The hospitality industry increasingly deploys AI-driven personalization systems to optimize ancillary revenue through targeted upsell offers. The poorly designed implementations risk guest trust erosion, algorithmic bias, privacy violations, and regulatory non-compliance. With the European Union AI Act (EU AI Act) entering full enforcement in 2026, hospitality technology architects face mounting pressure to transition from revenue-first designs to responsible-by-design systems. Existing frameworks OECD AI Principles and NIST AI RMF provide high-level ethical guidance but lack sector-specific operationalization for customer-facing hospitality upsell architectures. Using a Design Science Research (DSR) methodology, this paper introduces TRAIN (Trust, Risk, Architecture, Integration, Network). TRAIN is a governance-centered system design framework for trustworthy AI deployment in hospitality upsell contexts. TRAIN helps verify that every system feature satisfies its corresponding regulatory obligation. The compliance mapping function f : S × R → G maps every deployed system feature to an explicit governance control satisfying each applicable EU AI Act obligation. System compliance is defined by the completeness condition ∀si ∈ S, ∀rj ∈ R : f (si, rj )̸ = ∅, verified by inspection of the feature-to-obligation traceability matrix. Three design rules govern the matrix, the No-Gap, the No-Regression and the No-Waste Rules. The paper contributes: (1) a structured compliance traceability architecture aligning system design with EU AI Act Articles 9, 11, 14, 50, and 61; (2) a prescriptive layer conflict resolution protocol for five inter-layer governance tensions; (3) a hospitality touchpoint-differentiated transparency specification; (4) a revised framework comparison against five reference frameworks including peer-reviewed AI governance operationalization literature; and (5) a catalog of 21 prohibited design anti-patterns specific to hospitality AI.

Large language models hold significant promise for enhancing decision-support in complex domains such as healthcare. However, persistent challenges, including factual hallucinations and opaque reasoning, undermine the trustworthiness of or human trust in AI recommendations. Incorporating appropriate domain-knowledge in such a high-risk domain can be seen as strengthening safeguards against such challenges. This study explores the extent to which integrating such domain knowledge through neuro-symbolic reasoning and knowledge graphs in a retrieval-augmented generation can improve the veracity and explainability of AI systems in healthcare for a diabetes healthcare advice case. The architecture consists of an offline neuro-symbolic training phase combining complex embeddings with logic constraints, and an online inference stage performing knowledge graph-based retrieval with logic tensor network-based fact validation. The proof of concept-implemented pipeline was evaluated against the Mistral Small and Gemma 3 27B baselines using TruthfulQA-derived questions and the DeepEval benchmark. The model demonstrated comparable or improved performance in domain-relevant evaluation aspects, while enhancing transparency, without compromising fluency or factuality. While improvements in veracity were modest, the system generated more traceable and auditable recommendations grounded in structured domain knowledge. Overall, the results provide evidence that integrating domain-knowledge grounding, such as in the proposed approach, offers a viable pathway towards trustworthy AI-enabled decision-making, through enhanced transparency and veracity in decision-support contexts. The proposed approach serves as a transferable methodology for developing aligned, auditable, and explainable AI systems across other high-stakes domains.

This paper presents the design and implementation of a Technology Radar framework developed in the context of the AI-DAPT project. In fast-evolving and interdisciplinary fields such as Artificial Intelligence and data-driven systems, systematic monitoring of technological developments is essential for informed decision-making in research and innovation projects. Technology radars are a recognized instrument for this purpose, but existing approaches are often heterogeneous in design and differ substantially in process logic, update mechanisms, and degree of methodological formalization.
To address this challenge, the paper applies a structured literature review, following a PRISMA-inspired procedure, to identify relevant Technology Radar approaches and adjacent contributions in technology intelligence and foresight. Based on this review, a set of framework requirements is derived for a project-oriented radar design that combines scientific and practice-oriented evidence, supports iterative updates, and integrates stakeholder knowledge.
On this basis, the paper introduces an AI-assisted and project-driven radar framework that combines scientific publications, industry-oriented web sources, automated keyword and filtering support, and interview-based expert validation. The resulting approach enables the structured identification and classification of technologies by combining signals of long-term research dynamics with short-term developments from industrial practice.
The paper further discusses the practical value, current limitations, and transfer potential of the framework. The analysis highlights that the structured inclusion of additional evidence sources—such as patents, standards, and international research and innovation agendas—remains an important extension path. While this has been conceptually considered, it has not yet been fully integrated into the automated pipeline. Overall, the contribution provides a structured and transferable framework for Technology Radar development, which is instantiated and evaluated in the context of the AI-DAPT project.% \cite{Figueiras2026Roadmaps}

Predictive Business Intelligence (BI) systems have become central to customer retention strategies in the telecommunications sector, enabling organizations to anticipate churn and optimize commercial interventions. However, the growing reliance on large-scale personal data processing and automated decision-making introduces significant ethical, legal, and governance challenges. Existing churn prediction approaches largely prioritize predictive accuracy and operational efficiency, often underintegrating transparency, accountability, fairness, and human oversight into system architecture. This paper analyzes the ethical and regulatory implications of predictive BI retention systems and identifies gaps between technical performance and governance structures. In response, we propose the Ethical-Operational Framework for Business Intelligence Retention Systems (EO-BI-RF), a layered model integrating data governance, model governance, decision governance, and ethical oversight. The framework operationalizes principles such as privacy by design, explainability, algorithmic fairness, accountability, and information security within predictive infrastructures. By embedding governance directly into BI system design, EO-BI-RF offers a structured approach to responsible analytics in telecommunications, balancing innovation, regulatory compliance, and organizational legitimacy in AI-driven retention systems.

Machine learning services often involve sensitive data and limited mutual trust: clients seek privacy guarantees, while providers protect proprietary models. Zero-knowledge proofs (ZKPs) allow a prover to convince a verifier that a claimed statement about data or computation is true without revealing underlying secrets. In machine learning, this can include correct training or inference, as well as properties of the training data or model. Unlike differential privacy (DP), secure multiparty computation (SMPC), and fully homomorphic encryption (FHE), which primarily aim to protect confidentiality during data release or computation, ZKPs are primarily aimed at verifiability. This paper reviews recent advances in zero-knowledge machine learning (zkML), including representative protocols, applications, and open challenges.