The major AI firms are now entering healthcare with systems that can scan lengthy patient records, understand complicated medical terminology, and distill huge volumes of information. That technology exists. Still, healthcare executives should not mistake what a model can do for what an organization can actually put it to use.
Ensemble issues a warning in a sponsored analysis of where AI is heading next. The company claims that healthcare’s administrative problems stem from disorganized data rather than a dearth of it, citing scattered workflows, scattered responsibility, and a fragmented flow of information as the underlying causes.
The Fragmentation Problem
Decades of healthcare investment have centered on systems that capture activity, from electronic health records and billing platforms to payer portals, scheduling systems, call center platforms, and analytics applications. All of these tools record something important, but few were built with the full chain of decisions in mind. That chain determines whether patients receive timely access, clinicians hold the right documentation, and providers are reimbursed appropriately.
The outcome is an administrative workflow that is heavily divided. One claim alone can depend on patient insurance information, clinical records, coding guidelines, payer-specific rules, prior authorization needs, medical necessity standards, and numerous other data sources and operational systems. A failure in any single part of it can cause problems that show up weeks or months down the line.
This is the problem that AI must now confront.
The Revenue Cycle Test
Healthcare providers depend on the revenue cycle to receive payment for their services, a chain that begins with scheduling and registration and moves through coding, billing, payer follow-up, and payment collection. Ensemble points to this cycle as one of healthcare AI’s proving grounds.
Rigorous AI deployment finds unusual suitability in the revenue cycle because it brings together several distinct advantages: a large number of transactions, reasoning that calls for complex judgment, both structured and unstructured data, quantifiable results, and considerable variation in how operations are run. The revenue cycle also occupies a point where financial performance, patient access, and administrative workload intersect.
“Healthcare leaders should not confuse model capability with operational capability.”
This is why generic automation has frequently come up short. Robotic process automation built for traditional use performs well when workflows stay stable and rules remain predictable, yet healthcare administration is anything but. Insurer requirements shift. Documentation standards change over time. Exceptions occur regularly and can be significant.
Limits of Large Language Models
These models excel at interpreting written accounts, condensing documents, and assisting with logical analysis across extensive files. Yet deployed on their own, they carry significant weaknesses.
These systems can generate believable results while failing to provide adequate documentation of their decision-making process. They might also overlook local operational procedures. Furthermore, they could ignore payer-specific history or context that determines whether an action is likely to alter an outcome.
Healthcare is where the major AI firms are solving real technical problems. Improved context windows ease handling of longitudinal records. Enhanced reasoning aids the interpretation of complex clinical scenarios. More capable multimodal systems may eventually link text, imaging, structured data, and clinical signals in more useful ways. Safer model behavior and healthcare-specific tuning will keep improving adoption.
These tools will speed up healthcare, make it more consistent, and simplify navigation through it. They will not, however, fix deep-rooted administrative complexity by themselves.
Where Operational Knowledge Lives
What truly determines how hospitals run is rarely found in standard medical texts, coding guides, or public instructions from insurers. Instead, it comes from the practical experience that builds up over time once choices are put into practice.
Three queries capture the gap Ensemble describes. What makes one appeal strategy outperform another? Where do documentation omissions most commonly lead to delayed reimbursement? How does a particular payer react to a certain clinical argument?
The knowledge comes from actions taken over time, operations carried out, individual decisions made, and results observed. It covers how people act, how systems run, and how things develop across a span of years.
As foundation models grow more capable, the gap in baseline healthcare knowledge between leading systems will narrow. Most top-tier systems will be able to interpret ICD-10 codes, recognize medical terminology, summarize payer policies, and reason over public clinical criteria. What will set one organization apart from another is how it fuses model intelligence with its own proprietary operational data, structured knowledge, workflow context, and governance.
Orchestration Over Automation
Ensemble’s technical shift involves moving from automation to orchestration. With agentic orchestration, foundation model understanding is turned into coordinated action — intelligence capable of tracking work across systems, applying the correct rules, adapting when circumstances change, and continuing to learn from outcomes.
A prior authorization workflow can involve fetching clinical records via fast healthcare interoperability resources (FHIR) APIs, matching patient history against payer standards, pinpointing any missing proof, assembling a submission package, forwarding special cases to an expert, watching payer feedback, altering patient care paths as needed, and drawing lessons from the result.
Coordinating this kind of workflow demands a set of controls, including regulatory requirements, privacy standards, clinical policies, coding rules, payer criteria, and organizational risk thresholds.
A hybrid architecture that joins large language models with structured knowledge bases, symbolic logic, reinforcement learning, and deterministic validation layers is highlighted by Ensemble as one promising direction.
The Integration Test
It’s a significant and welcome development that major AI companies have entered healthcare, since it accelerates the technical foundation available to the field. These firms’ models are becoming more able to process lengthy clinical records, interpret complex medical language, compare documentation against evidence, and produce coherent summaries from large volumes of data.
These developments are assisting clinicians, operators, and administrative teams who spend considerable time hunting through scattered data by lowering the mental strain involved and making important details simpler to find.
Now comes integration testing. For years, the field has relied on isolated systems to record activity. What remains unclear is whether AI can draw conclusions spanning the entire sequence of choices that decides if patients receive prompt care, doctors hold the correct records, and medical staff receive proper payment.
Combining model intelligence with proprietary operational data, structured knowledge, workflow context, and governance is a more difficult challenge than building a better model.
Foundation Models: Necessary but Insufficient
The main case made by Ensemble holds that foundation models will grow into required tools for healthcare administration, yet they will remain inadequate on their own. Those models will continue to improve. Still, lasting superiority will depend on more than just model capability.
The source of this shift will be orchestration. It will rely on systems capable of tracking work as it moves across scattered platforms, applying the correct rules at the correct moment, adjusting when payer requirements change, and learning from results.
This happens within the revenue cycle, which joins large amounts of transactions with difficult thinking, clear results, and wide differences in how work gets done.
The Road Ahead
It is not a matter of picking one over the other between improved models and improved workflows. The way ahead involves constructing systems that combine both — models capable of grasping clinical and administrative language, encased within orchestration layers that understand the real workings of healthcare.
The source’s ideas can be organized into three levels of skill, not as a fixed line. Robotic process automation that follows set rules manages steady work flows. Large language models step in to read through files and sum up what they contain. Agentic orchestration then brings together actions across different systems.
| Capability | Focus | Key Technology |
|---|---|---|
| Automation | Stable workflows, predictable rules | Robotic process automation |
| Model capability | Processing records, summarizing data | Large language models |
| Orchestration | Coordinated action across systems | Agentic workflows, hybrid architecture |
The businesses that prosper will be those that merge foundation model intelligence with the practical experience of what happens after choices are put into action. That operational knowledge — based on behavior over time and difficult to express as rules — is what gives lasting benefit.
The workflow gap looms as healthcare AI’s next big hurdle. The models themselves are already here. But the systems around them still lag behind.
Source: technologyreview.com
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.