The Follow-Through Problem
Medicine has no shortage of breakthroughs. What it lacks, according to a new Viewpoint published in JAMA by Adam L. Beckman, MD, MBA, and Dave A. Chokshi, MD, MSc, is a reliable system for making sure patients actually receive them. The authors open with a sobering pattern: pre-exposure prophylaxis for HIV, antiviral treatment for hepatitis C, follow-up colonoscopy after a positive stool-based colorectal cancer screen, and blood pressure control all "reach less than half of the patients who are likely to benefit from them." As the authors put it, "Each is a breakthrough with a failure to follow through."
Cervical cancer is their central case study. Human papillomavirus (HPV) vaccines are nearly perfectly effective, and early-stage disease carries a survival rate above 90%. Elimination is not theoretical — Australia is on track to eliminate cervical cancer as a public health problem by 2035. Yet in the United States, cervical cancer incidence is rising among women aged 30 to 44, and more than a third of American adolescents remain behind on HPV vaccination. The historical precedent the authors invoke is smallpox: the first vaccine arrived in the 1790s, but eradication did not follow until two centuries later, once vaccination was finally paired with coordinated monitoring and containment.
A Hypothetical, Made Concrete
Rather than argue in the abstract, Beckman and Chokshi walk through a plausible near-future scenario. "Ms L," a 43-year-old Vietnamese-speaking woman working two jobs while raising a child, is flagged as 10 years overdue for cervical cancer screening by an AI system layered onto a state public health registry — one that reconciles claims data, scanned PDFs, and free text scattered across multiple electronic health records. A conversational AI agent reaches her by text and phone, explains the overdue screening in Vietnamese calibrated to her health literacy, and arranges an at-home HPV self-collection kit. When her result comes back positive for a high-risk genotype, a clinician — not an algorithm — delivers the news and coordinates next steps, while an AI agent books her colposcopy and arranges transportation through an insurance benefit she didn't know she had. Months later, when she misses a follow-up visit, a separate AI tool flags the gap to a community health worker.
The point of the vignette isn't novelty — every one of those steps could, in theory, be done manually. The value, the authors argue, is that "AI reduces the administrative friction hindering reliable delivery," making it possible to strengthen the safety net for every patient at once, not just the ones with the resources to chase down their own care.
Four Ingredients for a Public Health Agenda
The authors are explicit that realizing this vision requires more than good software. They outline four components that must move together:
Government leadership. Federal efforts — CMS interoperability rules, an HHS internal AI guide — have laid groundwork, but their impact is blunted by a weakened public health infrastructure and shrinking insurance coverage. Notably, ARPA-H's Health Care Rewards to Achieve Improved Outcomes model, one of the more promising federal efforts in this space, was recently shuttered. That gap elevates the role of states, which the authors note are "well suited to drive public health campaigns" — pointing to Alabama's newly launched cervical cancer elimination campaign as an example.
Technology at scale. AI's contribution is less about diagnostic novelty than administrative reach: identifying patients invisible to fragmented data systems and engaging them at volumes no human workforce could match.
Cross-sector partnership. State government may be the natural convener, but hospitals, health plans, and community organizations remain essential for execution — combining clinical infrastructure with a public health lens that can see patients clinical systems miss.
Capital. The authors call for public funding through vehicles like ARPA-H and the Center for Medicare and Medicaid Innovation, directed specifically toward community health centers, safety-net hospitals, and local health departments. They also point to philanthropic capital — citing the OpenAI Foundation's roughly $130 billion endowment as an example of resources that could be directed toward AI-enabled public health work.
The Risks Are Real — and Not a Reason to Retreat
Beckman and Chokshi don't romanticize the technology. They acknowledge that "algorithmic case finding can miss patients invisible to fragmented or biased data" and that automated outreach risks eroding trust in communities with long, justified memories of institutional surveillance. Their answer is not caution-as-paralysis but caution-as-design: human oversight, transparent monitoring of algorithmic performance, and regulatory guardrails built in from the start.
The Bottom Line
The authors are openly skeptical of the more grandiose claims circulating about AI in medicine — the idea that it might "cure all disease" or double human life span. But they pair that skepticism with a pointed reframe: the current status quo, in which the average American gets just one healthy birthday after age 65, is arguably just as hard to accept. Diseases like cervical cancer and hepatitis C are already scientifically solvable. What's missing is delivery infrastructure — and that, the authors argue, is precisely where AI, backed by government leadership, cross-sector collaboration, and sustained investment, could finally make the difference between knowing what works and making sure it reaches people.
