A useful chatbot needs a clear audience, supported tasks, authoritative data sources, refusal boundaries, and a route to human help. Test ambiguous requests, prompt injection, unsupported claims, sensitive data, multilingual input, and failure recovery. Retrieval responses should cite the underlying source and respect permissions. Minimize conversation retention and never place secrets in prompts or logs. Measure task completion and harmful error rates, not only fluency. Tell users when they are interacting with AI.
Akousa