Try the flows this design scopes: cancel a booking, ask about baggage, change a date, or
trigger an escalation. Watch the intent classification, confirmation gate, and audit log
respond in real time.
What this actually is: a rule-based simulation running entirely in your
browser — keyword matching stands in for intent classification, and the "tool calls" return
fixed illustrative numbers, not a live booking system. A public static site can't safely hold
an API key to call a real LLM (anyone could copy it from page source), so this demonstrates
the architecture and guardrails from the
full case study (T5)
— the confirmation gate, the escalation handoff, the audit trail — working end to end, not a
claim that this is production-grade language understanding.
Support Copilot
Try a quick-start below, or type your own
🧑💼 Connected to a human agent — the copilot has handed off with full context.
Agent Action Log
Autonomous tool calls only
No autonomous actions yet — this fills in when the copilot actually executes something (a cancellation, a refund, a date change), not just when it answers a question.
What to try: cancel a flight to see the confirmation gate (it never
auto-executes); ask about baggage to see a fully autonomous answer with no gate needed; ask to
change your date to see a lighter single-step confirmation with an intentionally-flagged
audit entry; or mention a dispute to see the hard-escalation path that never runs
autonomously, regardless of how confident the model is.