An AI assistant for a small online shop's customer service

Details anonymized and adapted for confidentiality.

The problem

An online shop run by two people. Around 40 messages a day, and the vast majority were about the same three things: “where is my package”, “how do I return an item”, “did my order ship”. The kind of questions you answer by pulling up the order number and the carrier tracking.

The result: a shared inbox, replies that sometimes went out 48 hours later, customers following up, and the real requests (a dispute, a broken product, a pre-sale question) buried in the noise.

What we built

A deliberately small assistant, wired into the support inbox and the e-commerce platform.

Front-door triage - Every message gets sorted: a repetitive case the assistant can handle, or a real case that needs the human. No grey zone, when in doubt it goes to the human.

Answers from real data - For package tracking, returns and order status, the assistant reads the actual order and the actual tracking, then replies with the exact details. It never invents a tracking number.

Clean handoff - When it’s a real case, the assistant doesn’t improvise. It passes the thread to the owner with a short summary of the situation, without having promised the customer anything.

Guardrails - Refunds, disputes and anything touching money are always escalated to a human. The AI never decides those on its own.

The result

• Around 65% of messages handled with no human intervention, mostly tracking and returns.
• First reply in minutes instead of hours, around the clock.
• Real cases surface clearly, no longer drowned in the flow.
• Two weeks to set up, not a six-month project.

The stack

• LLM pipeline (Claude API) for triage and drafting replies
• Direct connection to the e-commerce platform and carrier tracking
• Strict human-handoff rules on anything money-related
• Lightweight deployment on Cloudflare Workers, no infra to maintain
• A log of every reply, to keep an eye on it and tune as you go