BRD-02: Member & Provider Experience AI

Year 2 Workstream — This BRD defines the functional, non-functional, AI model, data, integration, and regulatory requirements for ACME Highland Health's conversational AI capability: a member-facing virtual assistant for self-service and a call-center agent-assist tool for real-time support during live calls. Developed through 7 JAD sessions during Year 2, with the same mandatory governance structure as BRD-01. All governance, testing, and model validation processes follow the Program Governance Model.

1. Executive Summary & Business Case

ACME Highland Health's member service center handles approximately 2.8 million calls annually across a 650-seat contact center operating Monday–Friday, 7 AM–8 PM ET. Average handle time (AHT) is 8.2 minutes. After-hours coverage is limited to a recorded IVR directing callers to emergency services or callback queues. Of the 2.8 million annual calls, approximately 55% (1.54 million) are routine inquiries — benefits verification, claims status checks, PA status lookups, provider directory searches, ID card replacements, and general plan information — that do not require clinical expertise or complex problem-solving. These calls consume approximately 210,000 agent-hours annually at an average cost of $38 per call (fully loaded: agent salary, facilities, telephony, supervision).

Simultaneously, member satisfaction surveys consistently identify three pain points: (1) long hold times during peak periods (average 6.4 minutes, peak 14+ minutes), (2) no after-hours service for routine inquiries, and (3) inconsistent information quality across agents (agents access 4+ systems during a call and sometimes provide conflicting or incomplete answers). Provider-facing inquiries follow similar patterns: 40% of provider calls are routine eligibility and claims-status checks that could be self-service.

BRD-02 delivers two AI-powered capabilities that address these problems without replacing human agents:

1. Member-Facing Virtual Assistant: A conversational AI available 24/7 across web, mobile, and voice channels that handles routine inquiries. The virtual assistant is not a chatbot in the traditional sense — it is a retrieval-augmented generation (RAG) system that answers questions by retrieving verified information from ACME's authoritative documents and generating natural-language responses grounded in that retrieved content. It does not guess, speculate, or generate information beyond what its source documents contain.
2. Call-Center Agent-Assist: A real-time AI tool that sits alongside human agents during live calls, providing response suggestions, policy lookups, member history summaries, and post-call case summarization. The agent-assist tool speaks to the agent, not to the member — the human agent always controls the conversation with the member.

1.1 Business Objectives

1.2 Investment & Timeline

Budget: BRD-02 delivery funded within SOW-02 ($41.58M Year 2, shared with BRD-03 and cross-cutting workstreams). BRD-02-specific delivery labor: approximately $8M of the Year 2 allocation.

Timeline: Year 2 phases follow the same 5-phase gate rhythm as Year 1. BRD-02 JAD sessions: Jan–Mar 2028. Requirements sign-off: Mar 2028. Build: Apr–Jul 2028. Pilot: Aug 2028. Production: Sep 2028.

Team: 22 people, led by X. Garcia (BRD-02 Lead, Pulaski). Includes 2 senior conversational AI engineers, 3 NLP engineers, 2 conversation designers, 2 data scientists, 1 integration engineer, 1 escalation protocol designer (ACME), 1 call-center ops liaison (ACME), and supporting UX research, QA, Scrum, DevOps, accessibility, and documentation roles.

2. Regulatory & Liability Context

2.1 The Air Canada Precedent & Hallucination Liability

2.2 HIPAA & PHI in Conversational Data

• Every conversation between a member and the virtual assistant is PHI — it contains member-identifiable health-related information (what they asked about, what the system retrieved from their claims/benefits records).
• Conversation logs are subject to the same HIPAA Privacy and Security Rule requirements as claims data: encrypted at rest and in transit, access-controlled, audit-logged, and retained per ACME's retention policy (minimum 7 years).
• If the virtual assistant is accessed via a third-party mobile app or voice platform, the platform vendor must have an executed BAA with ACME before any member data flows through the platform.

2.3 Accessibility Requirements

• The member-facing virtual assistant must meet WCAG 2.1 Level AA accessibility standards across web and mobile channels. This includes: keyboard navigability, screen reader compatibility, sufficient color contrast, text alternatives for non-text content, and captions for voice/video interactions.
• The voice channel must support speech recognition across major accent groups present in ACME's member population. Speech recognition accuracy must not vary by >5% across accent groups.
• Section 508 compliance required for any interface used by federal exchange members.

2.4 State Consumer Protection

• Several states require disclosure when a consumer is interacting with an AI system rather than a human. BRD-02's virtual assistant must clearly identify itself as an AI assistant at the beginning of every interaction: "Hi, I'm ACME's virtual assistant. I can help with benefits, claims, and general plan questions. If you'd prefer to speak with a person, just let me know."
• The virtual assistant must not impersonate a human or obscure its AI nature. No human names, no "I understand how you feel" empathy scripts that simulate human emotional response.

3. Current State Assessment

3.1 Current Member Service Channels

3.2 Current State Pain Points

• Hold times: Average 6.4 minutes during peak hours (10 AM–2 PM ET, Mon/Tue). Peak hold times exceed 14 minutes during open enrollment periods (Oct–Dec). Member abandonment rate: 8.3% during peak.
• Information inconsistency: Agents access 4+ systems during a call (core admin platform, claims system, PA system, provider directory) with no unified view. Different agents sometimes provide different answers to the same question depending on which system they consult first. Internal quality audit finds factual errors in approximately 6% of sampled calls.
• After-hours gap: Members calling outside business hours reach only a recorded message directing them to emergency services or the web portal. 18% of call attempts occur outside business hours. These members either call back during business hours (adding to peak volume) or do not call back (unresolved need).
• Agent burnout: Routine inquiries (benefits verification, claims status) consume 55% of agent time but require minimal expertise. Agents report that repetitive routine calls are the primary driver of job dissatisfaction. Annual agent turnover: 28%.
• Provider experience: Provider calls (eligibility checks, claims status, PA status) follow the same queue as member calls. Providers report frustration with hold times and prefer electronic self-service but find the current portal insufficient for their workflow needs.

4. Target State & Capability Description

4.1 Member-Facing Virtual Assistant

A conversational AI system available 24/7 across web, mobile, and voice channels. The virtual assistant handles routine inquiries conversationally — not through menus or keyword matching, but through natural-language understanding that determines what the member is asking and retrieves a grounded answer from ACME's authoritative data sources.

Core design principle: The virtual assistant answers questions by looking up verified information, not by generating information. It is a retrieval-first system — if the retrieval step does not find a relevant, verified answer, the system does not attempt to generate one. It escalates to a human agent instead. This architecture directly addresses the Moffatt liability risk.

Supported inquiry types (routine):

• Benefits verification ("Am I covered for physical therapy?")
• Claims status ("Where is my claim from Dr. Smith on June 15?")
• PA status ("Has my prior authorization for the MRI been approved?")
• Provider directory search ("Find an in-network dermatologist near me")
• ID card request ("I need a new insurance card")
• General plan information ("What's my deductible?" / "When does my plan renew?")
• Cost estimation ("How much will a specialist visit cost with my plan?")

Unsupported inquiry types (always escalate to human):

• Complaints and grievances
• Appeals of denied claims or PAs
• Complex billing disputes
• Clinical questions requiring medical judgment
• Fraud reporting
• Any inquiry where the member is emotionally distressed

4.2 Call-Center Agent-Assist Tool

A real-time AI tool displayed in a side panel alongside the agent's existing desktop interface during live calls. The agent-assist tool listens to the conversation (via real-time speech-to-text), understands the member's inquiry, and provides the agent with:

• Response suggestions: Recommended answers to the member's question, pulled from the same verified source corpus as the virtual assistant. The agent reads or paraphrases the suggestion — the tool does not speak to the member.
• Policy lookup: Relevant sections of the member's benefits summary, medical policy, or plan documents surfaced automatically based on the conversation topic.
• Member history summary: Concise summary of the member's recent interactions (calls, portal visits, claims, PAs) displayed at call start so the agent doesn't ask the member to repeat context.
• Real-time case summarization: As the call progresses, the tool generates a running summary of topics discussed, actions taken, and commitments made. At call close, the summary auto-populates the case management system, reducing after-call work (ACW) time.
• Sentiment monitoring: Real-time sentiment analysis flags when member frustration is rising, alerting the agent (and optionally a supervisor) to de-escalation opportunities.

5. Functional Requirements

Requirements classified using MoSCoW prioritization: Must (non-negotiable for production), Should (expected, can defer with justification), Could (included if capacity allows).

5.1 Virtual Assistant — Core Conversation

5.2 Virtual Assistant — Factual Grounding & Hallucination Prevention

5.3 Virtual Assistant — Escalation to Human Agent

5.4 Agent-Assist Tool

5.5 Audit, Reporting & Compliance

6. AI Model Specifications

6.1 Model Architecture

BRD-02 uses a multi-model architecture. Each model has a specific function, its own training/tuning approach, and its own validation criteria.

6.2 Hallucination Prevention Architecture (M-06 Detail)

M-06 is the highest-risk model in BRD-02. Its architecture is specifically designed to prevent the Moffatt scenario:

1. Retrieval step: When a member asks a question, the system first queries a vector database of verified ACME source documents (benefits summaries, plan documents, medical policies, provider directory, claims/PA records). The retrieval step returns the top-K most relevant passages with a relevance score.
2. Confidence gate: If no retrieved passage has a relevance score ≥ 0.90, the system bypasses generation entirely and escalates to a human agent. This is the primary hallucination-prevention mechanism: the model cannot generate a response when it has nothing verified to ground it in.
3. Constrained generation: If relevant passages are retrieved, the LLM generates a response constrained to use only information present in those passages. The system prompt explicitly instructs: "Answer using only the information in the provided context. Do not add information not present in the context. If the context does not contain enough information to answer, say so and offer to connect the member with a representative."
4. Citation verification: After generation, a post-processing step verifies that every factual claim in the response maps to a passage in the retrieved context. Claims that cannot be traced are removed and the response is regenerated or the system escalates.
5. Source corpus version control: The vector database is updated within 24 hours of any source document change (plan update, policy revision, benefits change). Stale-corpus incidents are treated as production incidents.

6.3 Model Training & Tuning

• M-06 (Conversational AI): Foundation LLM (vendor-provided) fine-tuned on ACME-specific Q&A pairs (generated from historical call transcripts and policy documents). Fine-tuning dataset: 50,000+ Q&A pairs, reviewed for accuracy by ACME's member services quality team.
• M-07 (Intent Classifier): Trained on 100,000+ labeled call transcripts (intent annotated by ACME call center quality team). Intent taxonomy defined during JAD sessions.
• M-08 (Sentiment Analyzer): Trained on 30,000+ labeled call segments (sentiment annotated). Additional tuning for distress detection using clinical escalation call samples.
• M-09 (Case Summarizer): Fine-tuned on 20,000+ call transcripts paired with human-written case notes. Evaluated on summary completeness and factual fidelity.

6.4 Production Monitoring

• Hallucination monitoring: Daily automated scan of virtual assistant responses for hallucination indicators (claims not traceable to source corpus). Weekly human review of 100 randomly sampled interactions for factual accuracy.
• Escalation rate monitoring: If escalation rate exceeds 30% for any supported inquiry type (indicating the system is failing to answer routine questions), the AI CoE investigates within 5 business days. Root cause may be: source corpus gap, model drift, or new inquiry patterns not in training data.
• Member satisfaction monitoring: Post-interaction CSAT survey offered to members after virtual assistant interactions. If CSAT drops below 3.5 / 5.0 for any channel, BRD-02 product team investigates within 5 business days.

7. Data Requirements

8. Integration Requirements

9. Non-Functional Requirements

9.1 Performance

• Virtual assistant response time: P95 ≤ 3 seconds for text responses (web/mobile); P95 ≤ 2 seconds for voice responses (measured from end of member utterance to start of system speech).
• Agent-assist suggestion latency: P95 ≤ 3 seconds from end of member utterance to suggestion display in agent side panel.
• Concurrent session capacity: Support 500 concurrent virtual assistant sessions (approximately 3× average peak) without degradation.
• Voice channel: Speech-to-text accuracy ≥ 92% across major accent groups in ACME's member population. Text-to-speech naturalness rated ≥ 4.0 / 5.0 by member test panel.

9.2 Availability & Resilience

• Availability target: 99.9% uptime for virtual assistant (24/7 service); 99.5% for agent-assist (business hours).
• Graceful degradation: If AI models are unavailable, the virtual assistant shall display: "I'm temporarily unable to assist. Please call [ACME service number] or try again shortly." It shall not display error codes, stack traces, or technical messages to members.
• RTO / RPO: 2 hours / 30 minutes (stricter than BRD-01 due to 24/7 member-facing nature).

9.3 Security

• All conversation data encrypted at rest (AES-256) and in transit (TLS 1.3).
• Member authentication required before any member-specific data is accessed or displayed.
• Adversarial input protection: prompt injection defenses (system prompt is not modifiable by member input), input sanitization, and output filtering (no PII/PHI leakage in non-authenticated contexts).
• Conversation data isolated by member — no cross-member data leakage in multi-session environments.
• Penetration testing before production and quarterly thereafter.

9.4 Accessibility

• WCAG 2.1 Level AA compliance for web and mobile virtual assistant interfaces.
• Screen reader compatibility (ARIA labels, semantic HTML, keyboard navigation).
• Voice channel: support for hearing-impaired members via real-time captioning/TTY integration.
• Section 508 compliance for federal exchange member access.

10. Human Escalation & Override Requirements

BRD-02's escalation model is confidence-based and member-directed, fundamentally different from BRD-01's adverse-determination-triggered model. The system escalates whenever it cannot provide a grounded answer OR whenever the member requests it — whichever comes first.

10.1 Escalation Triggers (Virtual Assistant)

10.2 Agent-Assist Override

• Agents are never required to use agent-assist suggestions. Every suggestion is a recommendation; the agent has full authority to ignore, modify, or override any AI suggestion.
• If an agent identifies an incorrect suggestion, a one-click feedback mechanism ("incorrect suggestion") flags the case for AI CoE review. Persistent incorrect suggestions in a topic area trigger model retraining assessment.

11. Fairness, Bias & Responsible AI Requirements

• Response quality parity: Virtual assistant response accuracy, completeness, and helpfulness shall not vary by member demographics. Testing includes stratified accuracy measurement across age bracket, geography (urban/rural), plan type, and language preference. Deviation > 5% triggers investigation.
• Escalation rate parity: Escalation-to-human rate shall not vary by >5% across demographic groups. If certain member segments are disproportionately escalated (e.g., members in rural areas, older members, non-native English speakers), this indicates a model gap that must be addressed.
• Voice recognition parity: Speech-to-text accuracy on the voice channel shall not vary by >5% across accent groups. Accent groups for testing are defined based on ACME's member population demographics.
• Accessibility as equity: Members with disabilities shall receive the same quality of service as other members. Accessibility is not a compliance checkbox — it is a fairness requirement.

12. Testing & Validation Requirements

13. Acceptance Criteria & Definition of Done

BRD-02 is accepted into production when ALL of the following criteria are met:

1. All "Must" functional requirements (FR-02.001 through FR-02.020) implemented and tested
2. All 4 AI models pass Independent Model Validation (zero Critical, zero uncorrected High)
3. Hallucination rate ≤ 0.5% on adversarial test set (500+ questions)
4. Factual accuracy ≥ 97% on benchmark Q&A set
5. UAT passed (≥ 95% pass rate, zero open Critical/High defects)
6. Pilot member panel: ≥ 90% task completion rate and CSAT ≥ 4.0
7. Accessibility testing passed (WCAG 2.1 AA across all channels)
8. Voice accent parity verified (≤ 5% STT variance)
9. Security testing passed (zero Critical/High; prompt injection defenses verified)
10. Escalation workflows verified end-to-end (confidence-based, complaint-based, explicit request, after-hours)
11. Agent-assist tool verified with live agent panel (suggestion quality, latency, case summarization accuracy)
12. All 3 governance boards sign off at Phase 3 gate
13. Production monitoring dashboards operational (hallucination, escalation rate, CSAT, latency)
14. Runbook and operational documentation delivered to IT Operations and call center management

14. Constraints, Assumptions & Dependencies

14.1 Constraints

• All conversation data (PHI) onshore US only
• Virtual assistant must identify itself as AI — no human impersonation
• Member escalation to human must be honored within one turn — no retention attempts
• RAG architecture: no response generation without retrieval confidence ≥ 90%
• Source corpus maintained by ACME business teams, not AI delivery team
• WCAG 2.1 AA accessibility — non-negotiable

14.2 Assumptions

• ACME's call center telephony platform supports SIP integration for voice channel
• Historical call transcripts (3 years, de-identified) available for model training
• Benefits summary documents and medical policies available in machine-readable format (or can be converted in Phase 1)
• Call center agents willing to adopt agent-assist tool (Change Management support via G. Okonkwo's team)
• Member adoption of virtual assistant follows industry norms (~30-40% adoption in first 12 months)

14.3 Dependencies

• DEP-03: Data & Cloud Platform Foundation must be live (inherited from Year 1)
• DEP-07: BRD-01 infrastructure and governance patterns reused — BRD-02 does not rebuild platform or governance from scratch
• DEP-08: Call center telephony platform integration API available for voice channel development
• DEP-09: Mobile app development team capacity available for virtual assistant widget integration

15. Regulatory Traceability Matrix

16. BRD-02 Delivery Team & Governance

Team size: 22 people. Lead: X. Garcia (Pulaski). Full roster in Resource Plan, Team 22.

Governance: BRD-02 operates under the full Project Catalyst governance framework. Sprint-level execution uses 2-week sprints with Scrum Master (T. Okoro5), Product Manager (K. Nakashima4), and daily standups. All model releases require Independent Model Validation sign-off and AI Governance Board approval.

Business Sponsor: M. Washington (VP Customer Service Operations, ACME) — owns business requirements, call-center workflow validation, and member experience standards. Works with the Escalation Protocol Designer (Q. Rutherford3, ACME) and Call Center Ops Liaison (R. Alvarado3, ACME) to ensure the AI capabilities integrate with existing call-center operations.

17. Sign-Off & Approval