Arianna and the promise of the agentic State: INPS launches an AI assistant for over 400 benefits

On 27 October 2026, just days after Legislative Decree 179/2026 implementing the AI Act in Italy came into force, INPS, the Italian national social security institute, presented its new digital assistant, Arianna, at its Directorate General headquarters in Rome. Already available on the INPS portal and in the INPS Mobile app, it currently works before login: it provides information and guidance on more than 400 benefits in response to natural-language queries. The Government has framed the initiative as a step towards the so-called agentic State championed by the Undersecretary for Innovation, Alessio Butti, while INPS President Gabriele Fava stated that “for generations it was the citizen who had to understand the State. Now it is the State that must learn to understand the citizen.”

According to the Institute, the assistant's answers rely exclusively on sources verified by INPS and run on a multi-agent architecture built with open-source frameworks and generative AI models. A report published by ANSAcom and produced in collaboration with INPS states that Massimiliano D'Angelo, Central Director for Technology, IT and Innovation, described a two-track strategy: Arianna handles only public information using “market-standard” models, OpenAI among those mentioned, adapted to INPS guardrails, while a separate model for citizens' data relies on open models, including France's Mistral and Italy's Velvet. The official INPS announcement does not name these specific vendors, though it does preview future developments for post-login assistance, filling in applications and simulators.

At the launch, INPS released a few performance indicators: 97% of use cases completed successfully and a Customer Satisfaction Score of 4.5 out of 5. The institute did not, however, disclose the testing methodology, the number of use cases or the testing period, and the results have not been independently verified. Nor does there appear to be any published data protection impact assessment or information on the costs of developing and maintaining the infrastructure.

Simplifying access to bureaucratic benefits is a crucial step for public service, but an assistant's effectiveness is not measured solely by how well it answers at the information stage: above all, it lies in the transparency of the criteria by which it is evaluated: testing methodology, models in use and impact on data. — Olya