Adrian Mak and Wilson Lui on Agentic AI and International Arbitration: A Three-Tier Taxonomy and Model Agentic AI Protocol (Journal of AI Law and Regulation)

“Agentic AI and International Arbitration: A Three-Tier Taxonomy and Model Agentic AI Protocol”
Adrian Mak (BSocSc(Govt&Laws)&LLB 2019; PCLL 2020) and Wilson Lui (Research Fellow, Centre for Private Law)
Journal of AI Law and Regulation, Volume 3, Issue 2, pp. 138–151
Published online: June 2026

Abstract: The rapid evolution from generative to agentic artificial intelligence (AI) presents both transformative opportunities and novel enforcement risks for international arbitration. Whereas ordinary prompt-response use of generative AI is largely reactive, agentic AI combines generative models with planning, persistent memory, tool use, and autonomous task execution, operating through multi-step workflows with minimal human oversight at each step. This article introduces a three-tier taxonomy: (1) Agent-Assisted, (2) Agent-Supported, and (3) Agent-Decided—to classify the deployment of agentic AI across the arbitral lifecycle, from conflict checks and procedural administration to deliberation and award drafting. It maps the principal enforcement risks under the United Nations Convention on the Recognition and Enforcement of Foreign Arbitral Awards 1958 (the New York Convention), particularly Articles V(1)(b), V(1)(d), and V(2)(b), and analyses the regulatory implications of the EU AI Act for high-risk AI systems deployed in dispute resolution. It situates the taxonomy against the principal existing AI-in-arbitration instruments—the SVAMC Guidelines and the CIArb Guideline—and argues that agentic AI requires a more granular, autonomy-based governance model than general AI guidance provides. To address the governance gap, the article proposes a Model Agentic AI Protocol structured around three pillars: (1) Disclosure, (2) Trajectory Logging, and (3) Human Certification—complemented by a technical toolkit comprising arbitration-specific benchmarks, mechanistic interpretability, adversarial robustness testing, and cognitive calibration techniques.