A Canadian WooCommerce store selling regulated products was running about 20 HelpScout tickets a day. Most of them were the same five questions: where is my order, how do I use this, can I take it with my medication, what do you ship, how do I return it. The support team was spending its hours on answers that were already written down somewhere.

I built a support bot that answers those questions from the store's own documentation, checks live order status in WooCommerce, and hands the conversation to a human in HelpScout when it cannot answer. Every question it fails on gets logged, so the knowledge base grows where customers actually push on it.

What it does

Answers only from approved content

Product and company questions are answered from a Pinecone vector knowledge base, never from the model's training data. The corpus is curated by hand from official documentation, FAQs and product descriptions, then embedded with OpenAI embeddings. For a store where a wrong dosing answer is a compliance problem, this constraint is the whole point. The system prompt treats the knowledge base lookup as mandatory, and the agent is instructed to escalate instead of guessing when retrieval comes back thin.

Looks up real orders

A customer can ask about their order inside the chat. The agent calls a WooCommerce sub-workflow that resolves the customer and returns a plain-language summary of their recent orders: status, dates, shipping address. Order data is only released to a logged-in customer whose email has been verified, so someone typing a stranger's address gets nothing.

Escalates with the full transcript

When the bot cannot answer with confidence, it offers to open a ticket. The HelpScout sub-workflow either finds the customer's existing conversation or creates a new one, and it attaches the complete chat transcript rather than a summary. The support agent sees exactly what was asked, what the bot said, and where it got stuck.

Logs what it could not answer

Every unanswered question lands in a Google Sheet with the exact wording, the bot's response, a timestamp and a topic label (Shipping, Product, Medical, Returns, General). The team reviews it weekly and turns the patterns into new knowledge base entries. That loop is what moved the numbers below.

Results

At launch, the bot resolved enough repetitive questions to cut HelpScout ticket volume by roughly 40%.

After several rounds of feeding the unanswered-question log back into the knowledge base, it was handling 18 to 19 of the roughly 20 daily tickets. One or two a day still reach a human, which is where a human belongs: the medical edge cases and the angry ones.

First-response time on order status and product questions dropped to seconds, and the business got something it never had before: a running list of what customers ask that the documentation does not cover.

The running cost is in the low double digits per month. The full breakdown is in What a RAG System Actually Costs.

Architecture

The system is three n8n workflows: the main agent, the WooCommerce order lookup, and the HelpScout ticket handler. The agent calls the other two as tools, so each one can be tested and updated on its own. Product catalog and FAQ content live in separate Pinecone indexes because they change at different rates and should never be confused with each other in retrieval.

The original build ran on GPT-4o. Answer generation has since moved to Claude; embeddings stay on OpenAI. The retrieval design did not change.

The technical walkthrough, including the prompt rules that force the retrieval call and the index separation, is in How I Built a WooCommerce RAG Chatbot with n8n.

Stack

Orchestration: n8n
Answer generation: Claude (originally GPT-4o)
Embeddings: OpenAI text-embedding-3-large
Vector database: Pinecone (two indexes: products, knowledge base)
Integrations: WooCommerce REST API, HelpScout API, Google Sheets

AI Customer Support Bot: RAG + WooCommerce Case Study