CoPlan: A Trustworthy Co-Intelligence Interface for... | AI Research

CoPlan is a human-AI interface designed to support care planning for aging-in-place. It addresses the limitation of many AI systems that provide fixed recommendations, which can conflict with clinical judgment, patient values, or real-world feasibility. By using a multi-agent workflow, CoPlan allows human care planners to inspect, challenge, and revise AI-generated care proposals, ensuring that human agency and clinical accountability remain central to the process.

A Collaborative Multi-Agent Workflow

The system functions through a multi-agent framework that mimics the interdisciplinary nature of healthcare. Initially, the system assesses patient data—including medical history and functional status—to recruit a team of specialized AI agents, such as nurses, pharmacists, and social workers. These agents generate candidate care interventions and provide both supporting and challenging arguments for each option. The system uses a Quantitative Bipolar Argumentation Framework (QBAF) to evaluate these arguments, assigning them strength scores based on clinical relevance, factual consistency, and reasoning transparency.

Enabling Human Contestation

CoPlan moves beyond simple explainability by providing an interface for "contestability." Instead of accepting AI outputs as final, human users can interact with an argument graph. This interface allows clinicians and caregivers to:

• Inspect the reasoning behind specific recommendations.

• Add their own arguments or counter-evidence.

• Modify or remove existing arguments generated by the AI.

• Review the distribution of support and challenges across different professional roles.
  Once the human reviewer has finalized the argumentative structure, the system uses this input to generate the final care plan.

Preserving Clinical Accountability

The design of CoPlan is built on the principle of co-intelligence, where AI agents and human stakeholders contribute complementary expertise. By requiring human validation of the argumentative graph, the system ensures that the final care plan is not just an algorithmic output, but a negotiated document that accounts for local workflows and patient-specific needs. This approach aims to maintain the necessary oversight for high-stakes healthcare decisions while leveraging the efficiency of multi-agent coordination.

Franklin Analysis

The evidence suggests that CoPlan is designed to solve the "black box" problem in clinical decision support by shifting the focus from automated recommendation to collaborative negotiation. By grounding the system in the interRAI Home Care assessment, the authors ensure that the initial data input is clinically validated. The use of QBAF provides a mathematical structure for handling conflicting viewpoints, which is a practical way to manage the inherent uncertainty in care planning. However, the effectiveness of this system relies heavily on the human user's ability to accurately evaluate and refine the AI-generated arguments, making the interface design a critical component of the system's overall trustworthiness.