Resolve support tickets, without letting AI touch the money.
AI agents resolve routine support tickets end to end, including refunds. AI proposes, code decides: no model output can move money unless the Policy Guard allows it, and anything over $100 waits for a person.
Plain code approves every refund$100+ waits for a personEvery action audited
AI proposes, code decidesRefunds over $100 wait for a personDelivered orders only30-day refund windowNo double refundsEvery action auditedAI proposes, code decidesRefunds over $100 wait for a personDelivered orders only30-day refund windowNo double refundsEvery action audited
✓ Your results
Inbox · 11 sample tickets
Approval queue · refunds over the limit wait for a person
Order book · sample data; today is 2026-10-07
Order
Customer
Item
Total
Refunded
Status
Delivered
Audit log · every action attempted, executed, refused, queued or approved
#
Ticket
Event
Action
Result
Evals · the project's test cases, run here against the release gates
Runs on the example engine, so it checks the guardrails and the test harness, not how good a live model is. The same cases run in Python with python evals/run_evals.py.