How the Italian Social Security and Welfare Administration (INPS) used artificial intelligence to streamline the sorting of certified emails (CEs)

The Responsible Organisation

The Italian Social Security and Welfare Administration (INPS) is the main public institution in the field of social security and assistance in Italy in terms of the volume of benefits and services provided and the number of users, as well as one of the largest at European level, with a budget that is the second largest within the Italian public sector, only behind that of the State. Today, more than 22 million Italians receive their pension, and more than 25 million workers are insured through INPS (INPS, 2023).  The Institute has had more than 20,000 employees and 448 territorial offices across the country as of 2022, representing one of the main points of contact between citizens and the national institutions.

The problem

One of the main channels that citizens use to communicate with INPS is via Certified Emails (CEs). Through them, citizens can contact the Institute to seek information or transmit relevant paperwork to apply for welfare services or entitlements (e.g., civil invalidity benefits, unemployment benefits). During and after the pandemic the number of incoming CEs increased significantly, from 3 million in 2019 to over 6 million in 2023 leading to a significant surge in the workload for INPS employees.  In fact, they used to manually examine each incoming e-mail and its attached files, classify them into various categories, and sort them to the departments and teams responsible for responding to specific inquiries on those topics. This task became increasingly time-consuming, repetitive, and potentially subject to clerk errors.

The solution and its implementation

In early 2021, the repetitive nature of the manual work of scanning and deriving the received CEs presented opportunities to test potential solutions by innovating with emerging technologies. Consequently, INPS started experimenting with Artificial Intelligence. The Institute began evaluating various machine learning models and ultimately selected BERT, an open-source model developed by Google that can be exploited by software engineers to create ad-hoc codes to process natural language. Fine-tuned with the organisation’s email data and employees’ feedback, the BERT-based solution was piloted in-house at INPS Data Centres to ensure compliance with GDPR regulations safeguarding citizen data privacy. 

The system analyses all CEs and their attachments, accurately classifies them into the relevant topics and directs them to the designated personnel, effectively eliminating the need for manual processing. Over two million emails have been successfully classified since its deployment and it has been implemented in the INPS’ offices of fifteen Italian cities, including Rome, Milan, and Naples. Given its technological scalability, it is planned to replicate the model in other cities. This solution was also awarded by IRCAI, UNESCO's International Research Centre for Artificial Intelligence, in the world Top 10 projects supporting the 17 UN Sustainable Development Goals (SDGs) through advanced Artificial Intelligence.

Expected benefits

The BERT-based model offers a wide range of benefits to the Institute:

• The communication between citizens and the Institute is quicker and CEs are now assessed by competent employees earlier than before. As of today, more than 2 million CEs have been already sorted by the solution, with an accuracy rate above 80%.
• When fully operational, it is estimated that between 30,000 and 40,000 working days per year will be saved across all local offices. 
• Clerks previously dedicated to manual sorting of CEs can now devote time to other higher value-added activities, thus bringing further benefits to citizens.
• The development and adoption of the solution created internal technological ‘know-how’ and capabilities that can be reused to enhance other services, such as the online customer centre which is currently being optimised using a similar AI sorting application.

Main challenges

One of the main challenges encountered in the development of the AI-based system has been the availability of the data necessary to train the model, both in terms of size quantity and in terms of temporal range. Moreover, the accuracy and the quality of data has been of paramount importance for the accuracy of the results: in fact, as the AI tool learns from data, if data is fallacious, so will be the outputs of the AI algorithm. 

A second challenge is on the degree of adaptability of the tool. As a Social Security and Welfare Administration, INPS services are everchanging to adapt to new needs of the citizens. Therefore, the AI model needs to be adjusted, the results are to be continuously monitored, and any potential degradation in the accuracy of the results must be fixed.

Finally, aside from technology, operational challenges have been encountered. On one hand, some of them have been already solved. For instance, at first, employees were curious to learn how the system would work in order to understand how it could help them. Thus, training sessions were organised to teach employees about the AI technology. 

On the other hand, other operational challenges persist. As an example, each of the INPS offices distributed across the country has a different offer of services depending on regional, demographic and socioeconomic differences: certain offices might handle mostly pension-related inquiries, while others might manage disability services, unemployment benefits, or other services. The diversity of services offered by INPS offices adds a layer of complexity to the training of the AI algorithm, that needs to be trained to recognise and categorise a wide range of topics and requests. In addition, offices often have unique administrative workflows. This entails wide differences in which departments are responsible for responding to specific requests in across offices and, as a result, in the way that the AI tool must be trained to understand variations in routing emails. Therefore, a “one-size-fits-all” model, even if preferrable, may be hard to achieve. INPS is tackling this challenge by grouping offices in terms of service similarity, and then applying tailored models to each cluster. 

• Article: “INPS: here is how we use AI to benefit citizens”
• INPS website
• IRCAI UNESCO Award for being among the top 10 AI projects worldwide
• YouTube video presentation