# AI That Pays: Lessons from Revenue Cycle Nathan Wan, Ensemble Health | AI Engineer World's Fair 2025 | 18:19 Source: https://www.youtube.com/watch?v=TquUsN1QsWs Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/ai-that-pays-lessons-from-revenue-cycle Published: 2025-07-24 Tags: enterprise, healthcare, human-in-the-loop, rag ## TL;DR - Revenue cycle management is a large, manual source of healthcare cost and lost revenue, with denials often caused by preventable technical errors. - Ensemble uses longitudinal data across the full revenue cycle to predict denials, correct problems before they happen, and automate parts of the process. - Generative AI can speed up clinical denial appeals, but clinical experts still need to review and approve the final letters. ## Summary Nathan Wan argues that healthcare AI should include the financial work that determines whether providers get paid. Revenue cycle management covers eligibility, registration, documentation, coding, prior authorization, billing, and denial management. The work is manual, rules-driven, inconsistent, and spread across many systems. Wan describes how Ensemble uses its end-to-end view of the process to connect early events with later denials. In prior authorization, this can mean flagging missing procedures or correcting documentation before a claim is rejected. For clinical denials, Ensemble built a generative AI pipeline with clinical experts instead of relying on an off-the-shelf model. The system helps assemble appeal packets from long medical records, guidelines, and payer policies, while a clinician makes the final quality decision. Wan reports a 40% reduction in appeal time and says the team can measure return on investment directly. He is also clear that automation alone is insufficient. The longer-term goal is a coordinated system that fixes errors upstream. ## Key ideas ### Revenue cycle problems are putting hospitals under financial pressure [00:00](https://www.youtube.com/watch?v=TquUsN1QsWs&t=0s) Wan opens with the financial side of healthcare, which he says has grown in size, cost, and complexity. He estimates that 40% of hospitals operate at a negative margin, and attributes much of the problem to broken manual processes around the revenue cycle. Delays, denials, rework, and lost revenue add friction after care has been delivered. Revenue cycle management covers the patient's financial journey through healthcare, from eligibility checks and registration to documentation, coding, and denial management. Wan says these administrative roles have grown 30-fold over the past three decades, while the number of clinicians has barely doubled. ### Administrative friction often changes the process without changing the outcome [06:09](https://www.youtube.com/watch?v=TquUsN1QsWs&t=369s) Wan defines friction as the inefficient communication between providers, payers, and patients. Much of the work still ends with the same basic result: a claim is paid or it is not paid. Denials create a large burden because providers must manage and appeal them while operating on slim margins. He gives an example in which a claim was denied and appealed four times. The provider sent documentation repeatedly through different interfaces and did not receive payment until 200 days after the procedure. Wan says AI could move effort away from this bureaucracy and toward clinical care or other productive work. ### Preventing technical errors is more useful than building a better appeal process [09:19](https://www.youtube.com/watch?v=TquUsN1QsWs&t=559s) Wan says many denials do not require a more intelligent appeal. They result from registration errors, missing data, or other technical problems that could have been avoided earlier. Ensemble's end-to-end position gives it access to data from the start of the process through the final denial. In prior authorization, payers may require permission for a procedure, but it can be unclear whether authorization is needed, and policies can change. By connecting prior authorization requests with later denial data, the team can predict risk and flag missing procedures or documentation before submission. ### Clinical denial appeals require evidence from several kinds of records [12:07](https://www.youtube.com/watch?v=TquUsN1QsWs&t=727s) Clinical denials happen when a payer and provider disagree about whether care was medically necessary. An appeal may require a review of the patient's medical record, clinical guidelines, and payer policies. Wan describes records that can run to hundreds of pages and include text, images, laboratory results, notes, and tables. Different clinical situations require different guidelines, and the work has to be completed under deadlines. This creates a limit on how many expert clinicians can prepare appeals. Generative AI can help gather the material and draft the appeal, but the task still depends on careful evidence selection. ### An off-the-shelf language model was insufficient for clinical appeals [13:53](https://www.youtube.com/watch?v=TquUsN1QsWs&t=833s) Wan says a general model can produce an appeal letter when prompted, but that was not enough for Ensemble's standards. The company worked directly with clinical experts to build a model and pipeline that supports the full process. The generated letter must meet the organization's quality requirements, and a clinical expert makes the final decision before submission to the payer. The system also accounts for differences across service lines and clients. This design keeps human judgment in the approval step while using generative AI to reduce the work required to assemble and write each appeal. ### The appeal system produced measurable operational results [14:34](https://www.youtube.com/watch?v=TquUsN1QsWs&t=874s) Wan reports that Ensemble has increased the speed of the clinical appeal process, with a 40% reduction in time and sometimes more. He also says the team has seen higher quality, measured through the overturn rate, which tracks how often denied claims are overturned on appeal. The volume of appeals has grown as well. Because the work is part of an operating service team, Ensemble can connect the AI system to financial results and measure return on investment directly. Wan presents this direct measurement as an advantage over AI projects where the value is harder to quantify. ### A shared data platform is needed before multimodal AI can work across the revenue cycle [15:39](https://www.youtube.com/watch?v=TquUsN1QsWs&t=939s) Wan says revenue cycle data is scattered across many systems and comes in inconsistent formats. Electronic medical records contain text, images, labs, notes, and tables, which can challenge multimodal language models. Ensemble has invested in a common infrastructure called EIQ that brings multiple data formats into one platform. The company is continuing to build agents for different parts of the revenue cycle, including prior authorization. Wan says the goal requires more than moving tasks through the system faster. It requires connecting events across the process so errors found during appeals or denials can be fixed earlier. ## Notable quotes - "Almost half the hospitals in the country are losing money." (00:30) - "They just need a way to avoid errors that cause these in the first place." (09:19) - "Automation alone isn't going to be enough." (16:59) - "We're tracking it very specifically and measuring it very concretely." (15:17) ## Tools & references mentioned - Google - Ensemble Health Partners - EIQ - Medicare - Medicaid ## Who should watch - Healthcare operators dealing with denials, prior authorization, coding, or other revenue cycle work will get concrete examples of where AI can reduce manual effort. - AI engineers working with messy medical records can see why domain workflows, longitudinal data, and expert review matter more than a generic model demo. - Healthcare leaders evaluating AI investments will find a direct connection between operational changes, appeal outcomes, process speed, and return on investment. ## Related talks - [The Billable Hour is Dead; Long Live the Billable Hour](https://aietalks.com/talks/the-billable-hour-is-dead-long-live-the-billable-hour) (Kevin Madura & Mo Bhasin, AlixPartners, 17:04) - [Healthcare's Agent Bytecode: X12 as the Harness for AI Agents](https://aietalks.com/talks/healthcares-agent-bytecode-x12-as-the-harness-for-ai-agents) (Vasant Kearney, Onlay, 20:25) - [From Ambient Documentation to Clinical Intelligence](https://aietalks.com/talks/from-ambient-documentation-to-clinical-intelligence) (Chaitanya Asawa, Abridge, 21:35) - [Shipping AI to a Million Patients Without an A/B Test](https://aietalks.com/talks/shipping-ai-to-a-million-patients-without-an-a-b-test) (Jared Joselowitz, Ufonia, 19:15) - [Trading Desks to Clinical Trials: Parallels in Applied Vertical AI](https://aietalks.com/talks/trading-desks-to-clinical-trials-parallels-in-applied-vertical-ai) (Ayush Bhardwaj, Allos AI, 20:02)