Finance agent

Inspiration

As the Finance Analytics team at the largest bank in Colombia, we constantly seek ways to transform financial data into strategic insights. Every month, teams across the organization request detailed breakdowns of financial performance by department — questions like “How did Sales impact net income this quarter?” or “Can we cut costs in Operations without hurting revenue?” are routine. But answering them takes time, manual effort, and domain knowledge.

We were inspired to create an AI agent that could instantly interpret financial statements, segment them by area, and answer these questions on the fly — just like a senior analyst would. We wanted a tool that could not only automate analysis but also communicate the story behind the numbers.

How We Built It

• We created a synthetic financial dataset with two years of income statements and balance sheets, enriched with business area mappings (Sales, Finance, Operations, etc.).
• Using Apache Spark on Databricks, we modeled this data as a clean DataFrame to simulate enterprise-scale financials.
• We built an AI agent that uses natural language to query and explain financial data, answering complex questions like YoY performance changes or departmental efficiency.
• To take it further, we integrated a RAG (Retrieval-Augmented Generation) pipeline to automatically generate narrative notes for financial statements — and exported these as PDF reports.
• The result is a self-service financial analyst that combines LLM reasoning, business logic, and automation.

What We Learned

• LLMs can replicate financial reasoning when grounded in structured data and context-aware prompts.
• RAG is a powerful approach for generating consistent and contextual narrative explanations of numeric data.
• With the right data modeling, AI agents can explain not just the "what", but also the "why" behind financial performance.

Challenges We Faced

• Designing prompts that mimic a financial analyst’s reasoning without drifting into hallucinations.
• Aligning the AI agent’s responses with real-world financial logic, including proper handling of negatives, margins, and costs.
• Balancing data realism with hackathon speed — creating a synthetic dataset that feels real but is fast to iterate on.

Final Thoughts

We started with a simple idea: bring financial intelligence to life with AI. What we built could easily evolve into a real enterprise tool — giving every team, from sales to the CFO, instant clarity into the numbers that matter.

Built With

• databricks
• openai
• python