CUSTOMER SERVICE / MULTI-AGENT WORKBENCH
A clear path from request
to verified action.
Three native service agents. Four specialized native judges. Immutable fictional cases. Every simulated action backed by a tool receipt, not an AI promise.
CONTROLLED ACCESS
Connect your session
Interim sign-in: paste a short-lived, delegated Entra access token for this API. Only explicitly allowed users can connect.
HTTPS required outside localhost. Token remains in tab memory, never storage or logs. No real customer data. Disconnect or close the tab to clear it. Never paste tokens into shared screens or chat.
01 / CHOOSE A JOURNEY
Realistic cases. Zero real transactions.
Connect to load the four synthetic scenarios.
02 / VERIFY THE OUTCOME
Simulation evidence
Choose a case and run a simulation. Calls are sequential and paced for shared quota. A run may take several minutes.
Judge calibration not established. Scores are provisional, not release-ready. Ordinary runs use one prompt per specialized criterion, on the same regional model.
Specialized judge rubrics / 10
Authoritative tool readback
Recorded simulation action only. Not final customer resolution and never a real payment.
Agent explanation
Trace, criteria and evaluation scope
03 / MEASURE & DRILL DOWN
System → journey → agent
Demo targets: 90% task completion, 8.5/10 goal achievement, 9/10 coherence. Not production SLOs.
Current replica-local snapshot only. State can disappear on restart or scale-to-zero. Content-free lifecycle metrics are emitted to existing Log Analytics; this page does not query historical logs.
Business impact, resolution rate and reopen rate: unknown until actual user feedback. Synthetic scores are not observed customer outcomes.
Connect to see your session's metrics.
Tier 2 / All journeys
Tier 3 / Agent metrics · all journeys
Completion denominator includes all attempts, including failed and pending. Goal/coherence means show scored denominator. Agent accuracy uses explicit task correctness labels, never 1 − error rate. Unknown values shown as unavailable. p95 uses nearest rank. Latency includes quota pacing.