Scholarlink

<p>Enterprises increasingly deploy agentic AI: LLM-based agents plan tasks, use tools, coordinate in multi-agent workflows and intervene in business processes. As autonomy grows, demands for transparency, traceability and accountability increase—especially under the EU AI Act’s risk-based requirements. Research remains fragmented: responsible-AI and governance frameworks often stay on principles and processes, while technical work proposes isolated mechanisms (e.g., logging, observability, policy enforcement) without systematically assessing their fit with audit and accountability expectations. The article conducts a&#xa0;structured literature review following Webster and Watson (<CitationRef CitationID="CR39">2002</CitationRef>) and synthesizes 33&#xa0;peer-reviewed studies (2022–2026) in a&#xa0;concept-centric way. Results (i)&#xa0;conceptualize traceability, auditability and accountability as a&#xa0;linked evidence-and-responsibility chain, (ii)&#xa0;map mechanisms across architectural layers and lifecycle stages, and (iii)&#xa0;identify recurring mismatches: insufficient audit-grade evidence, limited stakeholder-specific explainability, and ambiguous responsibility allocation in multi-agent settings. Based on these insights, the article adopts a&#xa0;risk-based perspective that links traceability infrastructure, governance controls and accountability models to guide the design of auditable agentic AI in enterprise contexts.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Governance für KI-Agenten: Ein risikobasierte Perspektive für Transparenz und Nachvollziehbarkeit

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月30日

查看全文
  • Bennet Santelmann

摘要

Enterprises increasingly deploy agentic AI: LLM-based agents plan tasks, use tools, coordinate in multi-agent workflows and intervene in business processes. As autonomy grows, demands for transparency, traceability and accountability increase—especially under the EU AI Act’s risk-based requirements. Research remains fragmented: responsible-AI and governance frameworks often stay on principles and processes, while technical work proposes isolated mechanisms (e.g., logging, observability, policy enforcement) without systematically assessing their fit with audit and accountability expectations. The article conducts a structured literature review following Webster and Watson (2002) and synthesizes 33 peer-reviewed studies (2022–2026) in a concept-centric way. Results (i) conceptualize traceability, auditability and accountability as a linked evidence-and-responsibility chain, (ii) map mechanisms across architectural layers and lifecycle stages, and (iii) identify recurring mismatches: insufficient audit-grade evidence, limited stakeholder-specific explainability, and ambiguous responsibility allocation in multi-agent settings. Based on these insights, the article adopts a risk-based perspective that links traceability infrastructure, governance controls and accountability models to guide the design of auditable agentic AI in enterprise contexts.