2026-08-21
ChatGPT Health, Amazon, Ant's Afu and Claude Healthcare wire LLMs into records, booking and pharmacy. Topol and Keane: public health must govern pathway control, not model scores.
Consumer health products are sliding from answering medical questions to putting a patient on an executable path. OpenAI's ChatGPT Health, Amazon Health AI, Ant Group's Afu, and Anthropic's Claude for Healthcare all attach large language models to records, scheduling, pharmacy fulfilment, payments, and clinical workflow. The shift bites hardest in resource-limited settings, where outcomes fail because patients never finish the sequence, not only because care is missing.
This Nature Health Perspective says public health should watch platform integration depth, not model leaderboards. Whoever controls how symptoms are read, where patients are sent, how payment happens, and whether treatment is completed, controls the pathway. The full article sits behind a paywall. What follows uses the abstract, the public framework figure, and the reference list. Product-level detail and any quantitative case evidence live in the unread body.
Four product trajectories are read as case studies and folded into a pathway-level accountability frame. Integration depth is tied to evaluation, procurement, routing transparency, data governance, and exit options. The figure splits the stack into four layers and tags each layer with pathway control, population-health opportunity, and governance.
The reference list also points at nearby products: Perplexity Health wiring records and wearables, Microsoft Copilot for Health as a consumer assistant, and Ant's AQ app, described in a 2025 press note as 15 million monthly users, with later reporting of more than 100 million users. Those are public materials the authors use to place the four trajectories, not a results table from the body.
A Perspective does not run a trial. The abstract flags three population-level consequences: care completion will change, triage power will concentrate, and new asymmetries will appear in data and operational control.
The figure hangs completion on the orchestration layer. Once scheduling and prescriptions are attached, the model is no longer only advice; it decides whether advice becomes a visit, a prescription, a filled box. Concentrated triage sits across the interface and orchestration layers: if entry, framing, and provider routing live on one platform, incumbent clinics become called backends. Data asymmetry sits in the data layer: longitudinal personal health data plus platform interaction logs that providers and regulators may never see in the same form.
The policy claim is blunt. Existing governance should cover the pathways and platform incentives these systems now shape, not models alone. The accountability list is evaluation, procurement, routing transparency, data governance, and exit options. Advertising control and conflict-of-interest review are drawn on the figure, so commercial incentives are treated as part of pathway design.
Medical LLM work still reports diagnostic accuracy. This piece moves the fight to who owns the path. Once a product touches records, booking, and pharmacy, evaluation has to add completion, triage concentration, data exit, and whether a buyer can leave, rather than another question-answering benchmark. If purchasers buy model quality alone, they hand the orchestration layer's routing rights to the platform by default.
For Chinese readers, Ant's Afu sits on the same diagram as ChatGPT Health and Claude for Healthcare. Super-app health entry points and chatbot entry points are treated as the same infrastructure problem.
This is a Perspective, not a new experiment. How far Afu or ChatGPT Health actually reach into booking and fulfilment should be in the body; that text was not retrieved, and press notes are not the paper. The figure is a conceptual stack and carries no completion or concentration numbers. Conflicts of interest belong in the reading: Keane co-founded Cascader and consults for several device and drug firms; Topol advises Microsoft AI, Perplexity AI, and Abridge AI. Grammar editing of the draft used Claude Sonnet 4.5, which the authors disclose.