Title:

Dependable Agentic AI in Regulated Environments: From Model Accuracy to System Accountability

Keynote speaker:

Nikolaos Korfiatis, FRSA, FHEA, Professor of Data Management at the Department of Informatics, Ionian University, Corfu

Nikolaos Korfiatis, FRSA, FHEA, is Professor of Data Management at the Department of Informatics, Ionian University, Corfu, which he joined in 2024. He previously held the Chair in Business Analytics and Artificial Intelligence at Nottingham University Business School, University of Nottingham. Before that he was Assistant and then Associate Professor of Business Analytics at Norwich Business School, University of East Anglia, where he remains a visitor, and Visiting Associate Professor at Warwick Business School. Earlier he was Assistant Professor at the University of Copenhagen (2009 to 2011) and directed the Big Data Lab at Goethe University Frankfurt (2011 to 2014). He holds a PhD from Copenhagen Business School, an MSc from the Royal Institute of Technology (KTH), Stockholm, and a BSc from the Athens University of Economics and Business. He has published more than 100 papers in data science and information systems, with more than 5,000 citations to his work, and has supervised more than 10 doctoral theses. Alongside his academic work he serves as an AI architect for enterprise AI platforms in manufacturing and a raporteur for OECD.

Abstract:
Large language models are moving from conversational assistants to agents that retrieve data, call tools and act inside enterprise workflows. In these settings the dependability question changes. A model that is accurate in a benchmark can still produce a system that is unsafe, unauditable or impossible to operate at scale. This talk draws on the deployment of agentic and retrieval-augmented AI platforms in manufacturing, where an error carries an operational and a regulatory cost. It argues that dependability in agentic AI is a property of the surrounding system, not of the model, and examines the engineering controls that deliver it, and more specifically identity and least-privilege access for agents, guardrail policies enforced in the deployment pipeline, evaluation against domain ground truth rather than generic benchmarks, traceability of every retrieval and action, and a defined point where the human decides. The talk closes with open research problems on measuring trust and explainability in agentic workflows, and with what public institutions and smaller organisations can adopt from these practices today.

Keywords: agentic AI, retrieval-augmented generation, dependability, AI governance, guardrails, human oversight, explainability, manufacturing