When Patients Hand Their Records to an AI Agent, Who Is Governing What Happens Next

Athenahealth now lets patients route their own records into Anthropic’s Claude through a new MCP connection, giving patients personalized answers drawn from their clinical data fiercehealthcare.com. Zocdoc has separately opened its scheduling infrastructure to AI agents and partnered with an AI personal assistant called Instinct fiercehealthcare.com. Neither is a pilot. Both are live, patient-facing, and built on the premise that a third-party AI system gets ongoing access to a person’s clinical history, not a single export.

HIPAA’s individual right of access lets a patient direct a provider to send records to a third party of their choosing, and once that direction is made the data generally leaves the covered entity’s custody. Information blocking rules under the Cures Act govern whether a provider or developer improperly interferes with that access. Both were built around a bounded transaction: patient asks, provider sends, transaction closes. Neither was written with a persistent agent connection in mind, one that keeps pulling and synthesizing a patient’s record on an ongoing basis rather than receiving it once.

A recent governance analysis of payer-provider data exchange flagged this seam directly, noting that no CMS or HHS guidance identifies autonomous agents in FHIR-based exchange, and that existing API and HIPAA obligations apply to agent architecture only by extension, leaving organizations to map the fit themselves trussed.ai. CMS-0057-F’s patient access API rules, which govern how payers and providers exchange data under the FHIR standard, were written for that same bounded model vernier.health. A standing agent connection does not obviously sit inside that frame.

The timing makes the gap harder to ignore. HHS just finalized HTI-5, and multiple analyses describe it as narrowing health IT certification scope while deferring the information blocking changes that would have been the most direct vehicle for addressing exactly this question reedsmith.com mcguirewoods.com. Stakeholder reaction has been mixed on what the rule does and does not resolve healthcareitnews.com. The one rulemaking cycle positioned to touch standing-agent access just closed without touching it.

Meanwhile the product pressure that makes agent access attractive is real. Patient history scattered across faxes, scanned-in outside records, and years of disconnected notes is the fragmentation problem vendors are now building tools to stitch together for clinicians and the AI systems they use rutlandherald.com, and ONC’s own work on patient access has pushed toward faster, more complete delivery of results and records healthit.gov. Product direction and policy direction are aligned. The rulemaking that would govern the connection between them is not.

The question in front of compliance committees is narrow but concrete. When a patient authorizes a standing link from the EHR to a third-party AI assistant, does the health system’s obligation end at authorization, or does it extend to the scope, duration, revocability, and downstream use of that connection. HTI-5 just confirmed that answer will not come from the next rulemaking cycle. It will come from whichever authorization flow generates the first dispute.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The argument is coherent and well-structured—identifying a genuine regulatory gap between bounded-transaction frameworks and persistent agent connections—though the final claim that resolution ‘will c
Source & Claim Verification	In-house seat · hosted fallback (qwen-2.5-72b-instruct)	held. Several key sources could not be opened, including those supporting claims about Athenahealth and Zocdoc’s AI integrations, impacting the verification of critical factual claims.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately reflects the regulatory gaps in HIPAA, Cures Act information blocking, and ONC health IT certification (HTI-5) regarding persistent AI agent access to patient records, but does
Technical Accuracy	Llama	cleared. The article accurately describes the current regulatory gap regarding standing agent connections between EHRs and third-party AI assistants, and correctly identifies relevant regulations and industry
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and highlights the lack of specific counterarguments or regulatory frameworks addressing persistent AI agent access to patient data, rather than relying on vendor h
Novelty & Non-Duplication	Grok	cleared. Athenahealth-Claude MCP and Zocdoc-Instinct launches supply a genuinely new patient-directed persistent-access fact pattern not covered in the recent HTI-5/permission-problem cluster, even though the
Validation	DeepSeek	cleared. The central claim that HIPAA and Cures Act rules were built around bounded transactions, not persistent AI agent connections, is validated by the cited governance analysis and rulemaking documents.

Sources cited: 44. Validation challenges: 0. Learn how these briefings are written and verified.