Omnichannel service test harness

What happens when a member asks for help.

This page documents the service test implementation. Voice is one channel, not the architecture. The harness validates voice/IVR, chat, secure messages, portal events, email/SMS handoffs, broker and group inquiries, agent workflows, CRM writeback, QA capture, and fail-closed handling for Contact Center AI.

Review contact-center blueprint Review evidence controls

Current scenario set Updating...

Scenario

Member asks why a claim is still pending and what documentation is needed. The same test can enter by voice, chat, secure message, portal workflow, or agent-created task.

System mode

Sandbox test harness with staged writeback. The agent must approve every customer-facing answer.

Contact center update

Spanish-language call flows are complete and ready for deployment. Claims call center workflows are in progress.

Approval gate

Deployment movement is currently on hold pending Vinod approval.

AI role

Listen, classify, retrieve, cite sources, draft response, and stage CRM disposition. No autonomous final action.

Fail state

Suppress suggestion and route to human-only service when identity, confidence, policy, or source checks fail.

Diagram depth levels

Service testing views from service execution path to fail-closed control logic.

These views let service leaders, engineers, QA, and compliance inspect the same service interaction at different levels of depth: operating flow, system architecture, runtime sequence, and exception handling.

L0 service flow

Customer service execution path

flowchart LR
  C["Member inquiry"] --> CH["Channel identified"]
  CH --> I["Identity and consent"]
  I --> T["Intent captured"]
  T --> R["Policy and claim context retrieved"]
  R --> A["Agent receives suggested answer"]
  A --> H{"Agent approves?"}
  H -->|Yes| CRM["CRM note and task written"]
  H -->|No| ESC["Manual service or specialist escalation"]
  CRM --> QA["QA evidence archive"]
  ESC --> QA
  QA --> BK["Knowledge gap and coaching backlog"]

L1 system

Omnichannel service test system context

flowchart TB
  subgraph Channels["Channels"]
    VOICE["Voice / IVR"]
    CHAT["Web chat"]
    SECMSG["Secure portal message"]
    EMAIL["Email / SMS handoff"]
    PORTAL["ServiceLink, broker/agent, GroupLink event"]
    DESK["Agent desktop"]
  end
  subgraph Control["AI controls"]
    STREAM["Transcript and message normalization"]
    INTENT["Intent classifier"]
    POLICY["Privacy and action policy"]
    ORCH["Conversation orchestrator"]
  end
  subgraph Knowledge["Grounding"]
    RAG["RAG knowledge base"]
    API["Core service APIs"]
    CRM["CRM case platform"]
  end
  subgraph Evidence["Evidence and QA"]
    TRACE["Trace log"]
    QA["QA scoring"]
    BACKLOG["Defect and knowledge backlog"]
  end
  VOICE --> DESK
  CHAT --> DESK
  SECMSG --> DESK
  EMAIL --> DESK
  PORTAL --> DESK
  DESK --> STREAM --> INTENT --> ORCH
  POLICY --> ORCH
  ORCH --> RAG
  ORCH --> API
  ORCH --> DESK
  DESK --> CRM
  RAG --> TRACE
  API --> TRACE
  CRM --> TRACE
  TRACE --> QA --> BACKLOG

L2 sequence

Runtime message flow

sequenceDiagram
  participant M as Member
  participant CH as Channel adapter
  participant ID as Identity service
  participant AI as AI orchestrator
  participant RAG as Knowledge RAG
  participant API as Claim API
  participant AG as Agent or workflow
  participant CRM as CRM and QA
  M->>CH: Claim-status inquiry via voice, chat, portal, or message
  CH->>ID: Verify identity, consent, and channel policy
  ID-->>AI: Verification and allowed data scope
  CH->>AI: Normalized transcript/message and intent context
  AI->>RAG: Retrieve policy and script sources
  AI->>API: Query claim status in sandbox
  API-->>AI: Pending document reason
  AI-->>AG: Draft answer, citations, CRM disposition
  AG->>CRM: Approve edited note and task
  CRM-->>AI: Evidence record and QA flags

L3 decision

Fail-closed service control

flowchart TD
  START["AI suggestion candidate"] --> ID{"Identity verified?"}
  ID -->|No| HUMAN["Human-only service script"]
  ID -->|Yes| SRC{"Approved source found?"}
  SRC -->|No| SUPPRESS["Suppress answer and create knowledge gap"]
  SRC -->|Yes| PII{"PII policy pass?"}
  PII -->|No| MASK["Mask redact or remove restricted fields"]
  PII -->|Yes| CONF{"Confidence above threshold?"}
  CONF -->|No| SUPPRESS
  CONF -->|Yes| APPROVE["Show to agent for approval"]
  APPROVE --> WRITE{"Agent approves CRM writeback?"}
  WRITE -->|No| HUMAN
  WRITE -->|Yes| DONE["Write CRM note and archive evidence"]
  HUMAN --> ARCHIVE["Archive reason and replay"]
  SUPPRESS --> ARCHIVE
  MASK --> APPROVE

Deep service flow

Omnichannel sequence from channel intake to QA evidence.

This test flow shows each system boundary, approval point, writeback path, and evidence capture point for a single member service interaction. A voice call is one replay scenario inside this broader harness.

Runtime service flow

Channel intake to governed service action.

The customer experience remains human accountable. The AI layer provides bounded assistance and generates traceable evidence for engineering, QA, and compliance.

Omnichannel service path

Service path under validation

Correlation ID follows the event through channel adapter, identity, transcript/message normalization, policy engine, RAG, core service APIs, agent desktop or workflow, CRM, and QA archive.

Channel ingress

Member inquiry Voice/IVR, chat, secure message, portal event, email/SMS handoff, language, and transfer context.

Intent routing Claim status, payment, enrollment, document, grievance, or unsupported path.

Identity boundary

Member verification Authentication, consent, plan context, and allowed-field policy.

PII/PHI filter Mask restricted data before it enters prompts, logs, or CRM summaries.

AI assist

Transcript and intent ASR transcript, sentiment, entity extraction, and service-policy selection.

Grounded response Retrieval-augmented answer with citations and confidence gate.

Human control

Agent approval Agent accepts, edits, rejects, or escalates the suggested response.

CRM staging Disposition, summary, task, next action, and call outcome are staged.

Evidence

QA archive Transcript, prompt hash, model hash, sources, tool calls, and agent edits.

Release signal Scenario result feeds readiness, defects, knowledge gaps, and control review.

Scenario replay

Traceable test transcript.

Each replay captures the customer request, retrieved source data, agent approval, system action, and evidence record.

Member

"I am checking on a claim that still says pending. Can you tell me what is missing?"

Intent candidate: claim status. Requires identity validation before detail retrieval.

System

Identity and consent checks pass in the sandbox profile. PII policy masks restricted identifiers before the model context is assembled.

Evidence: auth result, consent flag, data policy decision, correlation ID.

AI assist

Retrieves claim-status policy, pending-document rule, and sandbox claim record. Drafts a response with source citations and a follow-up task.

Evidence: source IDs, prompt version, model version, retrieval score, tool-call log.

Agent

Edits the response for tone, confirms the next action with the customer, and approves the CRM note.

Evidence: human approval, edit reason, final response, staged CRM payload.

Archives transcript and evaluates policy adherence, citation accuracy, CRM completeness, and knowledge-base gaps.

Evidence: QA flags, replay link, defect ticket if needed, release-readiness record.

Fail-closed controls

Unsafe conditions do not produce customer-facing automation.

These are the controls that make the service test credible to operations, security, and engineering.

Condition	System action	Evidence captured	Owner
Identity incomplete	No member details retrieved; agent follows manual verification script.	Auth failure code, blocked retrieval event, replay reference.	Service operations
Low confidence or source mismatch	Suggestion suppressed; knowledge gap routed to backlog.	Retrieval results, confidence decision, missing-source ticket.	Knowledge owner
PII/PHI policy hit	Field masked, redacted, or removed from prompt and CRM summary.	Policy decision, redacted fields, model-context record.	Privacy and security
CRM writeback exception	Manual update task created; final write requires agent confirmation.	Failed payload, retry state, manual task, agent signoff.	Platform engineering