Financial Services Policy Assistant

A tier-1 bank wants an AI assistant over internal policies, procedures and customer-support documentation. Support agents should get instant, cited answers to policy questions. Compliance must be able to audit every query and answer. Data access must respect role-based permissions.

• 1Which document sources matter most — policies, procedures, support tickets, regulatory filings?
• 2What are the access-control groups and how are they managed today (AD, Okta)?
• 3What is the acceptable answer latency, and what is the peak concurrent user count?
• 4What is the audit retention requirement, and what fields must be logged?
• 5Are there model-output restrictions (no advice, citation required, confidence threshold)?
• 6What is the fallback when the assistant is uncertain — escalate to human or refuse?
• 7Which languages must be supported?
• 8What is the change-management process for updating policies, and how fast must updates propagate?

• Regulated industry — full audit trail of queries and answers (7-year retention)
• Role-based access control — agents must not see content outside their permission group
• Citations required on every answer
• Model must refuse to give financial advice
• On-premise or private-cloud deployment preferred
• Answer latency < 2s P95
• Cost ceiling $0.05/query

• Hybrid retrieval over policies (keyword-strong) and procedures (semantic-strong)
• Access-control metadata filter at query time, not display time
• Reranking stage for precision
• Citation with document title, section, and chunk ID
• Confidence score with low-confidence escalation to human reviewer
• Full audit log: user, timestamp, query, retrieved chunks, model, response, confidence
• Policy update pipeline with freshness SLA (new policy searchable < 1 hour)
• Online evaluation sample for drift detection