Our summary
A customer asking about opening hours may need a saved answer, while an order question needs personal records. David Omotayo describes adding Jev to a hobby WhatsApp chatbot to help distinguish these needs before paying another AI model to write a reply.
The chatbot asks three separate questions: what kind of message arrived, whether it needs a newly written answer, and whether earlier conversation matters. Its own code chooses saved replies and handles failed requests. For an order question, the application must retrieve order records; Jev does not fetch them.
A wrong category can send a customer an irrelevant reply, even when the model's confidence is high. The example uses untested cutoffs and shows only part of the application. Teams considering this pattern should compare mistakes, fallback use, total costs and response times with their existing chatbot.
Key takeaways
- Ask about the message category, the need for writing and the need for conversation history separately.
- Give account-specific requests a route to the relevant records, whether or not AI writes the reply.
- Test the whole reply workflow with labeled messages; fewer writing calls alone do not establish useful savings.