Original research: Peterson Health Technology Institute, “Driving Adoption of High-Value AI in Healthcare,” October 2026 (37-page report)
Healthcare AI is getting better at producing notes, submitting authorizations, coding encounters, and appealing denials. A new Peterson Health Technology Institute report raises an uncomfortable question: what if we are simply making an inefficient healthcare system run faster? For revenue cycle management, the distinction between automating work and eliminating unnecessary work may determine which AI businesses create durable value.
In its October 14 report, “Driving Adoption of High-Value AI in Healthcare,” the Peterson Health Technology Institute (PHTI) argues that AI's real economic impact depends on the incentives that govern its deployment. If the system rewards documentation, transactions and billable activity, AI will increase those activities. If the system rewards outcomes, access and lower spending, AI can be directed toward those results.
That is a direct challenge to one of healthcare technology's most popular narratives: every administrative task removed from a provider's queue is automatically a win for healthcare. Often it is a win for that provider. But a faster process may create new work for insurers, patients or other providers, and it may leave the underlying rules untouched.
PHTI reports that most prior authorization transactions still depend on phone- or fax-based workflows and cites estimated costs of $20–$30 per submission cycle for providers and $40–$50 for health plans. These are cited industry estimates, not new measurements from PHTI or universal costs for every authorization.
AI can generate an appeal, assemble records, route an authorization request, classify a denial or review a claim faster than a human team. But providers and payers can automate opposite sides of the same conflict. When the underlying medical necessity rules remain fragmented, accelerating submissions and reviews can actually increase the volume of back-and-forth.
PHTI's report characterizes the risk as administrative tools layered onto broken workflows. A provider may reduce labor per appeal. A health plan may reduce labor per review. Neither has necessarily reduced the total number of reviews, denials, resubmissions, escalations or clinician interruptions.
For RCM leaders, that creates two different definitions of success:
These outcomes can align, but they are not the same. A vendor may produce a strong return on investment for one side of the transaction while leaving the healthcare system more expensive overall.
The next durable RCM moat may not be the best agent for working an exception. It may be the intelligence and integration needed to prevent the exception from existing at all. Automating broken workflows produces activity. Redesigning them can produce lasting operating leverage.
PHTI cites a stark inconsistency across commercial payer rules. Across four major U.S. insurers, the report notes that more than 5,000 procedure codes require prior authorization at one or more of those insurers, but only 3% of those codes require authorization across all four.
This is exactly the sort of underlying variation that a better voice agent or document extraction model cannot eliminate on its own. The hardest work is determining which specific coverage rule applies to which patient, payer, plan, service and setting, then reconciling that requirement with the clinical record.
The report describes early real-time prior authorization pilots with promising company-reported results, including an Optum Rx/Cleveland Clinic effort that reported an 88% reduction in appeals and a 68% reduction in denials caused by missing information. But PHTI cautions that such pilots are not yet proof of broad, affordable performance across many plans and conditions.
And PHTI makes an important observation: providers and payers have not yet established that AI prior authorization tools lower the average cost per claim after accounting for the cost of the AI system itself.
That creates a clear opportunity for differentiated vendors. The most valuable products may combine structured payer requirements, machine-readable policies, clinical evidence, secure data exchange, exception handling and feedback from actual decisions—not just faster form completion.
Revenue cycle investors are backing companies that accelerate eligibility, authorization, coding, denial management, appeals and collections. Many of these products can be commercially valuable even before national administrative simplification occurs.
But PHTI's argument raises a strategic question for evaluating an AI RCM business: does its value increase as the process becomes more complex, or as the underlying process becomes simpler?
A system paid per appeal might benefit from more appeals. A system paid for fewer avoidable denials has an incentive to fix upstream issues. Both models can create value in the current market, but they carry different long-term risks if payer policy infrastructure, APIs and reimbursement incentives change.
The report points to machine-readable medical policies, standardized API access, structured decision rationales, better denial monitoring and the elimination of repetitive steps. CMS's interoperability and prior authorization rule calls for specified payer APIs beginning in January 2027. The practical test is whether the infrastructure actually reduces conflicting requirements and duplicate work—not whether it simply allows both sides to send more messages electronically.
The report is equally pointed about AI scribes and coding. Better documentation can be clinically useful and can capture legitimate work that was previously missing from the record. But documentation technologies also have the potential to increase coding intensity and spending without proportionate improvement in outcomes.
PHTI describes one multihospital health system's research in which AI scribe adoption was associated with a 5% increase in Level 5 established-patient encounters and a 7% increase in Level 4 encounters, along with an average $1,004 increase in monthly revenue per clinician using the scribe. These figures describe a cited study, not a national estimate of AI coding impact or proof of inappropriate billing.
The issue is incentive alignment. A provider may regard more complete documentation and appropriate higher-complexity coding as accurate reimbursement. A payer may see aggregate utilization and coding intensity increasing. Patients and employers care whether additional spending corresponds to measurable improvement in care.
PHTI reports that insurers are responding with more automated review, reimbursement adjustments and scrutiny of coding outliers. The outcomes of these strategies remain uncertain. The danger is a new cycle of AI-generated documentation met by AI-generated scrutiny, which produces more administrative friction for both sides.
For RCM vendors, success cannot simply be maximizing billed complexity. Systems will need to support correct coding, evidence, clinical necessity, compliance and defensible payment—while recognizing that correctly paid care is not equivalent to maximizing reimbursement at any cost.
PHTI proposes coordinated action across four domains. The framework is primarily intended to accelerate high-value clinical AI, but parts of it are directly relevant to administrative and revenue-cycle software:
This is not a recommendation that an RCM software contract should be reimbursed like a clinical AI treatment. It is a broader lesson about incentives, measurement and workflow redesign.
The report urges three principles for payment of high-value clinical AI. First, payment should create system-level value by reducing spending while preserving or improving care. Second, compensation should reward measurable outcomes and appropriate use rather than each task performed. Third, payment levels should evolve as evidence improves, the technology scales and the cost of producing services changes.
PHTI points to CMS's ACCESS model as one effort to support outcomes-based care. According to the report, ACCESS launched in July 2026 with more than 160 organizations focused on cardiometabolic, musculoskeletal and behavioral health conditions.
The implications for future RCM infrastructure are significant. As AI performs more of the underlying clinical activity, reimbursement systems will need to distinguish human professional work, AI-supported services, autonomous actions, supervision and population-based outcomes. Traditional billing systems and payer policies were not designed for that division of labor.
That creates potential demand for new policy intelligence, coding controls, evidence capture, contract logic, payment models and auditable workflows. It also creates genuine regulatory and operational uncertainty. The report's proposals are not current universal reimbursement rules.
PHTI's report is not an argument to stop using administrative AI. Providers cannot simply opt out of payer complexity while waiting for policy reform. It is an argument to measure automation more intelligently and invest in systems that can improve the underlying process.
RCAI would prioritize five questions when evaluating an RCM AI deployment:
The uncomfortable truth in PHTI's report is that healthcare can automate its way into greater complexity.
In the near term, a strong RCM agent may be worth buying because the current reimbursement environment is complicated. In the longer term, the businesses with the greatest strategic value may be the ones that make that complexity unnecessary: payer-specific intelligence, defensible coding and payment, real-time decisions, fewer denials and workflows that stop generating repetitive human work.
There is no contradiction in saying that RCM AI can deliver strong provider-level ROI while questioning whether every use case improves the entire healthcare system. Both can be true. What matters is measuring them separately and designing incentives so that gains in one place are not simply costs transferred somewhere else.
The winning question for revenue cycle AI is changing from “How much work can we automate?” to “How much work should never have existed?”
Independent analysis of AI, payer policy, healthcare economics and revenue cycle transformation.
Source and methodology: Peterson Health Technology Institute, “Driving Adoption of High-Value AI in Healthcare,” October 2026, especially pp. 7–10 (executive summary and roadmap), 14–16 (prior authorization and billing), and 25–29 (policy and payment design). This is a synthesis and independent RCM-focused analysis of PHTI's policy report, not a newly conducted survey or an audited vendor ROI study. Statements about the report are attributed to PHTI; implications for RCM business models are RevCycleAI analysis.