Ten US public-health jurisdictions will test enterprise generative AI through a programme involving the Coalition for Health AI, OpenAI, Anthropic and Accenture. Called PULSE, the initiative will give as many as 2,000 practitioners access to donated licences and produce implementation playbooks in 2027.
The programme acknowledges a genuine capacity problem. Public-health teams are being asked to analyse more data, communicate faster and serve more communities with limited technology investment. Its announcement, however, leaves crucial governance details unresolved.
Five use cases carry very different risks
PULSE will organise state, local, tribal and territorial participants around five areas: biosurveillance and drug-wave prediction; mapping social determinants of health; operational efficiency and community-feedback analysis; public communications and multilingual translation; and automated clinical-data retrieval through a FHIR query engine.
These should not share one generic evaluation. A summarisation assistant used for an internal meeting has a different consequence profile from a system influencing outbreak surveillance or retrieving patient information. Each workflow needs a defined user, data boundary, decision point and human reviewer.
The Coalition for Health AI has not named the exact products, model versions or configurations involved, nor explained how OpenAI and Anthropic systems will be allocated. It has also not said whether pilots will use identifiable, de-identified, aggregated or synthetic data.
FHIR is a standard for exchanging healthcare information, but “FHIR query engine” does not establish what the model will do. It could generate a query, retrieve records, summarise them or perform all three. Incorrect queries and incomplete results require different controls from unsupported prose.
Public communication and translation also demand careful review. A fluent message can contain a subtle medical error or use language inappropriate for a local community. Approval by practitioners familiar with the audience must remain part of publication.
Enterprise licences are only one layer of protection
OpenAI and Anthropic say commercial customer inputs and outputs are not used for model training by default. That is relevant, but it does not define retention periods, access roles, audit logging, storage locations or whether protected health information may be submitted.
HIPAA will apply to some agencies and workflows but not automatically to every participant. The legal obligations depend on the organisation, data and function. Even when HIPAA does not apply, demographic and geographic data can expose vulnerable populations or produce harmful inferences.
Accenture will oversee onboarding and help create the playbooks. Before practitioners begin, each jurisdiction should document permitted information, prohibited uses, authentication requirements, logging, incident response and a process for challenging outputs.
NIST’s AI Risk Management Framework offers a useful structure: evaluate a system in its intended setting, identify affected people, define oversight and continue monitoring. PULSE has not yet published separate privacy, security, accuracy or human-review standards for its five tracks.
Procurement and cybersecurity readiness may also vary sharply. A large state department and a small tribal health organisation do not have interchangeable staffing, integrations or legal support. Shared guidance must preserve these differences rather than treating one successful pilot as a universal template.
Useful playbooks need evidence from ordinary failures
The pilots begin in autumn 2026, and the most valuable output will not be a collection of polished success stories. Other agencies need to know which tasks failed, how often staff corrected results, what training was necessary and where integration costs exceeded the benefit.
Evaluation should measure factual accuracy, retrieval completeness, translation quality, time saved, staff workload and public impact. High-risk uses need prospective approval and documented review before outputs influence operations. Model and prompt versions should be retained so results can be reproduced.
PULSE can give public-health practitioners practical experience that many agencies cannot fund alone. Donated access, peer groups and implementation support lower the entry barrier, while ten jurisdictions create an opportunity to compare different environments.
The programme’s credibility will depend on transparency. Playbooks should distinguish provider policy from local configuration, specify minimum safeguards and show when generative AI was less suitable than conventional software. If PULSE captures uncomfortable lessons as carefully as wins, it can help public agencies adopt AI thoughtfully. If it reports only adoption and enthusiasm, the most important governance questions will simply be passed to the next jurisdiction.
