The AI health boom is running ahead of the infrastructure it needs

Five major AI companies have launched dedicated consumer health products since January 2026: OpenAI's ChatGPT Health, Anthropic's Claude for Healthcare, Microsoft's Copilot Health, and Perplexity Health all arrived within the same quarter, and Google's Gemini-powered health coach (now folded into Google Health)launched on a longer timeline either side of that window. In the same year, the Pan American Health Organization and the Inter-American Development Bank published the most thorough maturity assessment the Americas has ever done of its health information systems, across all 49 countries and territories in the region. Its findings suggest most of the infrastructure these products assume simply isn't there yet.

The assessment, built from over 240 standardized indicators collected between 2019 and 2024, places 42.8% of countries at the most basic maturity level, where health data barely exists in digital form at all. Only 4.1% have reached the fourth of five levels, where governance is in place and systems are integrated. None have reached the fifth and final level.

To put that concretely: when an AI health product wants to pull a patient's lab results into a conversation, those results need to exist in a structured digital format somewhere. In close to half the countries in this region, they don't.

PAHO's companion Plan of Action to Strengthen Health Information Systems, 2024-2030 quantifies the distance still to travel. Only 4 countries currently have a regulatory framework for AI in health, against a target of 30 by 2030. Fifteen have interoperability governance mechanisms, targeting 35. FHIR-related interoperability standards are adopted in 12 countries, targeting 35. Just 5 have adopted ICD-11 for semantic interoperability, targeting 30. National digital health roadmaps exist in only 3 countries, with a goal of 30. And only 7 have a health-specific cybersecurity incident response strategy, against a target of 15.

The infrastructure gap is the visible problem — it can be pointed at, measured, funded. A harder one sits underneath it: in the countries that do have digital health records, the data in those systems is not neutral. It reflects decades of who received care, who got diagnosed, and who was represented in the clinical research that trained today's models. Dermatology AI trained overwhelmingly on light skin tones misses conditions on darker skin. Pulse oximeters have been shown to systematically overestimate blood oxygen levels in darker-skinned patients. Neither is hypothetical; both are documented, and both are the predictable result of building on data that was never representative to begin with. The risk is that closing the infrastructure gap without addressing this in parallel just scales the bias alongside the AI, rather than correcting it.

A year ago, the live question was whether AI could responsibly handle health data. Five companies have effectively decided that it can. The harder question now is whether the health systems underneath can handle AI, and the underlying maturity data suggests most of them can't yet, not from policy resistance, but because the digital plumbing isn't there, and where it is, the data flowing through it carries biases AI will amplify rather than correct.

The organizations doing the slower, less visible work — building interoperability standards, standing up governance frameworks, moving countries toward the standardized middle tier, establishing bias auditing before AI reaches clinical settings — are the ones determining whether these five product launches actually work outside their home markets. Infrastructure determines who benefits. Bias determines who gets harmed. For most of the world, as of today, neither problem is solved.