This exploratory, quasi-longitudinal study examines how accountability is allocated and prioritized in artificial intelligence (AI) projects operating under competing regulatory, financial, and societal objectives. The purpose of the study is to empirically test whether role-based governance structures reliably organize accountability in AI projects or whether system capabilities and delivery pressures shape accountability in practice. Using multi-wave survey data from experienced AI project participants collected between 2023 and 2025, the study evaluates accountability assignments across project roles and AI system scopes. Grounded in agency theory, it operationalizes an agent–principal accountability matrix by translating normative AI stakeholder expectations into 12 binary accountability indicators. AI system scope is modeled using both a binary autonomy classification and three capability-based dimensions: Generative/Agentic AI, Autonomous Algorithms, and Integrated Algorithms. In the 2025 wave, the study assesses self-reported project outcomes, including financial benefits, sustainability outcomes, societal impacts, and overall project success. The findings show that accountability in AI projects is shaped less by formal role definitions and more by how organizations prioritize competing objectives under different system scopes. Regulatory compliance, financial performance, and societal impacts are not mutually exclusive objectives; rather, they compete for attention under delivery pressure, leading to selective erosion of accountability. Strengthening accountability, therefore, requires explicit management of trade-offs and coordination mechanisms that align delivery incentives with societal and regulatory responsibilities. By empirically linking normative accountability theory with observed project practice, the study contributes a capability-sensitive perspective on AI governance. It proposes an evidence-based governance heuristic to support accountability management in complex AI initiatives. The results should be interpreted in light of limitations, including reliance on self-reported measures and cross-sectional outcome data.
The proliferation of heterogeneous sensing platforms in agriculture (UAVs, robots, IoT) makes transforming raw observations into interoperable representations a major engineering bottleneck. This paper reframes that challenge by treating data transformation as a first-class architectural concern and introduces an agentic approach for producing NGSI-LD entities which are syntactically valid, conform to the target JSON schema, and thus ready for ingestion into a context broker. We present an agentic architecture that combines three complementary design patterns – Coordinator, Production Line and Critic – and a proof-of-concept implementation in which specialized, task-scoped agents (including LLM-backed converters) perform extraction, semantic mapping and deterministic validation/repair. Using the critic (quality assurance) agent as an oracle, we experimentally evaluate eleven open-source LLM variants across multiple families and parameter counts, measuring syntactic correctness, retry behavior, latency, GPU energy and peak memory. Results show that the ability to produce ingestion-ready NGSI-LD output is architecture- and model-dependent rather than monotonic in parameter count: several mid-sized models achieve near-perfect correctness with far lower resource costs than the largest models, while some smaller or similarly sized models fail systematically. The main novelty of this work is threefold and explicitly practical: (i) an agentic pipeline architecture for structured transformation, (ii) a practical instantiation of LLM-powered agents with a deterministic validation/repair boundary that separates probabilistic generation from correctness guarantees, and (iii) a resource-aware methodology for selecting and deploying LLMs. These findings offer practitioners a methodology for balancing correctness, cost and deployment constraints. Limitations include the current focus on syntactic/schema compliance and a primary evaluation on imagery-derived metadata.
Accurately anticipating family consent to organ donation is essential for optimizing donor management, particularly in opt-in systems where delays prevent the timely execution of cost-intensive clinical assessments. In this study, we pursue a twofold objective: first, to model and identify the most important factors influencing the decision of relatives or next of kin to grant or deny permission for organ retrieval from a deceased individual; and second, to develop, based on these data, a tool for the early prediction of the outcome of this decision. To this end, we develop a data-driven framework to predict consent outcomes using real-world data collected from a regional transplant center operating within the Italian transplant system. We conduct a comprehensive exploratory analysis—including chi-square tests, Cramer’s V, and ordinal correlation measures—to identify statistically significant drivers of consent. A structured feature-reduction process is followed by the training and evaluation of several machine learning classifiers under data scarcity and class imbalance conditions. Our findings show that ensemble and non-linear models, particularly AdaBoost and neural networks, consistently outperform simpler approaches, with further performance gains achieved when family-related features are included. We also observe notable heterogeneity across nationality groups, with non-Italian donors exhibiting clearer predictive patterns. Even though our approach is based on a single-center dataset (with potential context-specific effects), the methodological framework is readily transferable where comparable data are available. Overall, the results demonstrate that predictive modeling can provide early, reliable estimates of consent likelihood. A primary implication is that clinicians are enabled to initiate preparatory activities sooner, hence reducing avoidable costs while enhancing the efficiency of organ procurement workflows.
Intermediary roles such as the IT Business Partner (ITBP) and Domain Architect (DA) are increasingly introduced to strengthen business–IT collaboration and strategic alignment. However, their specific objectives, task profiles, and required competencies remain underexplored. This study analyzes and systematically defines the roles of ITBPs and DAs within the business–IT collaboration process. Based on two case studies and nine expert interviews, we identify eleven key tasks, centered on relationship management and advice and guidance, and fourteen critical skills required for effectiveness. We further examine structural and operational challenges, including organizational integration, skill gaps, and limited strategic focus. A comparison highlights both overlaps and distinct responsibilities between the two roles. By clarifying role design and positioning, this study contributes to research on business–IT alignment. It also provides practical guidance for organizations seeking to structure, implement, and develop intermediary roles effectively and foster skills development of ITBPs and DAs.
The aim of this paper was to address the changes taking place in knowledge management processes in terms of multi-agent system (MAS) capabilities. The article, based on a literature review, addresses knowledge management with concepts of digital twin, software agent, and machine learning for creation of smart environments. The research study utilize AI agents, artificial intelligence technologies, and IoT-based data management solutions by supporting the design and simulation of device behaviors. This article aimed to conduct research addressing implementation of artificial intelligence in knowledge management with the following two research questions: RQ1: How does multi-agent software impact (RQ1.1) smart grid security, (RQ1.2) smart grid adaptability, (RQ1.3) smart grid user-friendliness; and RQ2: How to implement the multi-agent systems in a smart environment with current AI software? Additionally the paper’s research findings discussed the AI-based knowledge management model in the context of the GRAI framework usage, which stands for Generative, Receptive Artificial Intelligence, and presents a proposal of an updated model that takes into account recent developments regarding GenAI and its consequences for knowledge management (KM). The research methodology is focused on Web od Science analysis in the three areas (RQ1.1, RQ1.2, and RQ1.3) and the proposed agent-based algorithm (RQ2), with a JADE-based simulation with Isolation Forest, and grounding in the GRAI framework. Based on the conducted research, an algorithm for collecting information from devices was proposed and used in a simulation based on the Isolation Forest algorithm. The paper presents the example of implementation with some limitation concerning the lack od data set for literature analyzes after 2023, and conclusion for future research in context of an architecture of GRAI.
