Anthropic is integrating Claude more deeply into the US healthcare system. On August 27th, the company released Claude, which includes connectors to the US Federal Medicare coverage database, ICD-10 coding, and the national healthcare provider identification registry, among others. It also opened up access to health records and wearable data for some individual subscription users. The life sciences product line has been extended to include clinical trials and regulatory filings as well. On the surface, this seems like a combination of functions; however, the real change is that the model is now dealing with the most fragmented and sensitive data streams in the healthcare industry.
This release should not be interpreted as Claude obtaining independent diagnostic or treatment authority. Anthropic positions the product as a support provider, payer, medical technology company, and individual information organizer, emphasizing that HIPAA-ready infrastructure is used, user authorization is required, permissions are revocable, and final decisions are made by professionals. The model can provide suggestions, summaries, and drafts, but this does not mean that its output automatically constitutes a medical decision.
Connectors target administrative frictions; what may be rewritten first is not necessarily the consulting rooms.
In the American healthcare system, a large amount of time is spent searching for coverage policies, verifying codes, preparing pre-authorization and appeal materials. The newly added CMS Coverage Database connector can retrieve local and national coverage decisions; the ICD-10 connector is used to query diagnostic and procedural codes; the NPI Registry connector helps to verify the identity of service providers. Combined with the existing PubMed connector, organizations can incorporate policies, documents, and internal records into the same workflow for retrieval.
The typical scenario provided by Anthropic is prior authorization. Staff members need to cross-check insurance coverage requirements, clinical guidelines, patient records, and complaint documents. The model can first organize the relevant standards and check if there is corresponding evidence in the records, then generate a recommendation based on that for review by payment party personnel. The same method is also applicable to claims processing, patient message routing, and referral handovers. The efficiency here comes from reducing the need to repeatedly search for information, rather than having the model replace doctors in assessing the condition.
The product also adds Agent Skills for development and prior authorization review. FHIR is an important standard for medical system data exchange, and skill packs can help developers connect different systems; the prior authorization skills are a set of templates that can be modified according to institutional policies. Templating helps to unify processes, but it also means that institutions must clearly define their own rules, exceptions, and upgrade paths. If the underlying policies expire or patient records are incomplete, the model may still generate results that are fully formatted but lack sufficient basis.
The boundaries on the personal end are more sensitive. Users in the US with IDs Claude Pro and Max can choose to access laboratory results and health records. Connectors for IDs HealthEx and Function have entered the testing phase, while those for IDs Apple Health and Android Health Connect are being rolled out in batches on mobile devices. According to Anthropic, users can specify the scope of sharing, disconnect at any time, or modify permissions; their personal health data will not be used for model training. Claude allows for the interpretation of test results, summarization of medical history, identification of changes in health indicators, and assistance in preparing for medical consultations. However, the product includes prompts about uncertainties and guides users to seek personalized advice from professionals.
HIPAA-ready is not an exemption label; institutions still need to control its use and responsibilities.
Medical data compliance cannot be determined solely by whether suppliers provide the necessary infrastructure. Hospitals, insurance companies, and startups still need to decide which data can be included in models, for how long it can be stored, who has access to it, whether the output will be part of official medical records, and how errors can be detected and corrected. HIPAA emphasizes the boundaries for the use and disclosure of protected health information, but just because a set of technologies meets deployment requirements does not mean that every practical application is automatically compliant.
Even if models are connected to authoritative databases, it is not possible to eliminate illusions. There are regional differences in coverage policies; the ICD-10 coding needs to be combined with specific medical records, and the preprints have not yet undergone peer review. Patient-worn data may also be affected by device accuracy and wearing habits. Systems should retain the source links and query times, allowing reviewers to return to the original materials, rather than relying solely on a smooth summary. For high-risk matters, it is best to set a hard threshold that requires manual confirmation.
In the field of life sciences, new connectors such as Medidata, ClinicalTrials.gov, ToolUniverse, bioRxiv, medRxiv, Open Targets, and ChEMBL have been added. The scope of work has expanded from early-stage research to site selection for trials, enrollment progress, draft formulation of plans, and preparation of regulatory documents. Anthropic also provides skills in drafting clinical trial plans, and can be combined with the requirements of FDA and NIH to propose endpoints. It is important to note the word "draft" here: trial design involves statistical efficacy, ethics, patient safety, and regulatory communication. The acceleration of the model refers to the organization of data and the initial draft, not to replace the sponsors and researchers in taking responsibility.
Medical AI is most often promoted as "smarter doctors," but this time, the more realistic opportunities for the product come from the administrative level. There is a lot of repetitive work involved in prior authorization, coding, appeals, and trial documentation. The sources of information are relatively clear, which also facilitates manual review. If these processes can first prove their accuracy and time-saving benefits, and then gradually be applied closer to clinical decision-making, the risks will be more controllable.
Claude for Healthcare Currently, many tools are placed through the same entry point. The key indicators for the next step should not just be the number of connectors, but also the error rate, the proportion of manual modifications, processing time, the success rate of complaints, and whether patient waiting times have truly improved. Medical institutions need to use their own data for verification and to establish proper permissions, logging, and accountability chains first. The ability to read health data is one thing, but the real barrier for medical products lies in being able to use that data stably and traceably within sensitive systems.
