Andrew Ray, Chief Product and Innovation Officer at Ensemble – Interview Series

Andrew Ray, Chief Product and Innovation Officer at Ensemble, brings deep experience at the intersection of healthcare finance, revenue cycle operations, product strategy, and innovation. At Ensemble, he leads product, operational improvement, and innovation initiatives, including the company’s Rev Cycler partnership with Epic. His background includes healthcare provider transformation and strategic growth work at McKinsey & Company, revenue cycle leadership at Stanford Children’s Health, and earlier healthcare advisory roles at EY and Triage Consulting Group, giving him a strong foundation in both operational execution and technology-enabled healthcare transformation.

Ensemble Health Partners is a healthcare revenue cycle management company that works with hospitals, health systems, and physician practices to improve financial performance, patient experience, and operational efficiency. The company provides end-to-end RCM support across areas such as patient engagement, coding, claims management, denial prevention, payer performance, and account resolution, while using AI-driven decisioning, automation, analytics, and its EIQ revenue cycle intelligence platform to help providers reduce friction and accelerate reimbursement. Ensemble has also partnered with Epic through the Rev Cyclers Program and expanded its work with Microsoft (MSFT ) Azure generative AI and machine learning to advance AI-enabled revenue cycle operations.

What specific breakdowns in traditional RCM workflows convinced you that a purpose-built LLM was necessary? 

Health systems are operating on 2% to 4% margins while initial denial rates climb to 12% to 15%. That is the breaking point that showed us traditional revenue cycle approaches are not solving the core issue and that a fundamentally different approach is needed. That pressure affects every part of the organization, from staffing to technology investment to care capacity.

Most solutions do not address the end-to-end system as a whole. Organizations are left with fragmented tools, siloed data and disconnected activity that depends on stitching context together through prompt engineering on general-purpose models.

That approach is slower, more expensive, less consistent and hard to manage at scale.

A purpose-built model changes that by bringing historical and real-time data into a governed platform. It can anticipate payer behavior, surface relevant intelligence, help reduce denials and operate within compliant, auditable environments.

Has anything since your earlier writing reinforced or changed your perspective on general AI versus purpose-built models? 

Our continued development and deployment of end-to-end AI orchestration across the revenue cycle has continued to reinforce where frontier LLMs shine and where there are significant gaps that we are addressing with purpose-built models.

Frontier LLMs perform well with clinical reasoning, but we are investing heavily in providing the context and expertise from our physicians and nurses that have decades of experience bridging the clinical and administrative divide of healthcare. That is proving to be a powerful combination that delivers exceptional results by imbedding those AI solutions with experts in the loop.

Broad models continue to run into the realities of regulated revenue cycle environments and the lack of training data for navigating nuanced revenue cycle processes. Put simply, off-the-shelf AI identifies a means to an answer but often lacks the nuance to delivering intelligence and processes that drive outcomes and best in class performance.

Adapting general models still produces inconsistency, weak governance and limited ability to manage nuance. This is a structural limitation, not a tooling issue.

Ensemble is building capabilities and reengineering processes and data to deliver outcomes – denial prevention, improved patient engagement and access to care and improved care alignment with payers.

What are the most critical limitations of prompt engineering layered onto general-purpose models in RCM operations? 

The promise of AI in revenue cycle isn’t just about applying a general-purpose model to healthcare tasks — it’s about whether the AI actually understands the domain. The limitations of prompt engineering onto foundation models is that you achieve capable generalist capabilities that pass the reasonableness test on the surface but lack the nuance required to deliver best in class performance that continually adapts and improves. ‘Medical necessity’ means something different to a national commercial payer than it does to a Medicare MAC, and that difference determines financial performance. A model trained on the open internet doesn’t know that, prompting can close some of those gaps but misses complete and accurate contextual nuance.

By design, prompt-engineered models are static. They can’t learn when a payer quietly changes its adjudication behavior, when a new denial pattern emerges, or when an appeal strategy that worked last quarter stops working. At Ensemble, we’ve built the closed feedback loop that turns every claim outcome into intelligence — because we sit at the center of the revenue cycle across hundreds of health systems and payers, at a scale that no single provider could replicate internally.

What we’ve found is that the real moat in AI-powered RCM isn’t the model — it’s the data and the expertise encoded into it. We’ve spent years operationalizing what our best denials specialists, coders, and payer relations experts do to deliver best in class results. You can’t inject that into a system prompt — it must live in how the model was trained, what it’s evaluated against, and how it’s integrated into the workflow.

AI transformation in healthcare revenue cycle will separate into two camps: organizations using AI as a productivity layer on top of existing processes, and organizations that have rebuilt intelligence into the process itself. Ensemble is firmly in the second camp — and for health systems under margin pressure, that distinction is the difference between incremental improvement and transformation.

What does the accuracy ceiling of prompt-heavy systems look like in practice when dealing with payer rules and multi-step workflows? 

The ceiling shows up as inconsistency at the points where revenue cycle performance depends on precision. It is not just about raw accuracy. It is about reliability in real operating conditions.

Prompt-heavy systems try to introduce revenue cycle logic at inference time by loading context into each request. That can work in simpler scenarios, but it breaks down with payer-specific behavior, regulatory nuance and multi-step dependencies.

Outputs may be directionally right, but they are not reliable enough for day-to-day operations.

As complexity increases, cost rises and coherence declines, leading to gaps in sequencing, missed dependencies and loss of context across steps.

That is why we focus on models grounded in real operational data and embedded directly into revenue cycle activity rather than reconstructing context on demand.

Do you see payer AI and provider AI evolving into an arms race, and how should health systems respond? 

The arms race framing does not fully capture what is happening. Both payers and providers are rapidly applying AI across prior authorization and claims processing. In many cases, that increases friction rather than reducing it and can result in automated systems interacting without resolution.

AI on its own does not resolve fragmented revenue cycle operations. Applied on top of fragmentation, it can amplify the problem. The stronger response is to move from reactive to proactive operations. That requires visibility across the full revenue cycle, a stronger data foundation and the ability to anticipate payer behavior, identify risk earlier and resolve issues before submission.

When organizations operate with that level of coordination and insight, the dynamic shifts from friction to stronger performance. We are delivering real results, to that end, where we are achieving reduced friction and faster alignment with payers at every step – with AI reengineered at every step.

How do you ensure a model trained without identifiable patient data still captures real-world complexity? 

The absence of identifiable patient data does not mean an absence of real-world complexity.

Managing 80 million transactions a year, Ensemble has unified more than two petabytes of provider, payer and clinical data into an AI-ready asset. That foundation reflects how revenue cycle activity actually functions at scale, including operator expertise, payer-specific behavior and denial patterns, and it unlocks predictive modeling, real-time insights and agentic AI workflows that help protect providers against payment delays and denials.

We use high-quality synthetic and de-identified datasets from certified sources in a HIPAA-compliant environment. These datasets preserve the structure, variability and edge cases present in real-world scenarios.

We are not replicating the EHR. We take an EHR-native approach, adding intelligence where it matters most, especially in areas tied to payer requirements and decisions that sit outside the EHR.

We enhance this approach with world class experts (MDs, nurses, revenue cycle operators) that provide the approach nuance and expertise that AI and data cannot yet achieve independently.

This approach preserves real-world complexity, drives to outcomes and performance while protecting patient privacy.

How do you translate human decision-making and workflows into something a model can reliably learn and execute? 

What makes this work is the depth of operator expertise behind it. Translating that into our model is not about approximating expertise. It is about capturing how experienced teams actually work.

We train on real revenue cycle tasks and structure the system around the rules, dependencies and patterns operators use every day. The most effective systems combine multiple layers of intelligence. Deterministic logic supports steps that must remain consistent, such as validation and policy checks. Predictive models assess risk, such as likelihood of denial. Reasoning models address complexity, including interpretation and narrative generation.

Each layer reduces uncertainty and reflects how high-performing teams operate in practice.

What does end-to-end orchestration actually look like from patient intake to account resolution? 

End-to-end orchestration means moving from isolated task support to a system that operates across the full clinical-to-cash lifecycle.

It begins at patient intake, supporting accurate data capture, eligibility verification and authorization requirements in real time to prevent downstream issues.

As documentation and coding take shape, the system aligns activity with payer expectations and identifies risk early. Through billing and adjudication, it monitors for payer behavior, contract nuances and potential underpayments.

When issues occur, the system connects outcomes back to root causes and feeds insight upstream.

Today, many organizations rely on 15 to 30 different point solutions across the revenue cycle. Each improves a specific area, but they do not learn together. Gaps between functions allow revenue leakage to persist.

Orchestration brings this together through unified data, coordinated revenue cycle activity and clear accountability, creating a continuous learning system across the entire lifecycle.

Where do current EHRs fall short, and how does your approach extend their capabilities without disrupting existing infrastructure? 

EHRs are foundational. They serve as the system of record for clinical and financial data.

They were not designed to act as systems of intelligence for revenue cycle operations.

The gap is in connecting context across the revenue cycle. While EHRs store and organize information, they do not interpret payer behavior, anticipate denials or guide users through complex, multi-step activity that extends beyond their boundaries.

Much of the work required to manage payer requirements and exceptions happens outside the EHR, which is where fragmentation develops.

Our approach complements the EHR rather than replacing it. We work within existing environments, adding intelligence to the areas where work is actually happening.

This allows us to reduce administrative burden, improve accuracy and help teams focus on higher-value activity without requiring changes to core infrastructure.

How do you see the revenue cycle evolving over the next five years, and what role will human operators continue to play? 

Over the next five years, revenue cycle operations will shift from reactive and labor-intensive to proactive and coordinated.

Less effort will be spent on rework such as appeals and follow-up, and more will be focused upstream on prevention. Denials and underpayments will increasingly be identified earlier, improving both cost structure and financial predictability.

This shift expands opportunity for the workforce. Today, too much skilled labor is tied to repetitive tasks. As automation reduces that burden, teams can focus on exception management, process improvement, strategic decision-making and patient care

Roles do not disappear. They evolve. Human expertise does not diminish. It becomes more critical and more central to overall revenue cycle performance.

Thank you for the great interview, readers who wish to learn more should visit Ensemble Health Partners.