Agentic AI: Trustworthy Autonomous Decision Making Systems
Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 993-1001
DOI: 10.1609/aaaiss.v10i1.43518 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Kalpan Dharamshi.
Abstract (Originalsprache)
Large Language Models (LLMs) have opened up multiple avenues for innovation in the field of autonomous decision-making systems. The understanding of unstructured data and reasoning over it has allowed LLMs to power autonomous agents that can perform complex tasks. Agentic AI allows us to create a team of autonomous agents that can make complex business decisions to pursue a goal or an objective. Agents can use tools (functions) to receive inputs from the environment, reason over the retrieved data, and execute suitable actions with minimal or no human involvement. These capabilities, when embedded in business applications allow them to make efficient decisions to enhance their autonomy and adaptability to evolving business requirements. However, in high-stakes deployment, we need such autonomous agents to produce an explanation and a metric that can measure confidence in their decisions. A robust Agentic AI system considers quantifiable assurance as the central or core component of its design and implementation. Hence, we propose a modular Agentic AI framework that integrates assurance by defining a Trustworthiness Score (TWS)---a weighted, inference-time assurance metric composed of decision consistency, robustness, and explanation quality. The two distinct benefits of TWS include: system-level evidence for audit-compliance purposes, and an approach for involving the Human-in-the-Loop (HITL) for low-confidence decisions. In order to solidify the functioning of the modular framework, we present an Anti-Money Laundering (AML) transaction-classification experiment, which demonstrates its benefits over traditional manual or rule-based decision-making approaches. The experiment evaluation confirms our understanding that overall autonomy and adaptability of the system are enhanced, while the assurance metric TWS effectively routes low-confidence decisions for human intervention, thus providing an effective balance in Human-AI symbiosis and ensuring trust in the overall functioning of the system.
Zitieren
APA
Dharamshi, K. (2026). Agentic AI: Trustworthy Autonomous Decision Making Systems. Proceedings of the AAAI Symposium Series, 10(1), 993-1001. https://doi.org/10.1609/aaaiss.v10i1.43518
Vancouver
Dharamshi K. Agentic AI: Trustworthy Autonomous Decision Making Systems. Proceedings of the AAAI Symposium Series. 2026;10(1):993-1001. doi:10.1609/aaaiss.v10i1.43518
BibTeX
@article{Dharamshi2026,
author = {Dharamshi, Kalpan},
title = {Agentic AI: Trustworthy Autonomous Decision Making Systems},
journal = {Proceedings of the AAAI Symposium Series},
year = {2026},
volume = {10},
number = {1},
pages = {993--1001},
doi = {10.1609/aaaiss.v10i1.43518}
}
EUVIMED