Aplicación de IA generativa en el análisis automatizado de información y asesoría financiera: desarrollo de un prototipo orientado a Pymes
Year: 2025
Type: dissertation
Abstract: This work addresses a persistent challenge for small and medium-sized enterprises (SMEs): turning financial statements into timely strategic decisions with limited access to financial specialists.A prototype based on generative AI that ingests, standardizes, and interprets financial information and produces recommendations on operations, financing, and investment.This prototype integrates firm-level data, macroeconomic context, and domain knowledge.Implemented in Python with Streamlit for the front end, the system ingests financial statements (PDF/Excel), extracts and normalizes key items (currency/scale and period), and generates analyses and recommendations from a structured prompt.Retrieval is supported by RAG with FAISS to surface document fragments.Three models-GPT, Claude, and Gemini-are used to produce the analyses.The evaluation required each analysis to report 19 financial indicators per company, compar outputs against a ground truth built from manual extraction and calculation across 10 Colombian firms (7 SMEs, 3 large firms).Model performance was quantified as capture, coverage, and precisión.The results were as follows: GPT (capture 13%, coverage 13%, precision 100%), Claude (capture 14%, coverage 76%, precision 69%), and Gemini (capture 17%, coverage 67%, precision 88%).The results demostrate that the prototype can capture data, compute indicators, and generate decision-oriented diagnostics.Qualitative comparisons also assess interpretation and recommendations.While GPT-4o, Gemini 2.5 Pro, and Claude Sonnet 4.5 vary, preprocessing quality is the primary performance driver quality: Gemini balances coverage and accuracy best; Claude attains higher coverage with lower average precision and GPT is highly precise on a smaller set.Overall, the prototype meets the research objectives by delivering grounded analyses and recommendations.It identifies information requirements, implements a scalable solution, and demonstrates that reliable capture and normalization are critical for AI to add managerial value.
Author: Granda Alvarez, Juan Pablo
Language: Spanish
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7115807960