The Taxonomy of Autonomy: Mapping the evolution of AI agents in healthcare

The healthcare industry stands at a pivotal moment where AI agents are transforming from simple assistive tools to sophisticated autonomous partners in clinical care. As these systems grow in capability and complexity, researchers and regulators across the US and Europe have developed multiple classification frameworks to categorize them, each with its own perspective and purpose. Understanding these taxonomies isn't just an academic exercise; it's essential for healthcare leaders to effectively implement, regulate, and scale AI solutions that can truly enhance patient care while mitigating risks.

This is the first post in a series exploring the classification, applications, implementation strategies, and key takeaways of AI agents in healthcare.

The need for a unified understanding of AI agent classifications

Healthcare AI is evolving at breakneck speed, with over 880 AI/ML-enabled medical devices receiving FDA approval – more than 400 of those just since 2022. This explosion of innovation has led to a fragmented landscape of classification systems, each developed to address specific aspects of AI implementation. After analyzing the leading frameworks developed between 2023-2025, we've identified four distinct but overlapping classification approaches that together provide a comprehensive view of AI's role in modern healthcare.

The most significant finding is that these classifications measure fundamentally different aspects of AI systems: autonomy level, clinical integration, functional purpose, and risk profile. By understanding how these dimensions intersect, healthcare organizations can make more informed decisions about AI adoption and governance.

Key classification frameworks: Different lenses, common purpose

Autonomy spectrum: From assistive to fully autonomous

The American Medical Association's CPT Editorial Panel has established perhaps the most clinically-oriented classification system, dividing AI medical services into three increasingly autonomous categories:

• Assistive AI: Systems that 'detect clinically relevant data without analysis or generated conclusions' requiring physician interpretation
• Augmentative AI: Systems that 'analyze and/or quantify data to yield clinically meaningful output' still requiring physician interpretation
• Autonomous AI: Systems that 'automatically interpret data and independently generate clinically meaningful conclusions'

Autonomous AI is further divided into three levels of increasing independence:

• Level I: AI draws conclusions but requires physician action to implement
• Level II: AI draws conclusions and initiates actions with opportunity for override
• Level III: AI draws conclusions and initiates actions automatically, requiring physician initiative to contest

This classification directly informs medical billing codes, highlighting how AI autonomy impacts clinical workflows and reimbursement.

Integration depth: From foundation to pioneer

Researchers at Icahn School of Medicine at Mount Sinai have developed a framework that classifies AI agents based on their degree of clinical integration and autonomy:

• Foundation Agents: Basic data processing tools with minimal autonomy
• Assistant Agents: Support clinical workflows but require significant oversight
• Partner Agents: Collaborate actively with healthcare professionals
• Pioneer Agents: Operate with high autonomy in specific clinical domains

This framework offers a more nuanced view of how AI systems integrate into clinical environments. For instance, the Oracle Health Clinical AI Agent deployed across 70+ healthcare organizations represents a Partner Agent – it captures patient exchanges, proposes follow-ups, and generates discharge summaries, but still works collaboratively with providers. At AtlantiCare, this system reduced documentation time by 41%, saving providers approximately 66 minutes daily.

Functional purpose: Task-oriented classification

Beyond autonomy and integration, AI agents can be classified by their primary functions in healthcare:

• Information Processing Agents: Extract and analyze EMR data
• Decision Support Agents: Provide diagnostic and treatment recommendations
• Workflow Automation Agents: Streamline clinical and administrative processes
• Patient Communication Agents: Interact with patients for monitoring and education
• Clinical Documentation Agents: Generate and manage medical documentation

This functional approach helps healthcare organizations identify specific use cases for AI implementation. For example, Qventus's Inpatient Solution acts as a Workflow Automation Agent for discharge planning, saving OhioHealth nearly 1,400 excess days and approximately $550,000 within its first month of deployment.

Risk-based regulatory frameworks: US and EU approaches

Regulatory bodies have developed their own classifications focused primarily on risk mitigation:

• US FDA: Classifies AI medical devices into Class I (low risk), Class II (moderate risk), and Class III (high risk), with 96.5% of AI/ML approvals falling into Class II
• EU AI Act: Categorizes systems as Unacceptable Risk (banned), High Risk (strict requirements), Limited Risk (transparency obligations), or Minimal Risk (few obligations)

These regulatory frameworks affect how AI agents can be deployed in clinical settings. The EU's approach is notably more stringent, with most healthcare AI applications (approximately 75%) classified as High Risk, requiring extensive documentation, validation, and human oversight.

The convergence: Correlating classification systems

The true value emerges when we map these classification systems against each other to create a unified understanding:

This correlation reveals important insights. For example, as systems move from assistive to autonomous, they not only increase in capability but also face heightened regulatory scrutiny. Similarly, Pioneer agents with deep clinical integration almost always fall into high-risk regulatory categories, requiring rigorous validation.

The MASH framework: The future of coordinated healthcare AI

Perhaps the most forward-looking classification comes from the Multi-Agent Systems for Healthcare (MASH) framework. MASH envisions 'decentralized yet coordinated networks of specialized artificial intelligence agents' working together across healthcare domains.

The MASH approach recognizes that the future of healthcare AI isn't a single agent operating in isolation but rather an ecosystem of specialized agents collaborating on patient care. This represents a paradigm shift from classifying individual agents to understanding how they function as integrated systems, similar to how healthcare professionals work as coordinated teams.

Strategic implications for healthcare leaders

Healthcare organizations implementing AI agents should consider all four classification dimensions – autonomy, integration, function, and risk – when evaluating solutions. A comprehensive assessment should include:

1. Autonomy assessment: What level of independent action is appropriate for this clinical context?
2. Integration evaluation: How deeply should the AI agent be embedded in clinical workflows?
3. Functional alignment: Which specific tasks should the agent perform?
4. Risk analysis: What regulatory requirements will apply based on autonomy and function?

Organizations must also consider how these classifications will evolve. A system that begins as an Assistant agent may gradually develop Partner capabilities as it learns from clinical data and user interactions. Planning for this evolution, including the regulatory implications, is essential for sustainable AI implementation.

Conclusion: Beyond classification to transformation

The diverse classification systems for healthcare AI agents each illuminate different aspects of these complex technologies. By understanding how autonomy, integration, function, and risk intersect, healthcare leaders can make more informed decisions about AI implementation.

As we look toward the future of healthcare AI, the MASH framework's vision of coordinated agent ecosystems suggests that classification itself will evolve. The focus will shift from categorizing individual agents to understanding how networks of specialized AI systems can work together – and with human providers – to deliver more effective, efficient, and personalized care.

The most successful healthcare organizations will be those that look beyond classification to transformation, using these frameworks not just to categorize AI agents but to strategically deploy them as part of a broader evolution in how healthcare is delivered. In this future, the taxonomy of autonomy becomes not just a way to describe AI agents but a roadmap for healthcare transformation.