The best AI agent knows when to stop
An AI agent that resolves routine support tickets end to end and asks a human before anything risky.
Run a ticket through the agent
Pick a ticket. Watch the agent triage, look things up and apply the rules. Risky ones stop and wait for you.
A small browser version with a deterministic fake model — no API key, nothing leaves your device. The real project is in the repo.
What it solves
Support teams answer the same few questions all day: where is my order, I want a refund, change my address. Each one means opening the order system, checking the policy, acting and writing a polite reply.
How it works
- The LLM triages the ticket into a validated structure and chooses which read-only tools to call (LangGraph).
- Plain Python code applies the refund and shipping rules — the model never decides money.
- Risky cases (large refund, VIP, angry refund request, unverified sender) pause for one-click Approve / Reject; a different sender or an angry complaint is escalated instead. Every step is logged.
The honest verdict
The model is good at reading messy text; code is good at “refund over $100 needs approval”. The split held: none of Qwen 3.7 Plus's 11 wrong decisions ran an action the labels forbid. But the rules only act on what the model understood. On the 40 labelled tickets, intent accuracy was 97.5% for Claude Haiku 5.5 and 67.5% for Qwen 3.7 Plus (below the 90% of the keyword baseline), and the judge (Claude Opus 5.5) passed only 70% of Haiku's replies: the failed ones invented details or claimed actions that were never taken. The two misses found on gpt-5.4-mini were fixed by editing the prompt on those same tickets and re-checked on 4 cases, not on the full set.
Next demoDocuChat →