Concept projectA self-initiated brief, not client work. Product names are fictional.
Dispatch · AI support agent
A support desk that never sleeps.
An AI agent for online shops that answers order questions from real order data and hands anything unusual to a person.
The challenge
Online shops answer “where is my order?” and “can I change my address?” hundreds of times a week. The answers are already in their systems, but a person still has to look them up and type them out.
Goals
- Designed to answer routine order questions from live order and shipping data, at any hour.
- Designed to hand unusual cases to a person with a summary, so nobody asks the customer twice.
- Built so the team can see and adjust every action the agent takes.
Approach.
- 01
Read, then act
The agent looks up the order, the carrier and the policy before it answers. It never guesses an order status.
- 02
Clear limits
It can change an address before dispatch or resend a tracking link. Refunds over a set amount always go to a person.
- 03
A visible log
Every conversation shows the tools the agent used and why, so the team can review and correct it.
What we built.
An agent connected to the store, the help desk and the carrier, an inbox where people pick up escalations with full context, and a rules screen to set what the agent may do.
Hi, where is my order? It was supposed to come yesterday.
Hi Sam — your parcel is out for delivery today and should arrive by 6 PM. It was held one day at the depot because of the storm. Here’s live tracking: track.parcel/4821
Perfect, thank you!
Conversation closed by the agent · 1 min 40 s
Order
Timeline
- Ordered · Mon
- Shipped · Tue
- Held at depot · Wed
- Out for delivery · Today
Customer
Test a rule
I want a refund for my $240 lamp, it arrived broken.
Hands to a person with a summary, the photo and a suggested reply.
Evaluation suite
148 test conversationsAll passing after this changeDispatch
Priya R. · Order #4794
Refund for a damaged lamp — $240
Agent summary
- Photo shows a cracked base; delivered 2 days ago.
- Order is within the 30-day returns window.
- Refund is above the $100 limit, so it needs your approval.
Suggested reply
“Sorry about the lamp, Priya. I’ve refunded $240 to your card — no need to send it back.”
Concept screens, designed and built in HTML and CSS. All names, figures and data shown are fictional.
Design decisions.
Escalation is a feature
Handing over is designed as carefully as answering: the person who picks it up sees a summary, the order and the suggested next step.
Terminal green, used sparingly
A dark, quiet interface with one green for “resolved” keeps attention on the cases that need it.
Rules in plain language
Limits are written as sentences (“Ask a person before refunding more than $100”), not configuration.
How it would ship.
The plan we would run for a project like this, with a demo every week.
- Week 1Ticket review, pick the top question types
- Weeks 2–3Agent prototype on past tickets
- Weeks 3–5Integrations, rules, escalation inbox
- Week 6Supervised pilot, evaluation suite
Stack
- TypeScript
- Claude API
- Model Context Protocol
- PostgreSQL
- Shopify API
- Help desk API
- Evaluation suite
Deliverables
- Workflow audit
- AI agent with tool access
- Escalation inbox
- Plain-language rules
- Action log and review
- Evaluation suite