AI Engineer World's Fair 2025
AI That Pays: Lessons from Revenue Cycle
Loading slides…
Read the talk
AI That Pays: Lessons from Revenue Cycle
Nathan Wan follows healthcare’s financial workflow from repeated claim denials to prior authorization and expert-reviewed appeals, showing where AI can reduce work and where integration still limits it.
From a talk by Nathan Wan
When delivering care does not lead to payment
A hospital can deliver care and still lose revenue in the process of getting paid. Delays, denied claims and repeated administrative work consume resources while payment remains unresolved. Nathan Wan opens with a stark figure: he says 40% of hospitals operate at a negative margin. He attributes that pressure to broken, manual revenue processes rather than clinical costs—a stronger causal interpretation than the figure alone establishes.
Wan introduces himself as head of AI at Ensemble Health Partners, which he describes as serving hundreds of US hospitals and health systems with approximately 14,000 employees. Its work spans the full revenue cycle. That scope matters because an organization that sees both the beginning and the end of a financial workflow can potentially identify a problem before it becomes a downstream failure.
Revenue cycle management, or RCM, is the financial process surrounding a patient’s journey through healthcare. It is traditionally represented as a sequence of steps, each handing work to the next. The opportunity for AI depends on looking beyond those individual handoffs and understanding how earlier decisions affect later outcomes.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From language modeling to healthcare administration
Wan’s path into this problem began at Google, building software for operational teams and then working on speech recognition and language modeling. More than a decade before this talk, his teams compared traditional language models trained with Google-scale data against deep learning models. One project explored what is now called ambient documentation: using technology to reduce the hours doctors spend writing notes after seeing patients. That early effort did not succeed, but he points to later commercial systems as evidence that the underlying opportunity persisted as the technology improved.
He then moved into biotech startups, first building models and teams for blood-based cancer detection. The mechanism was to search for biomarkers across datasets and patients that could provide an early signal of cancer. After that diagnostics company grew from 30 to more than 300 people, he joined a smaller therapeutics startup, using novel datasets of interactions within complex microbiome communities to identify potentially valuable drug-discovery compounds.
Diagnostics, imaging, documentation, clinical decision support and drug discovery are the familiar faces of healthcare AI. They are difficult problems with potentially enormous benefits. The financial side of healthcare receives less research and media attention, but it offers another kind of opportunity: administrative improvements whose operational effects can be measured directly.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The work surrounding a patient visit
Wan estimates healthcare at roughly 20% of US GDP, with administration accounting for a large share of the work. Billing and insurance are only part of that workload. It begins with eligibility checks and registration, continues through documentation and medical coding, and extends into denial management. Patients often encounter this machinery only when coverage is missing or care becomes complicated.
The process combines manual work, extensive rules and inconsistent requirements. To illustrate its growth, Wan says administrative employment increased thirtyfold over three decades while clinician employment barely doubled. The comparison is presented without an underlying workforce definition or dataset, so it is best read as his characterization of administrative expansion rather than a precise staffing benchmark.
Three groups participate in these exchanges:
| Participant | Role | Examples |
|---|---|---|
| Patients | Receive care | People seeking treatment |
| Providers | Deliver care | Hospitals, specialty offices, doctors and nurses |
| Payers | Fund care | Private insurers, Medicare and Medicaid |
These roles distinguish the clinical encounter from the financial negotiations around it: the organization delivering care is often not the organization deciding whether to pay for it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Four denials before payment
Much of the cost comes from friction: repeated communication among payers, providers and patients. Those exchanges can consume substantial effort without changing the eventual result—the claim is paid or it is not. Denials are especially expensive because providers must manage the rejection, assemble an appeal and work through the response. With slim margins, changes in denial rates can materially affect a hospital’s finances. Reducing this work could free resources for clinical care or other productive uses.
In Wan’s example, one claim was denied four times and appealed four times; the provider received payment 200 days after the procedure. Documentation had to be sent repeatedly through multiple interfaces, probably involving multiple manual processes. The timeline makes the problem tangible: one patient encounter can generate a prolonged sequence of administrative exchanges after the care itself is complete.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Avoid the error before accelerating the appeal
Providers are not the only organizations adopting AI. Wan describes payers using it to adjudicate more claims and identify issues that can trigger denials, alongside increasing denial rates and larger provider backlogs. Faster processing on one side can therefore create more work on the other.
A better appeal is not always the right intervention. Wan characterizes most of the denials under discussion as technical problems—registration errors or missing data—rather than disagreements about medical care. If the initial submission contains the correct information, the subsequent appeal may never be necessary. That shifts the engineering target from generating stronger responses to preventing avoidable failures.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
An appeal is an evidence-assembly problem
Prevention cannot eliminate every denial. A clinical denial occurs when the payer and provider disagree about what care was medically necessary. Appealing it requires more than a persuasive letter: the provider must build a packet that connects the patient’s circumstances to the relevant clinical and coverage criteria.
That work draws on three sources:
- Review the patient record to establish the clinical facts.
- Consult standards and guidelines of care to determine what treatment the situation called for.
- Examine payer policies to establish what would or would not be covered.
Some electronic medical records span hundreds of pages, mixing text, images, laboratory results, notes and tables. The reviewer must also select the relevant guidance from hundreds of clinical guidelines. All of this happens under payer response deadlines, making access to expert clinicians a practical limit on appeal throughput.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Generate the letter, preserve clinical judgment
An off-the-shelf generative model can produce an appeal letter when prompted. Ensemble found that this alone did not meet its clinical experts’ requirements. The team worked with those experts to develop a model and pipeline that could meet the organization’s quality standard while leaving the final decision with a clinician before submission to the payer.
That quality decision sits inside a workflow that must accommodate different service lines and clients. The task is therefore not simply to generate fluent prose: it is to produce an appeal appropriate to the clinical situation and operational context, then make it available for expert validation.
Wan reports a 40% reduction in time for Ensemble’s clinical appeal process, sometimes more. He also reports improved quality assessed through denial overturn rate and increased appeal volume. He gives no numerical overturn-rate improvement, and the time result comes without a specified cohort, baseline or measurement period.
The operational setting makes these outcomes measurable. An organization that performs the work can track processing time, appeal volume and whether denials are overturned, rather than treating the existence of a generated letter as proof of value. Wan emphasizes that Ensemble’s service teams can measure ROI directly within the revenue-cycle process.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The integration underneath the model
These workflows inherit long-established processes, inconsistent rules, unstructured information and data scattered across systems. AI does not make those constraints disappear. Before a model can reason over an encounter coherently, the organization must bring the relevant information together.
Wan identifies Ensemble’s EiQ platform as that integration foundation: a single infrastructure bringing multiple data formats together. Consolidation does not guarantee correct interpretation, however. Electronic medical records still contain varied formats that challenge multimodal LLMs to parse accurately. The platform addresses access and consistency; understanding the records remains a separate technical problem.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Close the loop across the revenue cycle
Wan describes existing value in clinical appeal letters and prior authorization, with agents continuing to be developed across the revenue cycle. But moving work through the same process faster is insufficient. Clicking buttons and advancing individual tasks cannot, by themselves, resolve the dependencies that make the process complex.
The larger opportunity is to turn downstream failures into upstream corrections. Reasoning connects a denial or appeal problem to its earlier cause; connectivity makes that information available where someone can prevent a recurrence. The goal is a coordinated system that reduces waste across the process, rather than merely accelerating each isolated step.
That requires more than a dataset. Wan closes on the combination of longitudinal records, a multidisciplinary team and responsibility for the full RCM process. Ensemble’s claimed advantage is the ability both to observe what happened and to intervene on clients’ behalf. The useful loop ends with action: a discovered error changes how the next request is prepared.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Ensemble’s 2024 account of EIQ, including data harmonization, EHR integration and collaboration between operational experts and engineers.
Further reading
- The Cost of Caring in 2025Article
The American Hospital Association’s April 2025 account of hospital costs, reimbursement pressures and operational challenges.
Updates since the talk
Current EIQ capabilities spanning denial prevention, appeal drafting, analytics and coordinated revenue-cycle workflows, with company-reported results.
Research examining administrator employment relative to physicians, with discussion of occupational definitions and data limitations.
Full transcript
Loading transcript…