Name in original language
ANÁLISIS DE DOCUMENTOS MEDIANTE MODELOS DE INTELIGENCIA ARTIFICIAL PARA LA DETECCIÓN DE ANOMALÍAS EN PROCESOS DE COMPRA PÚBLICA
Initiative overview
Public procurement monitoring requires expert knowledge applied to a complex and increasing volume of legal and administrative documents. AI algorithms, including large language models (LLMs), could help to ensure broad coverage, as well as provide rapid and accurate analysis to detect compliance breaches, prevent corruption risks, diminish inefficiencies and reduce potential misuse of public funding.
The initiative pursues the following objectives:
- Extract, read and process unstructured data from text documents submitted by government agencies and suppliers.
- Develop and implement an LLM model trained on public procurement regulations that allows the detection of compliance breaches.
- Enable processing, updating and data analysis of information on a daily basis for large volumes of files in short timeframes.
This initiative is part of ChileCompra’s broader strategy to incorporate AI responsibly within the public procurement system. It includes complementary projects such as AI-based corruption detection tools, predictive price comparison instruments and models to promote supplier participation, especially among small and medium enterprises.
Moving forward, the ethical governance framework will be applied throughout the entire lifecycle of the institution’s AI systems, incorporating continuous risk assessments, bias monitoring, model validation, impact tracking and public transparency reporting. The goal is to institutionalise these practices and gradually extend them to additional platforms within the public procurement ecosystem.
Results, outcomes and impacts
The initiative started its implementation in January 2026. The first outcomes of the project include an overall design of the technological architecture, selection and prioritisation of red flags, as well as quality assurance indicators and strategies.
The proposed metrics to assess implementation results are organised into six main dimensions:
- Ground Truth or minimum evaluation baseline;
- rates of automatically detected anomalies;
- rates of confirmed compliance breaches;
- rates of correction of detected anomalies;
- OCR processing quality metrics;
- and language model (LLM) performance metrics.
