# AI Assurance Pack — Customer-Support RAG Agent ## 1. Executive assurance brief **Recommendation:** HOLD **Confidence:** LOW **Boundary:** Customer-support RAG agent; model, version, retrieval source, interfaces and deployment environment are MISSING. No evidence was supplied. This preliminary pack does not establish production readiness. The immediate decision is to hold approval while preserving a usable assurance structure and minimum evidence plan. ## 2. System/model card | Field | Current record | Status | |---|---|---| | Purpose | Customer-support retrieval and response assistance | PROVIDED — NOT VERIFIED | | Owner/version/model | MISSING | MISSING | | Intended users and uses | Support personnel and customer-support responses | INFERRED | | Data, retention and access | MISSING | MISSING | | Human oversight and fallback | MISSING | MISSING | | Evaluation, monitoring and incidents | MISSING | MISSING | ## 3. Assurance case | Claim ID | Claim | Evidence | Status | Residual uncertainty | |---|---|---|---|---| | CLM-01 | Responses are sufficiently accurate and grounded | None supplied | UNSUPPORTED | Hallucinated or outdated guidance | | CLM-02 | Customer data is appropriately protected | None supplied | UNSUPPORTED | Exposure through retrieval, prompts or logs | | CLM-03 | Failures are detected and escalated | None supplied | UNSUPPORTED | Silent harmful responses | | CLM-04 | Use is bounded by human oversight | None supplied | UNSUPPORTED | Excessive reliance or autonomous action | ## 4. Evidence register No evidence artifacts were supplied. Evidence IDs remain unassigned until artifacts or redacted summaries are provided; requests appear in section 9. ## 5. Limitations and usage boundaries Do not use for binding account decisions, safety-critical advice, identity verification or irreversible customer actions. Require human review for complaints, financial consequences, sensitive-data requests and low-confidence retrieval. Disclose that responses may be incomplete or incorrect. ## 6. Residual-risk register | Risk ID | Scenario | Current controls | Assessment | Treatment | |---|---|---|---|---| | RR-01 | Unsupported response reaches a customer | MISSING | Potentially high; unscored | PROPOSED grounding test and human escalation | | RR-02 | Sensitive data appears in retrieval or logs | MISSING | Potentially high; unscored | PROPOSED access, redaction and retention tests | | RR-03 | Prompt injection changes behavior | MISSING | Material; unscored | PROPOSED adversarial evaluation and tool boundary review | ## 7. Conditions and remediation plan 1. Assign accountable system and risk owners. 2. Supply architecture and data-flow evidence. 3. Run task, grounding, security and privacy evaluations with acceptance thresholds. 4. Define human escalation, monitoring, incident and rollback procedures. 5. Reassess all unsupported claims before deployment approval. ## 8. Executive sign-off page **Decision:** HOLD **Approval roles:** PROPOSED Product Owner, Security/Privacy Reviewer and Accountable Executive **Blocking conditions:** CLM-01 through CLM-04 remain unsupported. **Next review trigger:** Minimum evidence set received and evaluated. ## 9. Evidence requests and clarification questions Request: architecture/data-flow diagram, evaluation report, access and retention controls, monitoring/incident records, and human-oversight procedure. 1. What model, retrieval source and version are in scope? 2. What customer data can enter prompts, retrieval or logs? 3. Which actions require human approval? 4. What quality and grounding thresholds govern release? 5. Who owns approval and residual-risk acceptance?