AI customer support agent
An agent that answers customers from the business's real data — and knows when not to answer.
01The problem
In customer support, most contacts are the same few questions: where's my order, how do I return this, can I cancel. I handled them for over two years at Amazon, for the French market.
A typical chatbot handles them badly. Either it's too rigid and points to an FAQ, or — with a poorly constrained language model — it invents a plausible answer: a fake tracking number, a return policy that doesn't exist. In support, a made-up answer costs more than no answer.
02The solution
A conversational agent that answers only from the business's own data. It looks up the order live, searches the store's policies (returns, cancellations, shipping), and keeps context from one question to the next.
When a request shouldn't be handled by AI alone — a refund, a cancellation, a complaint — it promises nothing: it logs the escalation for the human team and tells the customer clearly.
03How it's built
- n8n Chat Trigger (self-hosted): the entry point of every conversation.
- AI Agent with tool-calling (OpenAI): the model decides whether to call a tool, and which one.
- Agentic RAG: searching the policies is a tool the agent chooses to use, not a step forced on every message. A “hello” or an order number doesn't trigger a pointless document search.
- Same embedding model for ingestion and retrieval: documents and questions are embedded by the same model — otherwise the vectors aren't comparable and retrieval returns noise.
- Order lookup tool: reads Google Sheets live, by order number.
- Escalation tool: writes the request (type, order, summary) to a dedicated log that the human team works from.
- Per-session memory: follow-up questions (“so when will it arrive?”) keep their context.
- Anti-fabrication framing: the agent answers only from its tools' results; if the information is missing, it says so.
04The result — what it proves
Tested end to end across three kinds of scenario — order lookup, policy question and escalation:
| Scenario | Expected behaviour | Result |
|---|---|---|
| Order lookup | Gives the order's real status from the data; if the order doesn't exist, says so instead of inventing one. | ✓ Tested |
| Policy question (RAG) | Answers from the passages retrieved from the store's policies; if the answer isn't there, says so. | ✓ Tested |
| Refund, cancellation, complaint | Doesn't act alone: writes the escalation to the log, then tells the customer. | ✓ Log checked |
These tests were run by me on a demo store, with fictional orders — this is not an external audit.
What it proves: a useful support agent isn't the one that answers everything — it's the one that knows when not to, and leaves a trail when it hands off.
