AI Intelligence for More Efficient Public Procurement
Executive Summary
Prodemge is the Information Technology company of the State of Minas Gerais, responsible for digitally supporting strategic secretariats such as Finance, Health, Justice and Public Security, Education, and Planning and Management. Operating in an environment where precision and transparency in public procurement are mandatory, Prodemge sought a solution capable of modernizing this process with the use of Artificial Intelligence.
The public procurement scenario presents structural challenges: bureaucracy, subjectivity in price definition, and complex workflows that hinder both supplier participation and the achievement of fair and consistent prices. This context generated differences between the prices researched and those effectively practiced, with a direct impact on costs and the quality of acquisitions.
As part of its Open Innovation initiative (TREM), Prodemge adopted Genius to develop an AI-based price prediction solution. The approach centralized government data, incorporated advanced semantic search with Elasticsearch and BERT, deep learning for future price forecasting, and human curation features for result validation.
The result was significant: the average price research time dropped from 36 to 8 days, with a simultaneous gain in the accuracy and reliability of the determined values.
The Challenge
Public procurement operates under strict compliance and transparency requirements, making the definition of reference prices a high-responsibility process. The scenario demanded structural evolution to make this process more agile, accurate, and auditable.
Bureaucracy and subjectivity in the pricing process: Price research workflows were extensive and subject to varied interpretations, hindering standardization and increasing the time required to complete each procurement process.
Discrepancy between researched and practiced prices: The absence of a robust analytical base generated divergences between the values collected in research and those effectively contracted, impacting costs and the quality of acquisitions.
Difficulty in identifying similar items: The variety of nomenclature for materials and services made item comparison an imprecise process, compromising the quality of the price references used.
The Solution
Prodemge, in partnership with Squadra Digital, developed an AI-powered public price prediction platform, integrated with the historical data of the National Public Procurement Portal. The solution was structured in two complementary fronts: data intelligence for analysis and information retrieval, and predictive modeling with human validation.
Genius in action: AI-Powered Price Prediction for Public Procurement
Data Intelligence: Centralization and Advanced Semantic Search
The first front structured the solution's analytical base, ensuring that historical procurement data was accessible, comparable, and visually interpretable.
Centralization of government data: Consolidation of information from the National Public Procurement Portal into a single structured database, enabling reliable historical analysis for any item, whether material or service.
Elasticsearch and BERT (semantic search): Advanced semantic search engine to identify similar items even with variations in nomenclature, expanding the coverage and accuracy of price research.
Interactive charts: Visualization of prices within the standard deviation, allowing clear reading of the historical value distribution and supporting more informed decisions.
Prediction and Curation: Price Recommendation with Deep Learning
The second front incorporated predictive modeling and a validation workflow that combines AI and human judgment.
Deep learning for future price forecasting: Models trained with public procurement history to recommend prices considering seasonality, period, and item similarity, without restriction on the volume of processed items.
Integrated human curation: Dedicated features for validating and adjusting predicted values, ensuring that AI accuracy is complemented by the judgment of public managers.
Results
Price research up to 4.5x faster: the average research time dropped from 36 to 8 days, freeing up teams for higher strategic value activities.
Greater accuracy in defining reference prices: the use of real historical data, semantic search, and predictive modeling reduced process subjectivity and brought researched values closer to those practiced in the market.
Auditable process with human curation: the combination of automated prediction and human validation ensured traceability and reliability of recommended prices.
Metric | Result |
Average price research time | From 36 to 8 days |
Reduction in process time | 78% |
Analyzable item coverage | Unlimited |
Gallery
Want results like these?
Talk to our experts and find out how AI can transform your business.