist within organizations and even though it has been described. Think of professional jargon, dictionaries or taxonomies, reference models, catalogues, data models of data in the systems, legal concepts, et cetera.

The missing puzzle piece is therefore a second technology to achieve a good output, says Ligthart. “That technology has been around for a while and is referred to as the 'knowledge graph'. Be hereby concepts are modeled and relationships between concepts are established such as customers, orders, places, events, etc. Concepts are thus placed in context. Synonyms and homonyms (words that look and sound the same, but have different meanings – think 'bank') can be modeled. In this way, for example, you can connect concepts from customer service processes to legal concepts. Knowledge graphs also provide a language model, but specifically of concepts and their mutual relationships.”

A semantic layer over all systems

With such a language model in your back pocket, you can then create an AI solution that can interpret the words in both structured and unstructured data much better. “If the word 'wage' is found, the AI ​​knows that there are many different variants of it, and will look further for words that, according to the knowledge graph, help to determine the specific meaning of 'wage'. If the AI ​​is looking for the intent of a customer calling or chatting, the AI ​​will look at the knowledge graph and then ask relevant questions to deduce which of the sometimes hundreds of intents the customer probably has.”

Such a language model can be regarded as a semantic layer, which you as a user can ask questions to, says Lighart, and which then consults structured and unstructured data to find the relevant answers. “You always make such language models in collaboration with the knowledge holders in the organization: senior employees of, for example, customer service, finance, laws and regulations or operations. And the nice thing is that LLMs can also be used to help compile language models and knowledge graphs from existing data and content. It is then important to validate the results together with the knowledge holders. This creates an 'enterprise knowledge graph (EKG), which represents the concepts and knowledge of an organization. Both old data from all existing systems and new data can be interpreted and processed; the knowledge graph grows over time and will adopt new concepts and knowledge.”

LLM + knowledge graph = promising

“Google was one of the first major tech companies to start using knowledge graphs. The knowledge graph that Google has built is very large, because it consists of large parts of the internet. Until now ChatGPT hardly looked at meaning and especially at the probability of the word order, when using knowledge graphs, the software also looks at the meaning of concepts, relationships, synonyms and related texts. The ideal is an intermediate form in which you combine both techniques.”

That combination offers prospects for customer contact, says Ligthart. “Anyone who would make a knowledge graph of all his customer contacts and map out the most important intentions, can then quickly find out the intention based on a customer conversation. Bring in the available data about that customer and tailor the conversation. Combine this with emotion detection and you can generate an artificial empathic response.” And if you add real-time digital translation to this, you could say that the customer contact of the future is ready.

Hidden intentions remain a problem

Then finally the intent recognition: what is the customer's question? Or what is the question behind the question? At this point, Ligthart considers it unlikely that humans will be completely defeated by machines. It is difficult for software to recognize hidden intentions, such as sarcasm or threats to cancel, while the customer wants recognition and perhaps a good offer.

“As I said, if you were to combine all the tooling that has been developed so far, you can realize AI solutions that come across as very empathetic,” says Ligthart. “The question, of course, is whether you should want to.” In other words: in addition to all kinds of ethical questions and questions about validity, quality and security, there is also the question of whether we should aim to replace employees with advanced AI solutions from an HR and CX perspective.

As far as Ligthart is concerned, the current speed of developments is also a problem: “The AI ​​world is now putting the turbo on all kinds of applications, but therefore also on all kinds of fake news and fraud. That is a huge risk factor. We are now at the top of Gartner's hype phase: we still have to find out where and how AI can be successfully deployed. The AI ​​toolbox is quite well-stocked for empathy,” concludes Ligthart, “But it is certain that AI applications will certainly affect the work of customer contact employees.” (Ziptone/Erik Bouwer)

Also read: “The Netherlands should have its own 'large language model'” 

Dossier, Featured, Technology