However, each additional chatbot in the chain impacts response times since the previous chatbot needs to finish generating its response, leading to a sequential dependency. The user request is first sent to the first chatbot in the chain, alongside the system prompt and previous conversation. Then, the full response from the first chatbot is awaited. Once completed, the second (next) chat bot can review the answer based on its instructions. If it is the last chatbot in the chain, the revised answer can be streamed (sending smaller paragraph chunks) instead of awaiting full completion to improve responsiveness. For simpler rules, such as stripping out specific words, a buffer for the response of the first chatbot could be used so that words can be validated before they are returned as part of a streaming response. The third strategy for quality assurance is to “put a human in the loop,” something that we will discuss at length in the second part of this paper. Part II: Integrating chatbots into workflows How much autonomy is granted to the chatbot and how much human labor should remain in the workflow depends on the business goals behind the chatbot implementation. In this second part, we will first discuss what organizations might aspire to get out of a chatbot implementation followed by different workflows that determine how human operators interact with the chatbot. Business Objectives Broadly speaking, a chatbot can exist for two reasons. First, automation can lead to efficiency gains while maintaining the same level of quality / providing a similar customer experience. When resetting a password or updating a mailing address, for example, a chatbot can achieve the same outcome for a fraction of the cost of a human to manually make such changes]. Second, automation can also be used to enhance the customer experience and the perceived quality of the service provided. For example, a chatbot can be used to engage patients and increase their compliance with their medication regiment in a way that would just not be possible with a non-automated process. Or a student studying geometry can be tutored by a chatbot and obtain a learning experience that otherwise might only be available for those students who can afford private tutoring. Another advantage of the chatbot is that the support of the customer can typically be provided immediately, saving the customer from spending endless time in a queue waiting to be served by a human. When deciding to implement a chatbot, it is critical to know that they can affect the accuracy of a service and its ability to consistently adhere to a quality standard. For example, hallucinations are a common quality concern for chatbots. Also, customers might have an inherent preference for human operators and human operators might be more knowledgeable and better able to resolve a customer problem. Humans might also be averse to receiving advice from a chatbot (Dietvorst et al. 2015) even though they might be unable to tell the difference between a human and an AI-powered response (Meincke et al. 2024).