Support

AI WhatsApp Customer Support Bot

Enabled 24/7 customer support with AI WhatsApp bot, saving 10 hours/week with Hebrew language support.

24/7Support coverage with 10hrs/week saved

The Challenge

An e-commerce brand with 500+ daily WhatsApp inquiries had a 3-person support team that could only cover 9 AM - 6 PM Sunday through Thursday (standard Israeli business hours). After-hours messages — roughly 35% of total volume — went unanswered until the next morning, by which point many customers had already purchased from competitors. During business hours, response times averaged 15-30 minutes during peak periods, and 60% of inquiries were repetitive questions about shipping times, product availability, and order status that didn't require human judgment. The team had previously tried a simple keyword-based chatbot, but it created a worse problem: bot-agent conflicts where both the automation and a human agent would respond to the same message, confusing customers and undermining trust.

Our Solution

We built an AI-powered WhatsApp support bot using n8n as the orchestration layer and GPT-4o-mini as the conversational engine, with native Hebrew language support including slang and informal writing styles common in Israeli WhatsApp conversations. The bot pulls real-time product information from a Google Sheets catalog (updated by the client's team), checks order status live via WooCommerce REST API, and answers shipping/return policy questions from a structured knowledge base. To solve the bot-agent conflict problem, we implemented a 24-hour human-handoff mechanism: when a customer asks for a human or the bot detects frustration/complexity beyond its scope, it seamlessly transfers the conversation to a human agent and pauses all AI responses for 24 hours on that thread. A 10-message escalation limit ensures that if a customer isn't getting the help they need within 10 exchanges, they're automatically transferred to a human. The bot identifies itself as an AI assistant in its first message to set expectations transparently.

Detailed Approach

The biggest challenge was making GPT-4o-mini work naturally in Hebrew WhatsApp conversations, which are informal, abbreviation-heavy, and often include transliterated English words. We spent the first week building and testing the system prompt through 100+ simulated conversations, refining the tone to be helpful and friendly without being robotic or overly formal. The prompt includes specific instructions for handling common Hebrew chat patterns like "מה המצב" (what's up) as a greeting vs. a status inquiry.

The n8n workflow architecture uses a webhook trigger from the WhatsApp Business API, passes through a conversation state manager (stored in Google Sheets keyed by phone number), and routes to either the AI engine or a "paused — human handling" bypass. The state manager tracks: conversation history (last 10 messages for context), handoff status, message count in current session, and the customer's identified intent. Before calling GPT-4o-mini, the workflow enriches the prompt with relevant context: if the customer mentioned an order number, it pre-fetches the order status from WooCommerce; if they asked about a product, it pulls current availability from the catalog sheet.

The human handoff mechanism monitors for explicit transfer requests ("I want to talk to a person"), frustration signals (multiple question marks, caps lock, negative sentiment), and the 10-message limit. When triggered, it sends the human agent a summary of the conversation so far — including the customer's intent, any order numbers mentioned, and the bot's assessment of the issue — so the agent can pick up seamlessly without asking the customer to repeat themselves. The 24-hour pause timer resets if the human agent tags the conversation as resolved.

Key Results

🕐 Knowledge base creation and Hebrew prompt engineering took 1 week

✓24/7 support coverage — after-hours inquiries now get instant responses
✓10 hours/week of human agent time saved by automating repetitive inquiries
✓Average first-response time improved from 15-30 minutes to under 5 seconds
✓Bot successfully resolves 68% of inquiries without human escalation
✓Zero bot-agent conflicts since launch — the handoff mechanism works flawlessly
✓Customer satisfaction score maintained at 4.5+ despite AI handling majority of first contacts

Knowledge base creation and Hebrew prompt engineering took 1 week. The n8n workflow was built in 2 weeks with extensive testing across 100+ conversation scenarios. A 1-week monitored soft launch (bot active, all conversations reviewed by staff) preceded full deployment.