Scholarlink

Autonomous AI agents are increasingly being adopted and integrated into organizational processes, with the trajectory of this adoption pointing toward agents that independently execute business decisions with minimal human intervention. This study introduces the Verified Adaptive Learning in Deployment (VALID) framework, a five-layer governance architecture designed to support the transition from human-in-the-loop oversight to human-on-the-loop supervision in autonomous contract execution contexts. Using Elaborated Action Design Research (eADR), the framework was developed and iteratively refined through three design cycles, integrating principles from AI governance, organizational trust theory, and agentic system design. A multi-AI agent simulation evaluated VALID against a simulated human baseline incorporating modeled cognitive biases, including fatigue, loss aversion, overconfidence, and anchoring effects, across 150 matched runs. Agents operating under VALID demonstrated consistent trust growth from a neutral baseline of 0.50 to a mean of 0.805, stable decision quality, reduced decision-avoidance behavior, and resilient trust recovery following shock events. The results demonstrate that autonomous AI agents operating within a governed, self-learning framework can systematically build, sustain, and recover interorganizational trust through task performance and compliance calibrations across the Formation-Shock-Repair lifecycle [9, 11]. A critical finding of this research is that B2B AI agents acting autonomously within this framework have demonstrated the ability to build trust between AI agents completing the contract lifecycle. These agents demonstrated the ability to execute the entire contract lifecycle of high-stakes contracts with the potential to outperform their human counterparts. Furthermore, AI agents that reach peak trust demonstrate the capacity to accept commitments beyond organizational de-livery limits, a finding that reveals trust calibration alone is insufficient for sustainable autonomous operations. This finding further underscores that human-on-the-loop supervision remains a critical safeguard, ensuring that humans retain the ability to monitor, intervene, and course-correct when autonomous agents approach or exceed operational boundaries. This study contributes preliminary design principles for interorganizational AI governance architecture, a simulation-based methodology for stress-testing governance frameworks prior to deployment, and evidence that AI agents build and repair trust through task performance and compliance rather than the social and relational mechanisms characteristic of human trust formation.

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

From Human-in-the-Loop to Human-on-the-Loop: VALID Framework for Resilient Trust in Autonomous AI Agent Contract Execution

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jonathan Ocasio,
  • Matthew Mullarkey

摘要

Autonomous AI agents are increasingly being adopted and integrated into organizational processes, with the trajectory of this adoption pointing toward agents that independently execute business decisions with minimal human intervention. This study introduces the Verified Adaptive Learning in Deployment (VALID) framework, a five-layer governance architecture designed to support the transition from human-in-the-loop oversight to human-on-the-loop supervision in autonomous contract execution contexts. Using Elaborated Action Design Research (eADR), the framework was developed and iteratively refined through three design cycles, integrating principles from AI governance, organizational trust theory, and agentic system design. A multi-AI agent simulation evaluated VALID against a simulated human baseline incorporating modeled cognitive biases, including fatigue, loss aversion, overconfidence, and anchoring effects, across 150 matched runs. Agents operating under VALID demonstrated consistent trust growth from a neutral baseline of 0.50 to a mean of 0.805, stable decision quality, reduced decision-avoidance behavior, and resilient trust recovery following shock events. The results demonstrate that autonomous AI agents operating within a governed, self-learning framework can systematically build, sustain, and recover interorganizational trust through task performance and compliance calibrations across the Formation-Shock-Repair lifecycle [9, 11]. A critical finding of this research is that B2B AI agents acting autonomously within this framework have demonstrated the ability to build trust between AI agents completing the contract lifecycle. These agents demonstrated the ability to execute the entire contract lifecycle of high-stakes contracts with the potential to outperform their human counterparts. Furthermore, AI agents that reach peak trust demonstrate the capacity to accept commitments beyond organizational de-livery limits, a finding that reveals trust calibration alone is insufficient for sustainable autonomous operations. This finding further underscores that human-on-the-loop supervision remains a critical safeguard, ensuring that humans retain the ability to monitor, intervene, and course-correct when autonomous agents approach or exceed operational boundaries. This study contributes preliminary design principles for interorganizational AI governance architecture, a simulation-based methodology for stress-testing governance frameworks prior to deployment, and evidence that AI agents build and repair trust through task performance and compliance rather than the social and relational mechanisms characteristic of human trust formation.