Conversational AI Agents Over Enterprise Trading Data

We built a multi-agent system where each agent specializes in a domain: a Portfolio Agent handles position and exposure queries, a Risk Agent interrogates VaR and Greeks, and a Performance Agent handles attribution and returns. Each agent connects to the appropriate data source — Snowflake for historical, PostgreSQL for live positions — through a secure query layer that enforces row-level security matching the user's existing Bloomberg and internal access permissions. Natural language queries are parsed by GPT-4, converted to parameterized SQL via a schema-aware intermediary (not raw LLM-generated SQL), validated against an allowlist of operations, and executed only after a deterministic security check. Every query is logged with user identity, time, and generated SQL for compliance audit trails.

In a regulated environment, the agent can never have broader data access than the human asking the question. We implemented a permission mirroring system that maps Bloomberg terminal access groups to database row-level security policies. An analyst with access to only European equity books gets the same restriction in the AI agent — verified at query time, not just login time. All agent infrastructure runs within the firm's Azure private network. No query data leaves the corporate boundary. The LLM receives only the schema structure and query intent — never raw financial data.

The agents were deployed through a React-based chat interface integrated into the firm's internal portal, accessible via SSO. We ran a two-week pilot with 12 power users, iterated on query quality and hallucination edge cases, then rolled out to 180+ analysts firm-wide. Adoption was organic — within 72 hours, agents were handling over 400 queries per day, routing the data team's queue to near zero for standard analytical questions.