Vincenzo Giliberti has a total of 4 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Governed Agentic Process Automation: A Floor-Safety Guarantee for Compliance-Critical LLM Routing

    by Massimo Pacella, Gabriele Papadia, Vincenzo Giliberti

    Agentic Process Automation (APA) extends Robotic Process Automation by delegating workflow construction and runtime decisions to Large Language Model (LLM) agents. In regulated business processes, this autonomy creates a specific safety risk: an LLM may recognize elevated risk while still failing to trigger the human approval required by policy. We propose Governed APA, a runtime decision architecture that bounds LLM autonomy through schema-validated inputs, a closed action space, a policy-derived floor guardrail, and a deterministic fallback. We formalize the governed decision and establish a floor-safety property ensuring that the executed action cannot under-escalate below the policy-mandated minimum. The architecture is evaluated on a structured employee-onboarding case study with 166 profiles and explicit human-in-the-loop (HITL) ground truth. We compare a deterministic baseline, an ungoverned LLM, and a guarded LLM using two local models, Qwen2.5 and llama3.1. Ungoverned Qwen2.5 identified elevated risk but failed to escalate any of the 43 sensitive cases to mandatory human approval, yielding HITL recall equal to 0. Ungoverned llama3.1 showed the opposite failure mode, achieving full recall but producing unnecessary human-approval escalations. With the floor guardrail active, Qwen2.5 HITL recall increased to 1.0, llama3.1 recall remained 1.0, and inter-run stability of the compliance decision increased from 0.73 to 1.0 for Qwen2.5. On this dataset and configuration, no false-positive HITL escalation was introduced for Qwen2.5. These results show that compliance-critical APA requires architectural governance rather than prompt-level reliance on LLM discretion. The policy-derived floor guardrail enforces the floor-safety property and prevents silent under-escalation. The guarantee is local: it constrains the compliance routing decision under validated inputs and a trusted policy, and does not cover input corruption, policy errors, extraction failures, prompt injection, downstream tool misuse, or end-to-end workflow correctness. The evaluation is a single-process, two-model proof of concept, and generalization to other regulated workflows is a hypothesis for future validation.


  • Open Access

    Stability-Aware Uplift Policy Selection for Customer Retention: From Predictive Scores to Actionable Segments

    by Massimo Pacella, Gabriele Papadia, Vincenzo Giliberti

    Uplift modeling optimizes intervention-based campaigns by identifying customers whose behavior changes exclusively due to specific treatments, moving beyond standard baseline risk predictions. However, in real-world deployments, algorithms that maximize traditional causal ranking metrics (e.g., the Qini coefficient) often fail to be optimal in practice. The inherent variance of Conditional Average Treatment Effect (CATE) estimators exposes critical trade-offs between expected economic value, algorithmic stability, and policy interpretability. To address this gap, this study proposes a stability-aware, value-driven computational framework for selecting an uplift policy. The pipeline evaluates multiple causal and non-causal algorithmic families, including traditional baselines, multimodel approaches, and transformed-outcome variants, within a repeated-run validation protocol. Candidate policies are assessed primarily through incremental revenue and target-set stability, whereas a post hoc surrogate tree distillation step is used to translate the selected policy into interpretable rule-based customer segments. An empirical evaluation of the publicly available Telco Customer Churn dataset under two distinct regimes (a causally controlled semisynthetic scenario and an observational proxy scenario) reveals that the highest-yielding causal policy frequently suffers from severe targeting instability, inducing a clear risk–return trade-off. Furthermore, uplift models outperform traditional baselines in the causally controlled regime, whereas traditional baselines remain economically superior in the confounded proxy settings. Overall, this study establishes that jointly assessing economic utility, algorithmic stability, and transparent segmentation is essential for deploying robust and defensible causal machine learning in production environments.


  • Open Access

    An Assessment of Digitalization Techniques in Contact Centers and Their Impact on Agent Performance and Well-Being

    by Massimo Pacella, Paride Vasco, Gabriele Papadia, Vincenzo Giliberti

    The role of contact centers in improving the operational efficiency of numerous organizations is of utmost importance. Presently, digitalization technology has enabled contact centers to deliver exceptional customer service and support, while minimizing the adverse impact on agent well-being. Artificial intelligence techniques such as topic modeling and sentiment analysis can aid agents in addressing specific queries, providing real-time support and feedback, and helping them build stronger relationships with customers. This study aims to investigate the advantages of integrating these techniques in the analysis of customer–agent conversations within contact centers. This study examines whether there is a discernible advantage in analyzing customer–agent conversations in real-time and whether it is worth using this type of digitization to enhance agent performance and well-being. Furthermore, this study explores the impact of these technologies on European privacy, business, real-time agent support, the value of conversation data, brand reputation, and customer satisfaction. The results of this study demonstrate the significance of incorporating topic modeling and sentiment analysis into the analysis of customer–agent conversations at contact centers.


  • Open Access

    A Comparison of Different Topic Modeling Methods through a Real Case Study of Italian Customer Care

    by Gabriele Papadia, Massimo Pacella, Massimiliano Perrone, Vincenzo Giliberti

    The paper deals with the analysis of conversation transcriptions between customers and agents in a call center of a customer care service. The objective is to support the analysis of text transcription of human-to-human conversations, to obtain reports on customer problems and complaints, and on the way an agent has solved them. The aim is to provide customer care service with a high level of efficiency and user satisfaction. To this aim, topic modeling is considered since it facilitates insightful analysis from large documents and datasets, such as a summarization of the main topics and topic characteristics. This paper presents a performance comparison of four topic modeling algorithms: (i) Latent Dirichlet Allocation (LDA); (ii) Non-negative Matrix Factorization (NMF); (iii) Neural-ProdLDA (Neural LDA) and Contextualized Topic Models (CTM). The comparison study is based on a database containing real conversation transcriptions in Italian Natural Language. Experimental results and different topic evaluation metrics are analyzed in this paper to determine the most suitable model for the case study. The gained knowledge can be exploited by practitioners to identify the optimal strategy and to perform and evaluate topic modeling on Italian natural language transcriptions of human-to-human conversations. This work can be an asset for grounding applications of topic modeling and can be inspiring for similar case studies in the domain of customer care quality.


  • Open Access

    Topic Modeling for Automatic Analysis of Natural Language: A Case Study in an Italian Customer Support Center

    by Gabriele Papadia, Massimo Pacella, Vincenzo Giliberti

    This paper focuses on the automatic analysis of conversation transcriptions in the call center of a customer care service. The goal is to recognize topics related to problems and complaints discussed in several dialogues between customers and agents. Our study aims to implement a framework able to automatically cluster conversation transcriptions into cohesive and well-separated groups based on the content of the data. The framework can alleviate the analyst selecting proper values for the analysis and the clustering processes. To pursue this goal, we consider a probabilistic model based on the latent Dirichlet allocation, which associates transcriptions with a mixture of topics in different proportions. A case study consisting of transcriptions in the Italian natural language, and collected in a customer support center of an energy supplier, is considered in the paper. Performance comparison of different inference techniques is discussed using the case study. The experimental results demonstrate the approach’s efficacy in clustering Italian conversation transcriptions. It also results in a practical tool to simplify the analytic process and off-load the parameter tuning from the end-user. According to recent works in the literature, this paper may be valuable for introducing latent Dirichlet allocation approaches in topic modeling for the Italian natural language.