It’s July 6, 2023, and here we are. The ChatGPT tsunami that started about nine months ago, while still quite potent, is abating. According to Similarweb, the traffic to the ChatGPT website decreased by 9.7% from May to June, and unique visitors dropped by 5.7%.

There is no denying an extraordinarily high interest in generative AI, and ChatGPT doesn’t cease to amaze. However, as many have found, there is no magic wand, and getting to professional usefulness requires a significant effort.

There are many use cases where ChatGPT brings value and enables productivity, and ample literature provides details. In the customer experience space, at this point, very few companies have implemented generative AI technologies and even less for customer service. Why is that?

Simply put, it isn’t effortless, and the necessary skills are in short supply. Most enterprises do not have the staff to pull this out, which is why Microsoft launched the AI skills initiative to help people to master the usage of AI.

To commission a ChatGPT or Bard-based customer chatbot, here is what you need to do in broad terms:

-       Identify the use case: what are you going to automate?

-       Make the data available to the application: which are the data sources, how do you access them, do you keep data in their native place or cache, duplicate, and so forth? In other words, like any well-behaved system, you must have a well-thought-out data and storage strategy.

-       Build the user interface, including dialogues and navigation. This is where things can get hairy. ChatGPT or Bard knows very little, if anything, about your product names or topics your customers want to chat or talk about. To inject this knowledge into the model that your dialogue will use, you must train it. It’s tedious and requires extracting phrases and vocabulary from recorded conversations or chats. Analyzers and topic modelers can help but are not free and take some training. Depending on the size of your project, annotating data manually or via tools is a cost, timing, and benefits decision.

-       Build the prompts; this is the core of using generative AI. Your application will communicate with ChatGPT or Bard via prompts – very much like you would do on the websites, except that the prompts will be sent via an API. How you construct these prompts is critical to the success of your chatbot. The better your prompts are, the better the response will be, and you will receive less generic mumbo jumbo and hallucinations. Concepts like prompt chaining and chain of thought are paramount for getting it right.

-       Decide if you need or want the ability to escalate a dialogue to a real person. Nothing is more frustrating than to be stuck in an endless loop, and unfortunately, this happens too often with customer service systems. Transitioning from a chatbot to a person implies behind-the-scenes integrations between the AI application and the contact center software, and it is most often not trivial.

-       Test the application. Assess the quality of the answers using either synthetic or real data. This means, for example, returning to the recordings, replaying the customer questions to the chatbot, and verifying that the answers are correct and helpful.

-       Monitor and optimize. Your enterprise data constantly change, as your products, services, and people do. Unless you constantly update and upgrade the model with the latest data, the risk of inaccuracy or of the chatbot not understanding the question increases, leading to unacceptable service.

This list is not exhaustive, but by now, you probably have a sense of what it takes to deploy generative AI for customer service. The good news is that tools that facilitate the process exist and can greatly simplify and speed up the development and implementation.

There is no doubt that mastering AI can provide competitive differentiation and lead to higher productivity levels. A sound understanding of technology and planning is essential for success, like always in IT.