Ensemble has published an argument that the next decade of AI in healthcare will be decided by integration with operational processes rather than by model capability — and the argument arrives with a product attached to it. The piece carries a disclosure that it was produced by Ensemble and not by the MIT Technology Review editorial staff. Its central claim is that healthcare executives need to separate two things that marketing tends to merge: what a model can do, and what a system can do inside real operational workflows. The distinction is real, and it deserves more scrutiny than a sponsored slot usually invites.
The administrative difficulty in healthcare is a data problem shaped like an organisational one. Over decades the industry built systems that each capture one kind of activity — electronic health records, billing, payer portals, scheduling, call centres, analytics. Each holds something important. Few were designed to reason across the whole chain of decisions that determines whether a patient gets care on time, whether a clinician has the documentation they need, and whether the provider is paid correctly.
Ensemble picks the revenue cycle as the test case, which is a reasonable choice and also a self-interested one. The revenue cycle is the process by which a provider gets paid for treatment: patient scheduling and registration, coding, billing, payer interaction, collection. It combines high transaction volume, complex reasoning, structured and unstructured data, measurable outcomes and wide variation in workflow. It also sits at the intersection of financial performance, patient access and administrative workload — which makes it one of the few places in a hospital where an AI claim can actually be scored.
A single claim can be affected by the patient's insurance details, the clinical documentation, coding rules, a specific payer's policies, prior authorisation requirements, medical necessity criteria, and other data sources and operational processes. A failure anywhere in that chain can surface weeks or months later, by which point the cause is hard to reconstruct.
This is why generic automation tends to fall over. Traditional robotic process automation works when scenarios are stable and rules are predictable. In medical administration the conditions move: payer requirements are updated, documentation expectations shift, and exceptions are frequent enough to change the outcome rather than merely delay it.
Large language models improve individual steps. They extract meaning from narrative text, summarise records and help reason through complicated documentation. On their own they retain limitations that matter here. A model can produce a plausible result without enough explainability, ignore the constraints of a particular workflow, or miss the history of interaction with a specific payer — the very thing that often determines whether an action changes the result.
Ensemble's sharper point is about where the useful knowledge actually lives. Most operational knowledge in healthcare is not in the general medical literature, in coding manuals or in published payer guidance. It accumulates in the observed consequences of decisions already made: why one appeal strategy works better than another, which documentation gaps most often delay reimbursement, how a particular payer responds to a particular clinical argument. That knowledge is tied to participant behaviour, operational process and change over time, and it is built up through years of working with operations, outcomes, exceptions and staff decisions.
Follow that to its conclusion and the strategic claim is straightforward. As foundation models improve, access to core medical knowledge yields less and less advantage. Most frontier systems will be able to interpret ICD-10 codes, recognise medical terminology, summarise payer policies and reason over published clinical criteria. The durable advantage, in this telling, comes from combining model intelligence with proprietary operational data, structured knowledge, workflow context, and governance and control.
The proposed mechanism is orchestration rather than automation: agents that carry work across systems, apply the right rules, adapt to change and account for downstream results. Prior authorisation is the worked example, and it runs eight steps:
1. retrieve clinical documentation through FHIR APIs; 2. match the patient's history against the payer's criteria; 3. find missing evidence; 4. assemble the submission package; 5. route exceptions to a specialist; 6. track the payer's response; 7. adjust the patient's care pathway; 8. record the final outcome.
Each of those steps needs guardrails: regulatory requirements, privacy standards, clinical policies, coding rules, payer criteria and the organisation's own risk thresholds. Ensemble's answer is a hybrid architecture combining large language models with structured knowledge bases, symbolic logic, reinforcement learning and deterministic verification layers.
That architecture is what the company calls EIQ, its intelligence layer for revenue cycle management. EIQ brings operational actions, clinical documentation, payer behaviour and reimbursement outcomes into a continuously learning layer integrated with the hospital's electronic health record, supplementing the system of record rather than replacing it. The approach is neurosymbolic: large language models and specialised small language models paired with rule-based reasoning. The language models interpret information and produce human-readable output; the symbolic layer encodes policies, rules, payer requirements and workflow constraints, which is what lets the system apply guardrails, make reasoning steps more traceable and recommend actions in a specific operational context. Ensemble says the architecture is built on one of the most complete datasets in healthcare, covering more than ten years of award-winning operations, transaction history, payer behaviour and operator decisions.
Here is where the piece should be read as what it is. "Your proprietary operational data is the moat, not the model" is the argument every incumbent workflow vendor makes the moment a general-purpose model starts encroaching on its category, and it is being made here by a company whose entire asset is proprietary operational data. That does not make it wrong — the observation that payer behaviour is not in the training corpus is correct, and the neurosymbolic framing is a real engineering position rather than a slogan. But it is a conclusion the author was always going to reach, and the reader should weight it accordingly.
What is notably absent from the piece is any number. There is no accuracy figure, no denial-rate reduction, no days-in-accounts-receivable improvement, no claim throughput, no named health system, no before-and-after on a single prior authorisation queue. For an argument whose stated premise is that revenue cycle is attractive precisely because it produces measurable outcomes, the absence of a measured outcome is the loudest thing in the text. "More than ten years of award-winning operational work" is the closest thing to evidence offered, and awards are not a data quality metric. A company claiming that results are what separate real systems from capable models could settle the question by publishing one.
The deeper tension is that Ensemble's moat and Ensemble's thesis point in opposite directions over time. If the durable advantage really is accumulated payer behaviour, then every frontier lab with a healthcare customer is currently accumulating the same thing, and the neurosymbolic scaffolding built to compensate for today's models becomes overhead the moment models stop needing it. Intelligence in healthcare does have to be embedded in the decisions that govern patient access, documentation, reimbursement and experience rather than living in a separate interface — that part is right. The question is whether the company doing the embedding ends up owning the layer, or merely renting it out until the model vendors learn the workflow.