The AI denials market is entering a more demanding phase. Generating a polished appeal letter is no longer the hard part. The hard part is proving that every medical and policy claim in the letter is supported by the right evidence, even when the relevant facts are buried across dozens of documents. R1's new research with NVIDIA suggests that specialized models, purpose-built workflows and evidence verification could matter more than the size of the underlying language model.
On October 15, R1's R37 research lab published the first installment of its work with NVIDIA on long-context medical reasoning for denied claims. The team used NVIDIA Nemotron 3.5 Super VL, R1's proprietary denial-appeal tools and domain knowledge, and NVIDIA's NeMo training software to build R1-super, a specialized family of appeal models.
The reported results are meaningful as early-stage AI research, but they are not evidence that a health system will overturn more payer denials tomorrow. The distinction between a passing synthetic appeal draft and a payer's actual decision is the most important qualification in the entire announcement.
In R1's initial 16-case synthetic evaluation, its final post-trained version passed 31.3% of single attempts and achieved a 71.9% four-attempt score. The company estimated $0.96 in generation cost per passing draft under projected serving assumptions. These numbers concern a strict internal drafting rubric—not appeal overturns, realized cash, or full production cost.
A complex hospital denial is an evidence-reconstruction problem. The appeal must connect clinical presentation, treatments, diagnostic results and hospital course to the specific payer rationale, contractual requirements and standards governing payment.
R1's benchmark includes complex diagnosis-related group (DRG)/coding and medical-necessity disputes. Each case can involve a denial notice and 25 to 52 supporting documents. An AI system might identify the right policy yet cite the wrong passage; discover a supportive lab result but ignore contradictory clinical evidence; or write a logically coherent argument that does not satisfy the insurer's specific payment rule.
Those are not cosmetic mistakes. If an appeal makes one material unsupported assertion, it may undermine the entire submission. R1's evaluation explicitly fails a draft with unsupported claims, omitted essential facts or incorrect required formatting. It does not give the model credit merely for including a citation if that citation does not substantiate the statement.
This is an important departure from the industry's frequent emphasis on task throughput. A faster appeal that cites evidence inaccurately can create rework, payer friction and compliance concerns rather than sustainable collections.
The defensible asset in denials AI may be the evidence environment surrounding the model. Payer policies, clinical record structure, appeal workflows, citation checks, exception handling and actual resolution data can be harder to replicate than the language model itself.
R1 built a specialized execution environment—a harness—that gives the model tools to read documents, attach citations, invoke bounded helper tasks, verify sentences and repair unsupported statements. The benchmark compares that system against a general-purpose Pi harness with R1's appeal skills.
Holding the main model's weights constant, R1 reports that the custom harness improved single-attempt pass rates by approximately 7 percentage points for both untrained Nemotron and GLM-5.3 in paired tests. That is an architecture lesson: better access to evidence and carefully designed verification can improve performance without increasing underlying model capability.
Next came supervised fine-tuning. The research team trained successive versions of R1-super from examples of successful appeals and corrections. Across three reported versions, the initial study's single-attempt pass rate rose from 17.2% to 18.8% to 31.3%; the four-attempt score improved from 37.5% to 50.0% to 71.9%. R1 notes that the first of these versions had already received post-training, so the increase is within a continuing specialization effort, not a comparison against a wholly untrained baseline.
The result illustrates the business case for domain-specific post-training. An RCM company with meaningful volumes of accurately labeled appeals, reliable human expertise and structured evidence workflows can potentially turn its operating experience into performance improvements in a model it controls.
R1's first-stage development evaluation used only 16 complex synthetic cases, with 589 supporting documents across the cases. Denial experts reviewed those cases as representative of difficult appeals, but the cases were still synthetic and relatively few in number.
In that setup, GPT-6 Sol achieved a 35.2% single-attempt score, ahead of R1-super's 31.3%. R1-super's 71.9% four-attempt score was ahead of the 56.3% reported for GPT-6 Sol and GLM-5.3 in their tested configurations. A four-attempt result represents the chance at least one draft meets the evaluation rules. It does not prove that a deployed system can reliably identify that draft or that a human reviewer will accept it.
R1 then ran an exploratory 50-case synthetic comparison. The larger comparison reported the following:
These exploratory differences are narrow, and the expanded sample was selected following earlier performance considerations. The scoring relied on a fixed GLM-5.3 evaluator with two coverage draws; R1 reports those draws agreed on final pass/fail outcomes for 85.4% of readable packets overall and 82.2% for R1-super. The study did not include randomized real-world payer adjudication, independent clinical outcome validation or live appeal cash-recovery results.
It would be inaccurate to say the research proves a proprietary RCM model outperforms all frontier AI systems. The stronger conclusion is that a small, specialized model and a carefully designed workflow can be competitive on a highly specific reasoning task.
Perhaps the most commercially relevant result is the estimated economics. In R1's initial study, the projected generation cost per passing R1-super draft declined from $1.80 to $0.96 across training versions, roughly a 47% reduction. Under the authors' serving assumptions, four R1-super attempts cost about $1.20, versus $7.08 for four GPT-6 Sol attempts in the tested setup.
There is a major caveat: these are estimated generation economics, not all-in hospital cost per successful appeal. The published assumptions exclude judging and selection, model training, human review, deployment overhead and first-time evidence extraction. R1's estimate uses projected NVIDIA B200 serving, whereas actual evaluation GPU cost was higher than the modeled production rate.
A hospital's true unit economic equation must include medical-record extraction, payer-rule retrieval, clinical and coding review, legal and compliance controls, staff escalation, software fees, integration, and the downstream success or failure of the actual payer appeal. Lower inference cost is valuable, but only if the resulting system also lowers those other costs.
R1-super is built on a Nemotron model whose weights can be adapted and whose serving environment can be controlled. For a scaled services-and-technology business, that offers strategic options beyond a subscription to a general-purpose API.
R1 can potentially retrain against its own appeal taxonomy, adjust verification workflows, choose different serving economics and bring domain-specific expertise into the execution loop. As the number of appeals grows, the difference between a few dollars per generated draft and much lower serving cost can become material.
But control is also responsibility. Maintaining a custom model stack requires evaluation infrastructure, expert labeling, guardrails, security, drift monitoring and ongoing adaptation to payer changes. A provider should care less about whether a model is open-weight than whether it produces reliable, evidence-grounded results at a lower fully loaded cost.
R1's advantage, if it materializes, will likely come from combining R37's research with actual operational casework, medical reviewers, payer-policy interpretation and collection outcomes—not from Nemotron alone.
In a real revenue cycle, an appeal that satisfies an internal writing rubric is only a candidate for submission. It must still be clinically appropriate, aligned with payer-specific rules and contractual terms, and filed correctly within required timelines. The payer can deny a well-supported argument; the provider can also submit an apparently successful appeal but fail to collect the correct amount.
RCAI would want five measures before translating these benchmark scores into a claim of material RCM improvement:
That final point may become more valuable than the appeals themselves. A model that repeatedly discovers a missing documentation element should help upstream teams fix that defect rather than endlessly produce better letters defending it.
Many vendors can offer an AI-generated appeal narrative. Fewer can show a rigorous, reproducible evaluation of whether every factual claim is actually supported by source material. R1's research provides a useful example of what stronger technical transparency can look like.
It also raises the bar for the category. Claims that an AI system “automates 90% of appeals” are difficult to interpret unless the vendor defines what completed means. Was a draft created? Did a human approve it? Was it sent? Did the payer overturn the denial? Was the money collected?
As domain-specific evaluation improves, buyers should demand evidence at each step. That may favor businesses with specialized datasets, service infrastructure and longitudinal payment outcomes over startups differentiated mainly by their choice of general-purpose foundation model.
R1's Nemotron research is not a breakthrough in payer overturn rates. It is a meaningful demonstration of how denial-specific reasoning systems can be engineered, benchmarked and made potentially cheaper.
The strongest strategic lesson is that the model is only one part of the product. Evidence retrieval, medical reasoning, payer-policy context, verification, controlled revisions and human oversight must operate together. Specialized training can improve that system further.
For a large RCM organization, the opportunity extends beyond drafting letters: every resolved denial is a chance to make the next denial less likely. The ultimate winner will not be the system that writes the most appeals. It will be the one that safely resolves disputes at the lowest all-in cost and systematically prevents avoidable rework.
Independent analysis of AI denials, payer policy, revenue cycle automation and healthcare economics.
Sources: R1 R37 Research, “Long-context medical reasoning to reduce denials: post-training NVIDIA Nemotron 3.5 Super VL” (October 15, 2026); NVIDIA Nemotron 3.5 technical overview (October 14, 2026). Benchmark results and costs are reported by R1 and derived from synthetic experiments with stated evaluation and serving assumptions. Revenue-cycle implications and diligence questions are RevCycleAI's independent editorial analysis. No operational denial overturn rate or realized cash increase is established by this research.