AI-Required CPT Coding Expanded Rapidly, but Use Stayed Sparse Across 11.6 Billion Encounters
AI-linked billing has moved from a policy idea to a measurable part of routine care, and this national analysis shows just how uneven that shift has been. Among 11,605,867,309 encounters in Epic Cosmos, only 335,349 encounters carried one of 65 AI-required CPT codes, a rate of 2.9 per 100,000 encounters. The low absolute rate matters because it shows that even as payment pathways open for algorithm-assisted care, billed use still occupies a tiny fraction of U.S. clinical volume.
The study spans January 1, 2021, through April 14, 2026, and it traces a steep rise over time. Odds of an encounter carrying an AI-required code were 9.88 times higher in 2025 than in 2021. In practical terms, the coding infrastructure for AI is accelerating faster than the overall clinical footprint of these services, which suggests that reimbursement rules, product availability, and quality measurement are each pushing adoption in different directions.
Medicare Patients Were More Likely to Receive Billed AI-Enabled Services
Age and payer status shaped use in a way that is hard to ignore. Encounters were more common among patients aged 65 to 85 years, and use was higher among those with Medicare or Medicare Advantage than among commercially insured patients. The unadjusted odds ratio for Medicare was 5.11, and for Medicare Advantage it was 5.12, both relative to commercial coverage.
Several of the AI-related codes currently in circulation fit older adults especially well. Automated retinal screening, brain imaging analysis, and other algorithm-assisted workflows often target conditions that become more prevalent with age, so a heavier concentration among older beneficiaries is clinically plausible. Reimbursement policy also likely plays a role: once a payer recognizes a service and creates a path for billing, the code can become embedded in specialty workflows far more quickly than the underlying technology would otherwise suggest.
One code category moved with unusual speed once a quality lever appeared.
Automated diabetic retinopathy screening increased after inclusion in a HEDIS quality measure, a reminder that quality programs can act as force multipliers for AI adoption. If a measure rewards detection or screening performance, health systems have a direct incentive to operationalize the technology, train staff around it, and build it into routine care pathways.
Racial, Insurance, and Neighborhood Gaps Point to Unequal Access
The distribution of AI-related billing was not even across patient groups. Use was lower among Black patients, with an odds ratio of 0.72, and among Hispanic or Latino patients, with an odds ratio of 0.34, both compared with reference groups in the analysis. Medicaid enrollment was also associated with lower use, with an odds ratio of 0.65. These differences raise concern that AI-enabled services may be reaching patients already better positioned to navigate specialty care, referral networks, and insurer-specific billing structures.
Geography showed a similar pattern. Residents of the most socially vulnerable areas had lower use, with an odds ratio of 0.53, and residents of the most rural areas had lower use as well, with an odds ratio of 0.86. Those are not abstract disparities in a claims table; they reflect the places where access to specialty imaging, digital infrastructure, and staff trained to deploy AI-enabled workflows may be thinner. In many settings, the decision to use AI is not made by the patient alone. It is embedded in how hospitals buy software, how payers recognize the code, and whether a health system has the operational capacity to bill it at all.
Newer Codes Reached Rural Areas Quickly, While Older Ones Took Years
The temporal pattern across individual codes was striking. Quantitative brain MRI appeared in rural areas within months of becoming effective, while older services such as retinopathy screening took years to spread. That contrast suggests that adoption speed depends not only on the maturity of the technology, but also on whether a code is simple to insert into a workflow and whether it meets an immediate clinical need in settings that already have imaging or screening volume.
A fast rural uptake of a newer code also complicates any easy assumption that AI diffusion always favors urban academic centers first. For some services, the code itself may be easier to deploy than older AI applications that require specialty referral pathways, patient recall systems, or repeated follow-up. In other words, code design, payment rules, and workflow burden can be as important as algorithm performance in determining where AI shows up first.
Payment Policy Now Sits at the Center of AI Adoption
This analysis arrives as the Centers for Medicare & Medicaid Services continues to reshape payment for software and expand outcome-based models. The policy stakes are substantial. If billing determines where AI-enabled care is visible in the claims stream, then payment structure will shape not only uptake, but also who benefits from that uptake and which communities are left behind.
The dataset used here, Epic Cosmos, is enormous, but it remains an EHR-based source rather than a complete census of all U.S. care. Billing patterns also do not fully capture unbilled AI use, and code adoption can differ across systems even when the underlying technology is similar. Still, the signal is strong enough to warrant attention: AI-enabled services are moving into routine care, but the path is neither uniform nor equitable.
For clinicians, administrators, and payers, the next question is not whether AI billing exists. It is which patients are actually receiving these services, in which settings, and under which payment rules. Those differences will determine whether AI functions as a broadly distributed clinical tool or a service that remains concentrated in the same populations that already receive more specialty care.