Healthcare AI’s next test is integration

Major AI companies are entering healthcare, bringing advanced tools that can process long clinical records, interpret complex medical terms, compare documentation against evidence, and summarize large amounts of information. These tools help clinicians, administrators, and operators by reducing the mental effort required to sift through fragmented data and making high-value information more accessible. For example, AI models can now handle longitudinal records (patient histories spanning years) and multimodal data (combining text, images, and structured data). However, the article notes that these technical improvements do not automatically solve healthcare’s deeper operational challenges, such as fragmented workflows and accountability.

The revenue cycle in healthcare refers to the process of getting paid for care, which includes scheduling, registration, coding, billing, payer follow-ups, and payment collection. This cycle is uniquely suited for AI because it involves high transaction volumes, complex reasoning, structured and unstructured data, measurable outcomes, and significant operational variation. A single claim (request for payment) can depend on multiple factors like patient insurance information, clinical documentation, coding rules, payer-specific policies, prior authorization requirements, and medical necessity criteria. A breakdown in any of these areas can lead to delays or denials weeks or months later. Traditional automation often fails in this environment because payer requirements, documentation expectations, and exceptions change frequently and unpredictably.

While large language models (LLMs) can extract meaning from text, summarize records, and support reasoning over complex documentation, they have important limitations when used alone. They may produce plausible but incorrect outputs without sufficient traceability (ability to explain how a decision was made). They may lack awareness of local workflow constraints (specific rules or processes unique to a hospital or system) or miss payer-specific history that determines whether an action will change an outcome. For instance, an LLM might generate a summary of a patient’s record but fail to recognize that a particular payer requires additional documentation for a specific procedure. These gaps highlight why generic AI models are insufficient for healthcare’s operational needs.

Foundation models are advanced AI systems trained on vast amounts of data that can perform a wide range of tasks. In healthcare, these models are becoming more capable, with better context windows (ability to process longer sequences of text), stronger reasoning, and safer behavior. However, even as these models improve, they will not solve healthcare’s administrative complexity on their own. Much of the operational knowledge in healthcare—such as why one appeal strategy works better than another or which documentation gaps cause reimbursement delays—is not found in medical literature or public payer guidelines. Instead, it comes from years of accumulated experience, transactions, and human judgment. This means that the real advantage will go to organizations that combine model intelligence with proprietary operational data, workflow context, and governance.

The next step in healthcare AI is agentic orchestration, which turns model understanding into coordinated action. This involves intelligence that can follow work across systems, apply the right rules, adapt to changes, and learn from outcomes. For example, a prior authorization workflow (process to get approval from a payer before a procedure) may require retrieving clinical documentation through FHIR APIs (Fast Healthcare Interoperability Resources, a standard for exchanging healthcare data), mapping patient history to payer criteria, generating a submission packet, routing exceptions to specialists, monitoring payer responses, and adjusting patient care pathways. This type of workflow requires coordination and guardrails, such as regulatory requirements, privacy standards, clinical policies, coding rules, and organizational risk thresholds. Hybrid architectures that combine LLMs with structured knowledge bases, symbolic logic, reinforcement learning, and validation layers are promising approaches to achieve this.

Some companies, like Ensemble, are using a neuro-symbolic approach to healthcare AI, which combines large language models (LLMs) with custom small language models and rules-based reasoning. This architecture is built on robust healthcare datasets informed by years of operational performance, transaction history, payer behavior, and operator decision-making. The language models interpret information and generate human-readable outputs, while the symbolic layer represents policies, rules, payer requirements, and workflow constraints. This allows the system to apply guardrails, make reasoning steps more traceable, and recommend actions that fit the specific operational context. For example, the system might use an LLM to summarize a patient’s record but rely on a symbolic layer to ensure the summary complies with payer-specific rules before generating a prior authorization request.