specific expertise, their geographic location, their availability of appointments, and so on. Which of these should be proposed to the patient? In the case of the third configuration, a GenAI tool can leverage the available information about the patient and the providers to recommend a few options to the human operator. The operator in turn has the responsibility to choose from these options. In other words, the “heavy lifting” is done by generative AI (narrowing down the choice set from thousands to a handful), but the final decision and responsibility rests with the human. Moreover, the human operators might have to add “a finishing touch” to the solution they recommend, such as explaining the choice to the patient. This configuration plays to the key strengths of an LLM: it is good at generating options, but, due to hallucinations, it benefits from having a human be the last point in the customer support journey. Configuration 4: Chatbot preparation with human operator responding This configuration is similar to configuration 3, except that the LLM only recommends one solution. Its focus therefore is not on generating alternatives for the human operator to choose from, but rather to prepare the final answer to the customer as much as possible. In our use case, imagine the LLM preparing a customized instruction message to the patient preparing for surgery. The message might include the arrival time, when to take (or not take) specific medications the patient is on, and post-surgical instructions. The role of the human operator in this case is to simply read and approve the message and potentially make minor edits. This configuration is similar to the use case of medical providers using an LLM to summarize an encounter with a patient and then only reviewing the documentation for accuracy before storing it in the patient’s electronic health record rather than typing up the report from scratch. Or, think of a radiologist who uses GenAI to read an image and prepare a first draft of the report automatically but who makes the final sign-off on the report. The key value proposition is to significantly lower the touch time of the human operator, thereby improving the efficiency of the customer support organization while leaving the “final word” with the human operator. Moreover, from the customer’s perspective, the experience is one of directly interacting with a human and receiving personalized support. In configurations 3 and 4, the chatbots are doing a lot of the heavy lifting but ultimately the human operator is responsible for providing the customer support. In the final two configurations, we see the primary responsibility for providing customer support shifting from human operators to the chatbots, with the human role becoming more supervisory. Configuration 5: Chatbot response with real-time auditing by human operator