헬스케어 AI의 다음 시험대는 '통합'

Sponsored Provided by Ensemble The entrance of major AI companies into healthcare is a meaningful and welcome development, accelerating the technical foundation available to the industry. Their models are increasingly capable of processing long clinical records, interpreting complex terminology, comparing documentation against evidence and generating coherent summaries from large volumes of information. For clinicians, operators, and administrative teams who spend significant time searching through fragmented data, these advances are helping reduce cognitive burden and make high-value information easier to access. But healthcare leaders should not confuse model capability with operational capability. Healthcare’s administrative challenges are caused by fragmented information, fragmented workflows, and fragmented accountability, not a lack of information. The industry has spent decades investing in systems that capture activity: electronic health records, billing platforms, payer portals, scheduling systems, call center platforms, and analytics applications. Each system records something important. But few were designed to reason across the full chain of decisions that determines whether patients get timely access, clinicians have the right documentation and providers are reimbursed appropriately. This is the problem that AI must now confront. Revenue cycle is becoming one of healthcare AI’s proving grounds The revenue cycle is the process healthcare providers use to get paid for care — from scheduling and registration through coding, billing, payer follow-up, and payment collection. It is unusually suited to rigorous AI deployment because it combines high transaction volume, complex reasoning, structured and unstructured data, measurable outcomes, and significant operational variation. It also sits at the intersection of financial performance, patient access, and administrative workload. A single claim can be influenced by patient insurance information, clinical documentation, coding rules, payer-specific policies, prior authorization requirements, medical necessity criteria, and many other data sources and operational processes. A breakdown in any one of those areas can create downstream consequences weeks or months later. This is why generic automation has often fallen short. Traditional robotic process automation works well when workflows are stable and rules are predictable, but healthcare administration is neither. Payer requirements change. Documentation expectations evolve. Exceptions are common and often material. Large language models improve part of the equation, extracting meaning from narrative text, summarizing records and supporting reasoning over complex documentation. But when used alone, they inherit important limitations. They may produce plausible outputs without sufficient traceability. They may lack awareness of local workflow constraints. They may miss payer-specific history or context that determines whether an action is likely to change an outcome. Why foundation models will become necessary but insufficient The major AI firms are solving real technical problems for healthcare. Better context windows make it easier to process longitudinal records. Stronger reasoning improves the interpretation of complex clinical scenarios. Better multimodal capabilities may eventually help connect text, imaging, structured data, and clinical signals in more useful ways. Safer model behavior and healthcare-specific tuning will continue to improve adoption. These capabilities will make healthcare work faster, more consistent and easier to navigate. But they will not, on their own, solve deep-rooted administrative complexity. Much of healthcare’s operational knowledge does not live in general medical literature, coding manuals, or public payer guidance. It lives in the accumulated experience of what actually happens after decisions are made. For example: Why does one appeal strategy outperform another? Which documentation gaps are most likely to cause reimbursement delay? How does a specific payer respond to a particular clinical argument? These insights are behavioral, operational, and longitudinal. They emerge from years of transactions, outcomes, exceptions, and human judgment. As foundation models become more capable, access to baseline healthcare knowledge will become less differentiating. Most leading systems will be able to interpret ICD-10 codes, recognize medical terminology, summarize payer policies, and reason over public clinical criteria. The durable advantage will come from how organizations combine that model intelligence with proprietary operational data, structured knowledge, workflow context, and governance. The technical shift: From automation to orchestration Agentic orchestration turns foundation model understanding into coordinated action — intelligence that can follow work across systems, apply the right rules, adapt when something changes, and keep learning from what happens next. A prior authorization workflow, for example, may require retrieving clinical documentation through fast healthcare interoperability resources (FHIR) APIs, mapping patient history to payer criteria, identifying missing evidence, generating a submission packet, routing exceptions to a specialist, monitoring payer response, adjusting patient care pathways, and learning from the outcome. This type of workflow requires coordination. It also requires guardrails: regulatory requirements, privacy standards, clinical policies, coding rules, payer criteria, and organizational risk thresholds. One promising approach is hybrid architecture that combines LLMs with structured knowledge bases, symbolic logic, reinforcement learning, and deterministic validation layers. At Ensemble, this is the design principle behind EIQ, our revenue cycle intelligence engine. EIQ brings together operational activity, clinical documentation, payer behavior, and reimbursement outcomes into a continuously learning intelligence layer that’s integrated with the hospital’s electronic health record (EHR). It supplements the system of record with a system of intelligence, designed to connect information and surface actions most likely to improve outcomes. EIQ uses a neuro-symbolic approach that combines LLMs and custom small language models with rules-based reasoning. That architecture is built on one of the most robust datasets in healthcare, informed by more than a decade of award-winning operational performance, transaction history, payer behavior, and operator decision-making. The language models help interpret information and generate human-readable outputs. The symbolic layer represents policies, rules, payer requirements, and workflow constraints so the system can apply guardrails, make reasoning steps more traceable and recommend actions that fit the specific operational context. What the next decade will reward The contribution of major AI firms to healthcare will be significant. Their models will become faster, safer, more capable, and more accessible. But the next decade of healthcare AI will be defined by integration, not model capability alone. The organizations that create the most value will be those that connect models to governed data, operational workflows, domain expertise, human oversight, and measurable outcomes. They will understand that healthcare intelligence cannot live in a separate interface. It has to exist inside the decisions that shape access, documentation reimbursement, and patient experience. This content was produced by Ensemble. It was not written by MIT Technology Review’s editorial staff. Deep Dive Biotechnology and health A startup claims it’s found a drug to make your blood young Generation Lab claims its drug combo can “stop the spread of aging” around the body. And it’s looking for influencers to give it a try. By Antonio Regalado archive page Montana’s plan to become an experimental medical hub just pushed forward The state’s effort to exp