Answer WhatsApp customer questions and escalate to humans
A 24/7 WhatsApp support agent that answers customer questions from a company knowledge base — and hands the conversation to a human when it cannot answer confidently. Task: A 24/7 WhatsApp support agent that answers customer questions from a company knowledge base — and hands the conversation to a human when it cannot answer confidently. Built with n8n for orchestration and OpenAI for the reasoning layer. Self-directed project, built July 2026. This repository documents the system design; see Repository contents below. The problem Most customer questions are the same handful of questions. Answering them manually costs a team its whole day, and questions that arrive at night wait until morning. But a support bot that answers everything confidently is worse than no bot — it will invent an answer on the cases that actually matter. How it works Design decisions Escalation is the feature, not the fallback. The interesting design question is not "can the model answer?" but "does it know when it cannot?" Unresolved cases route to a human with the conversation attached, so the handoff does not restart the customer from zero. Answers are grounded in the knowledge base. Replies are generated against company content rather than the model's open-ended knowledge. A support bot that improvises policy is a liability. Escalation goes to Telegram, not another WhatsApp thread. The operator channel is kept separate from the customer channel, so the team has one queue to watch. Stack Design goals These are the outcomes the system was built to achieve: Cover the majority of routine, repeated questions automatically Provide round-the-clock coverage without staffing gaps Keep a human in the loop for every case the model cannot handle reliably Repository contents This README documents the architecture and the reasoning behind it. The exported n8n workflow JSON is being added. Ziad Hady — AI Automation & Integration Specialist Portfolio: https://ziad-portfolio-omega.vercel.app/; tools: n8n. Human role: The original post does not say how a person stayed involved.; limitations: Prerequisites include access to WhatsApp Business API, Telegram Bot API, OpenAI keys, and a structured company knowledge base. The source is a self-directed project design, not a deployed production system, so integration costs and latency are unknown. Human approval is required for the escalation logic to ensure the confidence threshold is appropriate. The exported n8n workflow is incomplete, requiring manual construction of the routing logic..
Tools
Published on Agent Work Atlas:
Overview
A 24/7 WhatsApp support agent that answers customer questions from a company knowledge base — and hands the conversation to a human when it cannot answer confidently.
Workflow
A 24/7 WhatsApp support agent that answers customer questions from a company knowledge base — and hands the conversation to a human when it cannot answer confidently. Built with n8n for orchestration and OpenAI for the reasoning layer. Self-directed project, built July 2026. This repository documents the system design; see Repository contents below. The problem Most customer questions are the same handful of questions. Answering them manually costs a team its whole day, and questions that arrive at night wait until morning. But a support bot that answers everything confidently is worse than no bot — it will invent an answer on the cases that actually matter. How it works Design decisions Escalation is the feature, not the fallback. The interesting design question is not "can the model answer?" but "does it know when it cannot?" Unresolved cases route to a human with the conversation attached, so the handoff does not restart the customer from zero. Answers are grounded in the knowledge base. Replies are generated against company content rather than the model's open-ended knowledge. A support bot that improvises policy is a liability. Escalation goes to Telegram, not another WhatsApp thread. The operator channel is kept separate from the customer channel, so the team has one queue to watch. Stack Design goals These are the outcomes the system was built to achieve: Cover the majority of routine, repeated questions automatically Provide round-the-clock coverage without staffing gaps Keep a human in the loop for every case the model cannot handle reliably Repository contents This README documents the architecture and the reasoning behind it. The exported n8n workflow JSON is being added. Ziad Hady — AI Automation & Integration Specialist Portfolio: https://ziad-portfolio-omega.vercel.app/
Reported result
No measured results are available as the project is a design documentation with the workflow JSON still being added. The stated goals are to automate routine questions, provide round-the-clock coverage, and ensure human intervention for unreliable cases. A measurable acceptance check would be to verify that 100% of low-confidence queries successfully route to Telegram with context, and that high-confidence replies are strictly grounded in the provided knowledge base.
Human role
The original post does not say how a person stayed involved.
Limitations
Prerequisites include access to WhatsApp Business API, Telegram Bot API, OpenAI keys, and a structured company knowledge base. The source is a self-directed project design, not a deployed production system, so integration costs and latency are unknown. Human approval is required for the escalation logic to ensure the confidence threshold is appropriate. The exported n8n workflow is incomplete, requiring manual construction of the routing logic.
Try it
I need to build a WhatsApp support agent using n8n and OpenAI. My company knowledge base is located at [replaceable input: URL or file path]. I want to use WhatsApp for customer input and Telegram for human escalation. First, create a sample n8n workflow that takes a test message, retrieves relevant context from the knowledge base, and uses OpenAI to generate a response. Implement a confidence check step. If confidence is low, send the conversation to my Telegram chat [replaceable input: Telegram Chat ID]. If high, simulate sending the reply to WhatsApp. Run this with 5 sample queries: 3 routine and 2 complex. Show me the logs for each step. Do not connect to live APIs yet. Confirm the workflow structure and confidence logic with me before proceeding to integration.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.