You are adding LLM capabilities to an existing enterprise application that already has business workflows, APIs, permissions, and human support paths. The goal is to improve tasks like answering user questions, drafting responses, and taking limited actions, without breaking reliability or trust. Some requests can be handled with grounded retrieval, while others may need tool use against internal systems.
What experience do you have with integrating large language models into existing systems, and how would you approach that work in practice?