Context
The project was developed in a context of high operational complexity in the management of benefit requests handled by INPS, which require the submission and manual verification of large volumes of heterogeneous and non-standardisable documents. Every year, operators are required to review millions of documents of different types and contents within tight deadlines, resulting in a significant impact on workloads, error risk, and response times for citizens.
In this scenario, the Institute launched an innovation initiative aimed at modernising the document verification process through the adoption of artificial intelligence solutions.
Objectives
The objective of the project is to optimise and strengthen the document verification process by reducing the burden of repetitive manual activities, improving the consistency and reliability of checks, and accelerating the processing times of applications.
The solution is designed to support operators in document analysis and validation activities, ensuring greater uniformity in the decision-making process and an overall higher quality of service delivered to citizens, while preserving the central role of human decision-making.
Implementation
The project was implemented through a fully on-premise and open-source GenAI solution, integrated into the Institute’s document management systems. The platform combines OCR technologies, bidirectional encoder representations from transformers (BERT)-based classification models, vision models, and large language models to automate document classification, data extraction and verification, and to provide decision support to operators.
The implementation followed a progressive approach, starting with a pilot phase launched in April 2025, which enabled the analysis of thousands of documents, delivering significant results in terms of accuracy, error reduction, and operational efficiency improvements. Subsequently, the solution was extended to an additional benefit service starting in October 2025, with a potential workload of approximately 4.2 million documents.
From March 2026, the adoption of the solution is planned across two further benefit schemes, covering an estimated additional total of 260 000 documents.
The system demonstrates high quantitative performance levels, including:
- around 99.95% precision, with no false positives generated by the solution
- approximately 90% of AI extracted checks confirmed by human operators
- a maximum processing time of 15 seconds per document
- approximately 90 different checks and extracted information automatically managed
- approximately 16 different document types supported.