UI Trax: AI Customer Agent with Guardrails

The challenge Businesses want AI that answers customers instantly, but most chatbots either sound robotic or confidently make things up: wrong prices, fake policies, promises nobody keeps.

Our first version was rules-first. It passed thousands of tests, then broke on real messages: typos, slang, mixed languages and several questions in one line. You can’t write a rule for every sentence a customer types.

The approach: AI-first, with guardrails We rebuilt it the way serious support platforms are built. The LLM understands every message, but anything business-specific comes only from tools: product search, order lookup, staff handoff and customer memory. Facts, permissions and side effects stay in deterministic code.

Product answers from the store’s own catalog The team manages categories, prices, sizes, colours and links in the dashboard. The agent answers only what was asked and shares the saved link to see pictures. It never invents a link.

Order details only after verification The agent asks for the phone number or email used for the order. A mismatch gets a polite retry; two different wrong attempts and a human takes over. This is enforced inside the tool, so no clever message or prompt injection can get around it.

A guard on every reply Every number, link and email in a reply must trace back to store data, a tool result or the conversation. Invented discounts, quality claims, promises and order status are blocked; after two failures a deterministic engine answers instead.

The most useful bugs came from live model runs. Each one became a guard rule and a regression test.

The agent remembers the customer’s name, handles small talk naturally, and says “I don’t have that” instead of guessing about shipping, live news or anything the store hasn’t approved.

Damaged items, refunds, exchanges and complaints go to the right person with an AI summary. While the customer waits, every message is acknowledged; the moment a team member replies, the bot goes silent.

Products, orders, alerts and the staff inbox in one place. The agent reads products and orders instantly, and Shopify orders are read directly when connected.

• 580+ unit tests and 1,500+ scripted dialogue scenarios • 160+ simulated customer journeys: typos, anger, bargaining, prompt injection • Every reply checked before it’s sent, with a deterministic fallback • Typical AI reply time of 1–3 seconds in live runs