AI in Revenue Cycle Management: The ROI Evidence from Health Systems

The ROI Challenge in AI-Powered Revenue Cycle Management

The pitch from AI vendors has become familiar: deploy our platform, automate your revenue cycle, and watch the savings accumulate. Yet for the hospital CFOs and financial leaders who must sign the checks, the reality has been far murkier. A 2024 poll by HFMA and FinThrive of 101 healthcare organizations found that while 63% have already integrated AI-powered automation into revenue cycle workflows, only 15% report seeing a positive return on investment so far. That gap between adoption and realized value is the central tension this article addresses.

The skepticism is not unfounded. The cautionary tale of Olive AI — which raised over $900 million, reached a $4 billion valuation, served 900+ hospitals across 40 states, and then ceased operations in October 2023 — looms large. A devastating investigation by Axios found that Olive generated only a fraction of the savings it promised, and its "AI" often turned out to be 1990s-era screen-scraping technology. The company strategically pivoted 27 times, according to TechCrunch. Stories like this have made health system leaders justifiably wary of vendor claims.

But the evidence base is maturing. A growing body of published, attributable data from named health institutions — not vendor white papers — now provides concrete, quantified outcomes that can inform investment decisions. This article synthesizes that evidence: aggregate survey data from HFMA, AHA, and Oliver Wyman; detailed case studies from Auburn Community Hospital, Methodist Health System, Stanford Health Care, and a Fresno community health network; and a practical framework for evaluating AI RCM investments on your own terms.

What the Surveys Tell Us: Adoption Rates, ROI Expectations, and Persistent Barriers

Three major surveys conducted between 2024 and 2026 provide a macro-level view of where the market stands. While each uses a different methodology and sample, together they paint a consistent picture: adoption is accelerating, ROI expectations are cautiously optimistic, and the barriers to success are well understood.

The HFMA-FinThrive poll also identified the top barriers to AI RCM adoption. IT infrastructure limitations topped the list at 51%, followed by lack of budget (44%), integration challenges (43%), difficulty demonstrating ROI (42%), and vendor reliability concerns (42%). These numbers suggest that the problem is not a lack of interest — it is a lack of organizational readiness and clear evidence to justify the investment.

The HFMA-AKASA survey of 519 CFOs and revenue cycle leaders added another critical data point: respondents estimated that 8.49% of total revenue is at risk due to documentation or coding issues. For a health system with $1 billion in net patient revenue, that represents nearly $85 million in potential leakage — a figure that makes even a modest improvement in coding accuracy or denial reduction financially compelling.

The Oliver Wyman survey of 200+ decision-makers found that 92% of respondents agree there are "no-regret" AI investments. The top no-regret moves identified were ambient documentation, clinical documentation improvement (CDI), coding automation, and electronic prior authorization. Some studies cited in the survey show AI accuracy reaching 90%+ in specific clinical domains and up to a 46% reduction in coding time for complex cases.

Named Case Studies with Quantified Outcomes

Survey data establishes the macro context, but the most compelling evidence for AI RCM ROI comes from individual health systems that have published their results. The following case studies are drawn from sources including the American Hospital Association (AHA), HFMA, Fierce Healthcare, and HealthTech Magazine — not from vendor marketing materials.

Auburn Community Hospital: 10x ROI from a 99-Bed Rural Hospital

Auburn Community Hospital, a 99-bed rural access hospital in upstate New York, deployed an AI-powered coding and CDI platform and achieved results that would be impressive for any institution, let alone a small rural facility. According to a case study published by the AHA, the hospital saw a 50% reduction in discharged-not-final-billed (DNFB) cases, a more than 40% increase in coder productivity, and a 4.6% rise in case mix index. The return on investment exceeded 10 times their initial expenditure.

For a hospital of Auburn's size, a 4.6% increase in case mix index is particularly significant. Case mix index reflects the acuity and complexity of patients treated; a higher index translates directly into higher reimbursement under prospective payment systems. Combined with the DNFB reduction — which accelerates cash flow by closing the gap between discharge and final billing — the financial impact is substantial.

Methodist Health System: 14 FTEs Freed, 5,559 Hours Saved

Methodist Health System in Omaha, Nebraska, implemented an AI platform from AKASA to automate revenue cycle workflows. The results, reported by Fierce Healthcare, are among the most detailed publicly available: the AI removed 71% of accounts from staff queues, freed the equivalent of 14 full-time employees of work, and resolved 56,118 accounts over eight months, saving 5,559 hours of staff time.

The 14 FTEs freed is not a theoretical projection — it represents actual work that no longer requires human intervention. In a labor-constrained environment where experienced revenue cycle staff are difficult to recruit and retain, this kind of capacity reclamation has both financial and operational value. Methodist did not lay off staff; it redeployed them to higher-value tasks that AI cannot handle.

Stanford Health Care: 17 Hours Saved in a Two-Month Pilot

Stanford Health Care took a more targeted approach, piloting an AI tool in January 2025 to handle patient billing queries. According to HealthTech Magazine, 10 billing representatives processed 1,000 messages using the tool, which generated draft responses that considered each patient's insurance policy and the organization's brand voice. Human reviewers checked every draft before sending. The result: approximately one minute saved per message, totaling 17 hours over two months.

Seventeen hours may seem modest compared to the Methodist or Auburn numbers, but the pilot was intentionally narrow in scope. The billing staff were reportedly "super excited" about the tool, and Stanford expanded it to the entire billing staff by March 2025. The pilot demonstrates that even small, well-scoped AI deployments can generate measurable time savings and staff enthusiasm — both of which are prerequisites for broader adoption.

Fresno Community Health Network: 22% Reduction in Prior-Authorization Denials

A community health network in Fresno, California, used an AI tool to review claims before submission, targeting the prior-authorization process that is a major source of administrative burden and revenue leakage. According to the AHA, the network achieved a 22% decrease in prior-authorization denials by commercial payers and an 18% decrease in denials for services not covered — all without hiring additional RCM staff. The tool saved 30 to 35 hours per week on back-end appeals alone.

For a community health network operating on thin margins, the ability to reduce denials without adding headcount is transformative. Prior authorization is one of the most labor-intensive RCM functions, and the 30–35 hours saved weekly on appeals represents a significant operational win.

Hard Savings vs. Soft Savings: A Framework for Evaluating AI RCM ROI

Not all savings from AI RCM are created equal. When building a business case, it is essential to distinguish between hard savings — directly measurable reductions in cost or increases in revenue — and soft savings, which are real but harder to quantify. The following table provides a framework for categorizing and evaluating each type.

The Olive AI story is a cautionary reminder that not all vendor claims hold up under scrutiny. Olive promised transformative savings across the revenue cycle, raised nearly a billion dollars, and served hundreds of hospitals — but when independent journalists investigated, they found that the promised savings were largely unrealized and the technology was far less sophisticated than advertised. The lesson is not that AI RCM is a mirage, but that health systems must demand evidence, run pilots with clear success criteria, and verify vendor claims against independent data.

Key Metrics for Measuring AI RCM ROI

To evaluate AI RCM investments rigorously, health system financial leaders should track five essential metrics. These metrics are already used to measure RCM performance; the key is to establish baselines before AI deployment and measure the delta after implementation.

• Discharged-Not-Final-Billed (DNFB) Rate. DNFB measures the lag between a patient's discharge and the final coding and billing of that encounter. A high DNFB rate means delayed cash flow and increased risk of write-offs. Auburn Community Hospital's 50% reduction in DNFB cases directly accelerated revenue realization. Target: reduce DNFB days by 20–50% depending on baseline.
• Cost to Collect. This is the total cost of revenue cycle operations divided by net patient revenue. It includes labor, technology, and outsourced services. AI that automates manual tasks should lower this ratio. Methodist Health System's 14 FTEs freed directly reduces cost to collect. Target: reduce by 10–30% over 12–24 months.
• Clean Claim Rate. The percentage of claims that pass payer edits and are accepted on first submission without manual intervention. Low clean claim rates drive up denial rates and rework costs. AI pre-submission review tools, like the one used by the Fresno network, directly improve this metric. Target: 90%+ clean claim rate.
• Days in Accounts Receivable (A/R). The average number of days between claim submission and payment. Faster cash realization improves liquidity and reduces the need for borrowing. Vendor-reported benchmarks suggest AI can reduce days in A/R by 40–50%, but independent validation is limited. Target: reduce by 10–20 days from baseline.
• Net Revenue Impact. The ultimate measure: did AI increase net patient revenue? This can come from improved coding accuracy (higher case mix index, as Auburn saw), reduced denials, faster cash realization, or a combination. The HFMA-AKASA survey's finding that 8.49% of revenue is at risk from documentation/coding issues provides a ceiling for potential improvement. Target: 1–5% net revenue improvement.

These metrics should be tracked at monthly intervals, with a clear pre-deployment baseline of at least six months. Without a baseline, it is impossible to attribute changes to AI versus other factors such as payer policy changes, staffing changes, or seasonal volume fluctuations.

Building the Business Case: A Practical Guide for Health System Leaders

The evidence from surveys and case studies provides a foundation, but every health system's circumstances are different. Building a defensible business case for AI RCM requires a structured approach that accounts for your organization's specific revenue cycle challenges, IT infrastructure, and change management capacity.

Step 1: Identify High-Impact Use Cases

Not all RCM functions are equally suited for AI automation. The survey data points to three areas where health systems are seeing the strongest results:

• Prior authorization. The HFMA-FinThrive poll found that 73% of organizations believe AI will have the biggest impact on prior authorizations. The Fresno network's 22% denial reduction demonstrates the potential.
• Medical coding and CDI. Auburn's 40%+ coder productivity gain and 4.6% case mix index improvement show the direct revenue impact. The Oliver Wyman survey identified coding automation as a top no-regret investment.
• Denial management. 67% of HFMA-FinThrive respondents said AI will drive the most significant impact on denials and underpayment management. Methodist's 71% queue reduction and the Fresno network's 30–35 hours saved weekly on appeals are concrete examples.

Step 2: Set Realistic ROI Expectations Based on Peer Benchmarks

The 15% positive ROI figure from the HFMA-FinThrive poll is sobering, but it reflects early-stage deployments. The same survey found that ROI expectations are expected to rise to approximately 30% in the next quarter as deployments mature. The Oliver Wyman survey's finding that 70–90% of decision-makers expect to increase spending on AI-enabled RCM suggests confidence is growing.

When building your business case, use the case studies in this article as benchmarks — but adjust for your organization's size, complexity, and starting point. A 99-bed rural hospital like Auburn cannot expect the same absolute savings as a multi-hospital system like Methodist, but the percentage improvements (50% DNFB reduction, 40%+ productivity gain) may be replicable.

Step 3: Structure a Pilot with Clear Success Criteria

Stanford Health Care's approach offers a model: start with a narrow, well-defined use case (patient billing queries), run a short pilot (two months), measure a specific metric (time saved per message), and use the results to justify expansion. Key elements of a successful pilot include:

• A clearly defined scope that is large enough to generate meaningful data but small enough to manage
• Baseline measurement of the target metric for at least three months before deployment
• A control group or parallel process to isolate the AI's impact
• Predefined success thresholds (e.g., "reduce denial rate by 15%" or "save 10 hours per week")
• A plan for collecting qualitative feedback from staff — Stanford's billing staff enthusiasm was a key signal

Step 4: Avoid Common Pitfalls

The barriers identified in the HFMA-FinThrive poll — IT infrastructure limitations, integration challenges, and vendor reliability concerns — are not abstract. They are the reasons why 85% of organizations have not yet seen positive ROI. Specific pitfalls to watch for include:

• Over-reliance on vendor projections. The Olive AI story is the extreme case, but even reputable vendors may present optimistic scenarios. Insist on seeing results from organizations similar to yours, and verify them independently where possible.
• Underestimating integration costs. 43% of organizations cite integration challenges as a top barrier. AI tools must connect to your EHR, practice management system, and payer portals. Integration is often more expensive and time-consuming than the software license itself.
• Failing to account for change management. Revenue cycle staff have spent years developing expertise. Asking them to trust an AI tool requires training, communication, and leadership buy-in. Stanford's approach — having human reviewers check every AI-generated draft — built trust before expanding.
• Treating AI as a one-time fix. Revenue cycle processes, payer policies, and coding guidelines change constantly. AI models require ongoing monitoring, retraining, and updating. Budget for this ongoing cost in your business case.

The Bottom Line: Where the Evidence Stands and What Comes Next

The evidence reviewed in this article supports a nuanced conclusion: AI in revenue cycle management can deliver substantial, measurable ROI, but outcomes vary widely by use case, deployment maturity, and organizational readiness. The case studies from Auburn Community Hospital, Methodist Health System, Stanford Health Care, and the Fresno community health network demonstrate that when AI is applied to the right problems with the right implementation approach, the returns are real and quantifiable.

At the same time, the survey data makes clear that most organizations are still in the early stages of their AI RCM journey. Only 15% report positive ROI today, but that figure is expected to rise to approximately 30% in the next quarter as deployments mature. The 80% of health systems exploring or implementing generative AI for RCM, per the HFMA-AKASA survey, suggests that the industry is betting heavily on these tools — and that the evidence base will continue to grow.

For health system financial leaders, the path forward is clear: start with a well-scoped pilot in a high-impact area like prior authorization or coding automation, measure rigorously against a pre-deployment baseline, and use the results to build a data-driven case for broader investment. The evidence from peer institutions is now strong enough to justify the first step — but only if that step is taken with eyes wide open to the risks, costs, and organizational effort required.