Most of the literature on AI-assisted text simplification focuses on the English language, making the simplification of Italian texts a relatively unexplored field, despite its importance in the context of public administration. This study aims to provide an innovative contribution to this area, focusing on the simplification of administrative texts in Italian.

Existing metrics often fail to capture all aspects of text simplification or correlate with human judgment. This study incorporates user evaluation to assess whether analytical metrics actually correlate with improved user perception and introduces GULBERT, a novel simplification metric that balances readability and semantic preservation, specifically tailored for Italian administrative texts.

By incorporating real user assessments on factors such as clarity, engagement, and confidence in the content, the study ensures that AI-generated simplifications align with user needs.

The experimentation on text simplification has been developed as a prototype and has successfully passed the evaluation stage; it is now transitioning into an operational business process by INPS and its technology partners, integrating a graphical interface to support the human-in-the-loop approach, in the form of a collaborative workflow.

Alternative promising developments are in the evaluation stage, such as the generation of Service Pages from scratch, using exclusively technical documents and regulations via RAG (Retrieval Augmented generation), while ensuring compliance with completeness and accessibility criteria, or the introduction of a source-referencing system. Moreover, the currents results are being prepared for publication as a comparative analysis.

INPS (Institutional Partner):

• Defined the strategic goals, use case and evaluation criteria
• Coordinated all phases of the experimentation, enabling access to real-world administrative documents
• Offered domain expertise

Almawave:

• Study of the literature, conceptualization of the experimentation
• Development and industrialization of the prototype

Almaviva:

• Process management support

Unguess:

• Developed and administered the user validation phase

• INPS, the main stakeholder, optimized the document simplification process, improving its communication effectiveness, fostering inclusiveness, transparency and trust
• Citizens, the primary beneficiaries of a more inclusive communication ecosystem, which democratizes access to complex administrative information, especially for individuals with reading difficulties, limited language skills, or lower educational backgrounds.

The AI effectively simplifies text, delivering expert-level results in a fraction of the time. Its robust metrics framework accurately evaluates various aspects of simplification and highlights areas for improvement.

A large-scale survey, including A/B testing with 1620 participants, confirmed the AI’s effectiveness. The AI-generated text was preferred in more than 70% of responses vs. the original text, and outperformed the original across all metrics, scoring significantly higher in fluency (+14.1%), readability (+13.5%), clarity (+11.11%), and engagement (+12.3%).

These findings highlight key benefits: scalable AI-driven simplification for public administration, improved accessibility to administrative information, and seamless integration into broader communication strategies.

AI-driven text simplification for administrative documents must navigate challenges like maintaining linguistic consistency, preserving key information, and balancing clarity with depth. Traditional evaluation metrics, such as BLEU and ROUGE, fall short in capturing these complexities, highlighting the need for more advanced assessment frameworks.

To address these gaps, strategies like the novel GULBERT metric, large-scale user testing, and iterative model enhancements have been successfully implemented.

Looking ahead, efforts will focus on strengthening AI-human collaboration, expanding datasets, and refining simplification techniques to ensure continuous improvement and more effective, accessible communication.

Success requires a public administration that truly puts users at the center. AI adoption must be guided by a governance framework that prioritizes usefulness and user impact over hype and novelty. Clear policies ensure legal compliance, while strong leadership aligns innovation with public value. Robust infrastructure and adequate funding enable scalability, but it is skilled staff and a shared commitment to accessibility that ensure quality. Only by focusing on citizen needs can AI become a sustainable tool for better services.

This AI-driven text simplification model demonstrates significant potential for broader application, particularly in sectors such as healthcare and legal services, where complex documents require greater clarity.

Its success in streamlining bureaucratic Italian texts highlights its adaptability. Within public administration, the model could scale across departments—such as social services and taxation—or be implemented in other governmental systems, accommodating diverse languages and regulatory frameworks.

Ensuring a successful expansion requires maintaining a careful balance between automation and human oversight, preserving both accuracy and public trust in official communications.

Key insights from this AI-driven text simplification initiative:

• Human-AI collaboration is critical—automation excels at scaling simplification, but human oversight ensures legal accuracy and nuance.
• User validation is not negotiable: metrics alone (e.g., readability scores) don’t capture users’ preference; direct feedback is needed to effectively assess the efficacy of the model.
• Context matters—effectiveness hinges on tailoring prompts/models to institutional language (e.g., Italian bureaucracy).

Future adopters should prioritize hybrid workflows, localized testing, and clear success metrics.

This case was inspired by a need-driven innovation approach, a strategic pillar in INPS. It started from a well-identified user need—making administrative texts easier to understand—and an organizational pain point: a manual simplification process that was time-consuming and resource-intensive. Generative AI was tested as a targeted solution to address both. The project defined measurable goals, applied a rigorous validation process, and demonstrated concrete value. It shows how innovation can be effective when guided by real needs, not by the technology itself.