Who Controls the Agent Is the Governance Question, Not Whether It Works

Health systems are moving fast on agentic AI in clinical workflows. UT System just signed an enterprise-wide agreement with Abridge to expand ambient AI across its hospitals, and HHS has put $1.5 million in LEAP awards behind MedStar and Regenstrief to build agentic AI directly into clinical care onhealthcare.tech. The deployment question everyone is answering is whether the tool works. The governance question almost nobody is answering deliberately is who controls what the agent is allowed to do.

The control question nobody is deciding on purpose

The HHS-funded design is notable precisely because it is not the market default. It puts the patient in control of the agent’s keys, while the commercial trajectory elsewhere hands that control to EHR vendors, health systems, payers, or the AI lab itself onhealthcare.tech. A chatbot that drafts a note and waits for a clinician to sign off never needed a real answer to this. An agent that retrieves records, triggers orders, or shares data autonomously does. That requires identity, authorization, audit, provenance, and a clear line of clinical responsibility, and most enterprise AI contracts signed this year were not built around those five requirements as a checklist onhealthcare.tech.

Vendors building toward “strip it down” agentic deployments inside hospitals are already discovering that narrow, well-scoped agents work better than ambitious general ones medcitynews.com. That is a useful engineering lesson. It is also an admission that scope, and who sets it, is the actual control lever.

HTI-5 raises the stakes on autonomous access

This is not a theoretical debate. Cures Act information blocking enforcement began in September 2025, with penalties up to $1 million per violation for health IT developers eltexsoft.com. HHS’s proposed HTI-5 rule, released in December 2025, signals regulatory intent to let autonomous AI systems retrieve and share health data through FHIR APIs without a human in the loop for every transaction eltexsoft.com. ONC has framed its interoperability push around reducing physician burden and expanding data liquidity, not as an abstract technical exercise healthit.gov. Once autonomous retrieval and sharing become permissible by rule, the question of which party’s agent is doing the retrieving, under whose authorization, stops being a product design choice and becomes a compliance exposure.

The lock-in is happening by default

Modern Healthcare’s reporting on EHR switching costs points to the same dynamic from a different angle. As AI capabilities get built deeper into Epic’s and Oracle’s platforms, the architecture itself starts to determine who holds control over agentic functions, making vendor switching harder precisely because governance was never separated from the underlying system modernhealthcare.com. Vendor transparency notes, like EverHealth’s disclosure for its Scribe AI, are a start, but a disclosure is not an authorization framework support.drchrono.com.

For compliance committees evaluating any enterprise agentic AI deal this quarter, the diligence question is not benchmark accuracy. It is a contractual one: does this agreement specify, in writing, who authorizes agent actions, who audits them, and who is accountable when an autonomous retrieval goes wrong. If the vendor can’t answer that before go-live, the health system has already ceded the decision, and it will not get the chance to make it again later.

---

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 core argument—that governance control over agentic AI is being ceded by default through vendor contracts—is coherent and timely, but the piece overstates regulatory certainty by treating HTI-5’s p
Source & Claim Verification	Qwen · hosted fallback (qwen-2.5-72b-instruct)	cleared. Most factual claims are supported by citations, but a few lines lack specific references, such as the discussion on the ‘strip it down’ approach and the implications of HTI-5 on autonomous access.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately identifies control and governance gaps in agentic AI but understates specific compliance requirements of ISO 42001 (Clause 6.1.2), NIST AI RMF (GOVERN 1.1), and HTI-5’s FHIR AP
Technical Accuracy	Llama	cleared. The article accurately discusses the technical requirements for agentic AI in clinical workflows, such as identity, authorization, audit, provenance, and clinical responsibility.
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and critiques the lack of deliberate governance in agentic AI deployments, highlighting vendor-driven defaults and potential compliance exposures without succumbing
Novelty & Non-Duplication	Grok	held. Core thesis, framing, and key claims are lifted almost directly from the outlet’s own prior onhealthcare.tech piece on the HHS LEAP awards and agent control, making this a repackage rather than net-ne
Validation	DeepSeek	cleared. The central claim that governance of agentic AI is being neglected in favor of functional performance is strongly supported by current industry deals, regulatory developments, and expert commentary on

Sources cited: 44. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.