Chatbots fail when teams write happy-path dialogs first and call it AI. watsonx Assistant forced us to start from intents shoppers actually use: order status, returns, sizing, and escalation to a human.
Entities and synonyms mattered more than clever copy. The same question arrives as “where is my package”, “track order”, or “delivery late” — the assistant had to route all of them reliably.
Fallback and escalation paths were not edge cases; they were core UX. A confident wrong answer is worse than a clear handoff.
The project aligned with IBM’s certificate track, but the durable skill was conversational systems design — something I can reuse outside watsonx on any NLU stack.