Why the New Generation of AI Customer Experience will be Better than We Think
Most of us have had some interaction with an AI chatbot in the past, and found the experience frustrating, and not much better than using the phone-based decision tree (press 1 for billing, 2 to report a service interruption, etc.) This is largely because earlier versions of AI that have been in use for some time, are often based on the same model, using pre-scripted prompts, and searching for keywords in your response to sort your issue into one of these predefined categories. And that is how we have come to perceive the capabilities of AI in customer service.
The so-called first wave of AI was based on sorting and categorizing. Applications of AI included search results, ads served to you, etc. The term “generative” means generating content, such as text, but often that has been in the form of pre-scripted prompts to aid in sorting, both the category of your issue, and the FAQs presented to you as a solution.
The coming wave of AI Customer Service will be using the Transformer-based Foundational model, trained on domain-specific data, and will be better at creating new solutions to newly presented issues.
The handoff to a human agent will be more seamless because the human agent will be prompted with the information you already gave to the AI entity, and in some cases, AI may be integrated into business processes to allow you to complete basic transactions without the need for a handoff to a human agent. The use of AI for customer service applications is a fast-growing field with both perceived advantages in cost and accuracy, and concerns about the loss of human feel in customer interactions.
Leaders in AI Customer Support
Current leaders in the AI Customer Support field include NICE with their CXOne platform, and Forethought with their Product Support GPT, each with several Fortune 500 clients, and each using a foundational model, which is then given the domain specific data, to be effective in a given business model.
Transformer AI Models
A big shift in the efficacy of Natural Language Processing began around 2017 with the introduction of neural networks called Transformers, based on the Self-Attention principal, as described in the paper Attention is All You Need written by a Ashish Vaswani and his team at Google Brain.
The problem that Transformers excel at solving, is taking each word in a passage, and referencing it to all of the other words in the sentence or paragraph to give it context. For example, to understand who or what is being referred to by the words “it” or “He”, your mind must be able to reference the rest of the sentence, when it arrives at that word, to assign the intended meaning to it.
Without explaining a lot of the computer science behind it, Transformers provide a more powerful model than the previous RNN Recurrent Neural Network, and LSTM Long Short-Term Memory models, with regard to Natural Language Processing. This Transformer model became the foundation for BERT (Bidirectional Encoder Representations from Transformers) which is used in Google's search algorithm to process the meaning of user input, as well as in GPT (Generative Pre-Trained Transformer). In Chat GPT, after processing the meaning of the user input, the AI engine generates a customized “story” by adding one word at a time grabbing more context as it does so, from the huge knowledge base it has been given (in some cases, the entire Internet).
The Need for Domain-Specific & Company-Specific Data
There are certain areas where creativity is positive, for example, in writing poetry or fiction, and GPT is useful for this. However, in certain realms, such as legal or healthcare, it is critical that the provided answers are correct. While the “pre-trained” part of Generative Pre-Trained Transformer may contain enough information to make a conversation believable, it will not know a specific company's product offerings, return policy, etc.
The process of introducing usable data pertaining to a business to an AI application, can be a lot like training a new employee. Previous call center interactions may be logged until it begins to develop a sense of the correct responses to various situations. It's sort of like having a new call center employee shadow and observe existing employees on their first day, but on a much grander scale. With Generative Pre-Trained AI, however, you can go much farther than conventional customer support, by incorporating knowledge from the company’s engineering, manufacturing, technical documentation, as well as knowledge from the company’s marketing and customer experience departments about customer roles and personas and contexts. Quality Control?
Paranoia as a Useful Source of Checks and Balances
How do you keep AI from generating the Wrong Answers? In an interview with Forbes, Deon Nicholas, founder of Forethought, responded to a statement from the Center for AI Safety often paraphrased as predicting AI's potential role in future human extinction.
These concerns are very much a common premise in science fiction writing. For example, in the post-apocalyptic Terminator movies, the day that fictional AI system SKYNET gained sentience, is viewed as the beginning of the battle between humans and androids, with an imperative that human agents be sent back from this future to stop AI world domination at all costs. Science fiction writers only need a thin premise on which to base a human struggle. The part that they leave out of the logic is primarily motive. Does the enslavement of humanity benefit a sentient software entity? Not really.
Nicholas points out that the process for developing early forms of generative AI was based on pitting networks against each other, in a configuration called GANs (Generative Adversarial Networks). In this scenario, there is a network in charge of testing another network and searching for flaws, or output that looks unconvincing or inaccurate. As the tested network improves, the testing network becomes redundant, and the output of the tested network is deemed ready for use. He suggests that similar adversarial networks could, or should be used in the future, to test that the output of Generative AI remains factual and useful, and that this type of paranoia may be a valid source of test criteria.