WhatsApp support agent

AI agents and chatbots
  • n8n
  • WhatsApp Business
  • OpenRouter
  • Pinecone
  • Gemini
  • Gmail
The n8n workflow behind WhatsApp support agent, as built
The flow as built in n8n. The diagram below explains each step.

Routine WhatsApp questions answered from the knowledge base in under a minute; refunds and complaints reach a person every time

Outcomes

  • Routine WhatsApp questions answered from the company's own documents in under a minute, in the same chat.
  • Refunds, complaints, cancellations and anyone asking for a person reach the support team every time, with the full history.
  • The team maintains one thing: the documents. Update the document, and the next answer is right.

Challenge

Support teams answer the same WhatsApp questions all day: opening hours, how a return works, where the invoice is. The urgent complaint sits behind twenty routine questions and waits longest.

Solution

  • We built an n8n workflow on the company's WhatsApp Business number that takes text messages only; voice notes and images are left to a person.
  • We gave the agent the last turns of each conversation and the company's documents in Pinecone with Gemini embeddings, the model through OpenRouter; when the documents do not cover a question, it says so instead of inventing an answer.
  • We routed every refund, complaint, cancellation and request for a person to the support team by e-mail, unanswered by the agent.
  • Built as a demo of the pattern; the same flow runs on website chat and e-mail.