The Evolution of AI in Healthcare Operations

Administrative costs in the healthcare industry total $265.6 billion per year, nearly 7% of all healthcare spending. AI programs like large language models (LLMs) excel at generating text or analyzing documents and have helped the industry realize efficiencies in documentation and communication, but LLMs can only take us so far.

The healthcare industry stands at the precipice of an AI revolution that extends far beyond chatbot interactions and prompts. Next-generation AI technologies—from retrieval-augmented generation to autonomous agents—offer truly transformative capabilities. These technologies support autonomous process management, real-time decision support, and coordinated workflow optimization, all of which are necessary to significantly reduce administrative burdens.

Understanding this evolution is essential for healthcare organizations seeking to leverage AI models to enhance operations while maintaining appropriate governance and oversight.

How AI Has Evolved

The rapid technological progression of AI is defined by four distinct evolutionary stages: foundational LLMs, context-aware retrieval-augmented generation (RAG) systems, action-capable tool use, and autonomous goal-driven agents, which are enhanced by Model Context Protocols (MCPs). Building upon the capabilities of prior stages, each stage introduces new possibilities for healthcare automation and optimization, from simple document summarization to complex self-managing workflows across domains.

Large Language Models (LLMs)

Large language models, or LLMs, established the foundation for AI adoption in healthcare through their natural language processing (NLP) capabilities. These systems excel at understanding and generating human language, making them valuable for documentation, communication, and analysis tasks that form the backbone of healthcare operations. LLMs have also proven effective for correspondence generation, policy analysis, and development of educational materials.

Despite the benefits, LLMs—particularly standalone LLMs—are not without their shortcomings. LLMs learn in a vacuum: they are trained on a predetermined pool of data, which often does not include the latest clinical guidelines or regulatory requirements. By the same token, LLMs cannot access current healthcare systems or databases, limiting their utility for real-time decision support. Most significantly, LLMs are limited to analysis and recommendation, rather than operational execution.

Healthcare processes require current information and system integration, and to achieve that, AI models had to evolve.

Retrieval-Augmented Generation (RAG)

Retrieval-augmented generation, first formalized by Facebook AI Research, addresses the static-knowledge limitation by connecting LLMs to current, authoritative information sources. Rather than relying solely on training data, RAG systems retrieve relevant, up-to-date information before generating responses in order to deliver more reliable and contextually appropriate outputs for healthcare applications.

The RAG process operates through three coordinated steps: query processing to understand information requirements, retrieval from current databases and knowledge sources, and generation that combines retrieved information with AI analysis to provide current, accurate, and contextually relevant information that standalone LLMs cannot.

RAG systems deliver measurable value through improved accuracy, reduced manual research time, and enhanced decision-making support. However, RAG systems—like LLMs—are informational rather than operational: they provide excellent analysis and recommendations but cannot execute actions within healthcare systems.

Tool Use / Function Calling

Tool use, also called function calling, is now implemented across major LLM providers including Anthropic’s Claude, Google’s Gemini, and others. Tool use represents a fundamental shift from information provision to action execution. While RAG-enhanced LLMs access current information, tool-enabled systems interact directly with healthcare applications, databases, and workflows to execute tasks autonomously.

Tool use enables AI systems to interact with external systems through defined interfaces. Rather than merely providing information, these systems can execute database queries, submit forms and applications, process transactions, update system records, and trigger workflows and alerts.

Tool-enabled systems can autonomously execute complex, multi-step healthcare processes that previously required significant manual effort and coordination, such as prior authorization. Traditional prior authorization requires multiple manual steps: clinical need identification, requirements research, form completion, submission, and status monitoring; tool-enabled systems allow for those steps to be performed autonomously.

Healthcare organizations implementing tool use typically integrate multiple categories of tools across different operational areas, including:

Clinical	Financial
Risk score calculators (HCC, CDPS, ACG)	Medical cost calculators
Clinical decision support	Medical loss ratio (MLR) tracking
Diagnostic coding assistants	Settlement processors

Administrative	Quality
Eligibility verification	HEDIS measure calculators
Prior authorization	Stars rating tracker
Claims processing	Performance scorecards

The next challenge emerges when healthcare workflows require coordination across multiple systems and processes simultaneously. Complex healthcare operations often involve parallel activities across clinical, administrative, and financial systems, creating the need for standardized coordination mechanisms enterprise-wide through Model Context Protocols.

AI Agents

AI agents, also known as agentic AI, represent the culmination of AI evolution to this point: autonomous systems that understand complex goals, plan multi-step approaches, execute workflows independently, and adapt based on outcomes. Unlike previous implementations which required human guidance at each step, AI agents can operate independently within defined governance parameters.

Healthcare AI agents combine goal understanding, strategic planning, independent execution, adaptive learning, and oversight maintenance. Agents are able to interpret high-level objectives like “optimize value-based care contract performance,” break complex processes into manageable sequences, take coordinated actions across multiple systems, improve performance based on outcomes, and operate within governance boundaries with appropriate checkpoints.

As an example, consider the complexity of a Medicare Advantage risk-sharing contract involving 50 primary care providers serving 15,000 members. Managing this contract requires continuous performance monitoring against cost, quality, and utilization metrics; real-time identification of intervention opportunities; provider performance analysis and improvement recommendations; financial tracking and settlement calculations; quality measure compliance and reporting; and risk adjustment with member attribution management.

A value-based care AI agent managing this contract might operate as follows:

Agentic AI systems are typically implemented as specialized agents working in coordination across multiple operational areas:

Clinical	Financial
Episodes of care coordination	Medical loss ratio (MLR) monitoring
Disease management	Risk adjustment
Care gap identification	Claims processing
Quality measure tracking	Settlement Calculations

Quality / Compliance	Member Engagement
HEDIS measures	Personalized outreach
Stars ratings	Care coordination
Regulatory compliance	Health education
Performance measurement	Experience optimization

The most sophisticated implementations involve agent orchestration, wherein a coordinating agent manages multiple specialized agents through task distribution, priority management, conflict resolution, human escalation protocols, and performance optimization across the system.

Healthcare AI agents operate within comprehensive governance frameworks that include operational boundaries with defined autonomous action scope, clinical safety controls with validation requirements, financial risk management with spending limits and approval thresholds, and regulatory compliance encompassing HIPAA requirements and audit processes.

Model Context Protocols: Supporting Agent Enablement

Agentic AI allows for autonomous execution of complex workflows; however, those workflows are rarely limited to a single system, which means coordination is required between claims processing, clinical data repositories, quality measurement platforms, financial settlement systems, and provider communication portals. Until recently, coordinating those systems required custom integrations to connect each individual system—but that’s changing with the introduction of Model Context Protocols (MCPs).

Introduced by Anthropic in 2024, MCPs establish standardized interfaces that allow AI systems to seamlessly manage interconnected workflows seamlessly. Rather than requiring custom integrations for each system combination, MCPs provide universal protocols for data access, action execution, communication, error handling, and audit compliance.

As an example, consider comprehensive episode of care management for a Medicare Advantage member undergoing total knee replacement surgery. Traditional management would require manual coordination across multiple departments and systems; however, MCP-enabled systems can coordinate all aspects simultaneously:

This example demonstrates an MCP’s capabilities for simultaneously managing complex healthcare processes spanning multiple operational domains. Well-structured MCPs are universal, allowing agents with disparate purposes to access data, services, and actions of the core system they represent. From a healthcare standpoint, MCP implementation is typically focused on clinical workflow coordination, administrative process integration, financial operations synchronization, and quality compliance orchestration.

The coordination infrastructure provided by MCPs creates the foundation necessary for autonomous agents to manage entire healthcare processes end-to-end.

The Path Forward

As the evolution from chatbots to autonomous healthcare agents demonstrates, the future of healthcare operations is intelligent, autonomous, and achievable, and the technology infrastructure required to get there exists today. Organizations that understand these technologies and implement them thoughtfully and systematically will gain significant competitive advantages in efficiency, quality, and member outcomes.

With appropriate guardrails to ensure, compliance, and continuous optimization, today’s healthcare leaders can help define the next generation of healthcare delivery excellence—and in our next article, we’ll discuss the keys to ensuring responsible and ethical AI implementation and use.