Designing Human-in-the-Loop AI: Why an AI-Generated Reply Should Be a Draft, Not an Action

In a recent technical deep dive, developer Ramji Tripathi argues that integrating AI into customer support workflows requires a strict architectural separation between generation and execution. Rather than allowing an LLM to automatically commit replies to a conversation, developers should treat AI output as a candidate draft. By implementing a 'Human-in-the-Loop' design, the system forces a clear boundary: the model provides a suggestion, a human reviews and potentially edits it in a composer, and a separate, authorized operation commits the message. This approach prevents accidental side effects, maintains clear ownership of the communication, and ensures that authorization checks remain robust. Tripathi details his implementation using React and Express, emphasizing that keeping generation and sending as distinct operations improves system reliability, simplifies error handling, and ensures that human oversight is a structural component of the application rather than just a UI suggestion.