The potential for a chatbot to form a long-term relationship with a customer is enormous. Imagine, for example, a user with a learning disability or someone who simply struggles with specific mathematical content (e.g., the concept of compound interest rates). Not only might a relationship-focused bot customize its answers to a form that proves effective over time with the user, such as the terminology used, but it could also help diagnose that the user is experiencing difficulties in the first place. The bot will “remember” what the user struggles with and thereby can adapt its interactions. A long-term relationship might also make the user more comfortable interacting with the bot. Though such increased comfort comes at the risk of “humanizing” the chatbot and entering an unhealthy relationship (a phenomenon that has been reported especially among teenagers, Roose 2024), there exists enormous potential for customization in utilizing the history of prior interactions. Technical considerations. It is simple not to store any user history (“stateless”), as only the current question needs to be processed by the bot. No additional resources for storage or authentication are needed. For long-term relationships, the challenge lies with storing past interactions and providing relevant pieces to the chatbot, so it remembers key facts. Due to context window limits (the amount of text an LLM can remember), it is often not feasible to provide the entire conversation history in the prompt to the LLM. Then, a second step is necessary that extracts meaningful insights from previous conversations and summarizes them to provide context for future conversations. In addition, certain customer actions or information from other sources might also be provided as part of the context to the LLM to improve the long- time relationship. Extensive testing is necessary to establish how much information can be provided before performance deteriorates, such as the LLM starting to forget previous facts or responses becoming slower and more expensive since more text needs to be processed with each query.ii This is especially important for cases where the customer might expect that all previous conversation history is considered for future requests; clear communication is necessary in cases where this is not possible. Dimension 3: Proactive vs. Responsive Does the user reach out to the chatbot or the chatbot to the user? Most chatbots are responsive, i.e., they wait until the user takes the initiative and approaches the chatbot with a request. However, there is no reason why the chatbot shouldn’t be proactively taking the initiative (“Hi Joe, your upcoming flight to Paris is in 3 days. It looks like it will be rainy for the first few days. Do you want to learn more about must-visit indoor spots in the city?”). Such proactive bots might increase engagement by offering timely information or reminders, something that is well-studied in the medical domain (Volpp et al. 2017, Lekwijit et al. 2024). Currently, most chatbots do not reach out proactively by themselves. There are several reasons for this. In general, chatbots mimic the workflow of a customer support center or a help desk, which is by nature responsive. In addition, deploying a responsive chatbot is very simple on a technical level, as that is precisely what the chatbots are designed to do. Some recent examples of companies moving toward proactive chatbots include platforms such as Character.ai that have experimented with chatbots sending messages to users who have not engaged with