The button below starts a genuine Amazon Connect chat. The flow hands a verified identity to a dispatcher Lambda, answers you in under a second, and a Claude agent on Bedrock AgentCore keeps working after the chat ends — writing its findings to a ticket in a sandboxed demo desk. The right-hand panel narrates what is actually happening.
What the caller sees
Pick a task and press start — this opens a genuine Connect chat as demo-user, a synthetic identity in the demo tenant.
What is actually happening
The glass box. Every step of the real path gets narrated here as it happens — no payloads, no identifiers, just what the system is doing and why it is built that way.
The same conversation a WhatsApp number carries: Amazon Connect chat in front, a Claude agent on Bedrock AgentCore behind it, answering only from an operator-edited answer set. Ask a question sideways and it still finds the approved answer; ask something off-list and it hands you to a human rather than improvising — on a public channel, an invented answer about payments or legal duties is worse than none.
The answer set here is a fictional licensing service, seeded as demo data. Point the same flow at a WhatsApp Business number and this exact conversation happens in WhatsApp.
Demo Licensing Support — WhatsApp-style Q&A
Ask it like a customer would — try “my gran is 80 and gets pension credit, does she have to pay?” Answers come only from an operator-approved set; anything off-list is handed to a human, never improvised.
The interesting part isn't that it works — it's what it refuses. These rails are enforced server-side; the page cannot loosen them.
The browser sends an identity key. The email address it maps to lives in a server-side directory the page cannot edit — so nobody can type someone else's address and borrow their authority.
You pick a task id from a fixed list. There is no free-text way to talk the agent into arbitrary work — the flow chooses what runs, not the visitor.
Every action the agent can reach re-checks its own authority at execution time, in the action itself. The demo desk is a sandbox with synthetic identities — not a customer environment.
Each run costs real model spend (roughly $0.03 per gated scenario), so the endpoint is rate-limited per visitor and per day. If it says it's busy, that's the limiter doing its job.
The same architecture on the voice path adds deterministic OTP verification before the flow ever runs — verification is not authorisation, and the platform treats them as different jobs.