Report
https://www.dronahq.com/ai-agents-vs-chatbots-ecommerce/
In ecommerce, chatbots handle FAQs and order status well enough. The limit shows up when the workflow needs action, like refunds, stock checks, delivery exceptions, or offer logic. Curious where others first saw chat stop being enough.
The limit usually shows up when the intent is compound. "Handle refunds" is actually 3-4 separate decisions (eligibility, amount, method, exceptions) that a flat chatbot prompt squashes into one undifferentiated instruction.
One signal I use: if I can't write the full behavior in a single objective block, it probably needs an agent with separate prompts per sub-task. I built flompt to structure those individual prompts, 12 typed blocks (objective, constraints, chain_of_thought, etc.) that make each decision point explicit before you start building. github.com/Nyrok/flompt