Automated Financial Report Generation

A mid-sized asset manager produced monthly investor reports manually. Analysts spent three to four days each month extracting figures from the ERP, writing variance commentary, and formatting output for distribution. The process was bottlenecked by a single analyst, creating delivery risk.

G|AI Works designed a three-stage pipeline:

1. Data extraction layer: A scheduled job pulled portfolio performance metrics, benchmark comparisons, and attribution data into a validated JSON schema. Validation rules enforced completeness and value ranges before any LLM call was made.

2. Narrative generation: A versioned prompt template — reviewed and signed off by the client's compliance team — injected the validated data into a structured generation request. Output was constrained to specific narrative sections: performance summary, attribution commentary, and risk indicators.

3. Human review interface: A lightweight web interface presented draft reports side-by-side with the source data, allowing the reviewing analyst to approve, edit, or reject individual sections before final assembly.

• Around 70% reduction in analyst time per reporting cycle (from 3.5 days to under 1 day)
• No data errors logged in generated reports across the first six months of operation (grounding validation catches mismatches before output)
• Consistent terminology across all reporting periods, eliminating manual style review

• Data extraction: Python + SQLAlchemy against existing ERP
• LLM: claude-sonnet-4-6 with structured output
• Validation: Pydantic schema enforcement on both input and output
• Review interface: lightweight Astro + server actions