AI customer-support agent
Customer messages needed fast replies without losing control of orders.
- Client
- Independent project
- Service
- AI automation & agents
- Status
- Live
01 / Build
What I built
Agent that answers customers, places orders as pending for staff review and hands off to a human with an issue ticket.
- Claude
- MCP
- Chrome extension
02 / Decisions
Key decisions
- Orders the agent places are saved as pending for staff review, never confirmed on their own, so a misread message can't send the wrong product.
- The agent works through MCP tools with narrow, named actions (find an order, place a pending order, open an issue) instead of open access to the store.
- Every hand-off to a human comes with an issue ticket, so staff pick up the conversation with its context instead of starting again.
03 / Notes
Architecture and hindsight
Architecture
Claude reads each customer message and answers it, or calls one of a few MCP tools when the message needs data or an action. Those tools are the only way the agent touches the business: they look up orders, create pending orders and open issue tickets.
What I'd do differently
Log every tool call with the message that triggered it from the first day. When an answer is wrong, the log shows in one lookup whether the model misread the message or a tool returned the wrong data.
04 / Outcome
Outcome
Live in production; the agent answers customers, and every order it places waits for staff to confirm.
Have a problem like this one?