About NHID-Clinical

NHID-Clinical is an independent, open governance framework for non-human actors — AI voice agents — that operate under delegated authority and cross the boundary between healthcare organizations.

It applies at one specific moment: when an AI voice agent acting for one healthcare organization contacts another. A provider's agent calling a payer about eligibility, claim status, or prior authorization. The model behind that agent, its training and its clinical reasoning sit inside the organization that runs it, and are explicitly outside this framework's scope. NHID-Clinical applies at the crossing.

1. Who is this agent?

   Identity of the non-human actor.

2. Who authorized it?

   The delegating organization, and the authority it granted.

3. What scope does it hold?

   The bounds of that authority, and whether they narrow across hops.

4. Did it disclose that it is automated?

   And before or after protected data was exchanged.

5. What record exists of what happened?

   An auditable trail the receiving organization can review.

Who authorized the agent? is the question the framework is organized around. A disclosed agent may still be unauthorized; an undisclosed one may be perfectly authorized. The two are separate properties, and the framework keeps them separate.

NHID-Clinical was created and is maintained by Brianna Baynard, an independent AI governance and security researcher (AIGP, ISC² CC).

It grew out of healthcare payer operations. She worked live eligibility, claims and prior-authorization phone lines, and was on the receiving end of hundreds of calls from AI voice agents acting for provider offices — including agents that disclosed their automated status only after protected health information had been exchanged, or not at all. NHID-Clinical documents those failure patterns and defines a testable baseline for addressing them.

LinkedIn · GitHub · contact@nhid-clinical.org

The distinction matters, so it is stated rather than blurred.

AOriginating with NHID-Clinical

Impersonation Latency — the elapsed interval between interaction start and the point at which a non-human actor discloses its non-human identity to the human recipient. Reported in seconds and in conversational turns, with turns as the normative unit.

It measures disclosure timing only. It does not establish that impersonation occurred, determine intent, detect an impersonator, prevent impersonation, establish authentication, or establish authorization. Disclosure is observed through a transcript, so a reported figure is only as accurate as the transcription path that produced it.

This is not an established industry term. It is this project's name for a specific measurable interval, and it should not be presented as standard vocabulary.

BEstablished work, not ours

Delegated authority, agent identity, and scope attenuation are prior art. This project did not invent them; it applies them to a healthcare-specific case and is designed to compose with authenticated-delegation approaches and with SPIFFE/OAuth-style stacks.

• NOT

  A standard

  No standards body has adopted it and no formal standards process has started.

• NOT

  A certification

  Not a badge, a licence or a credential. It certifies nothing and no one.

• NOT

  A regulatory requirement

  And not an officially endorsed framework of any agency.

• NOT

  HIPAA-compliant software

  It confers no compliance status. It is mapped to NIST AI RMF and ISO/IEC 42001, and designed to support the disclosure obligation described in EU AI Act Article 50(1) — mapped, not certified.

• NOT

  Production-ready infrastructure

  It is a reference implementation.

• NOT

  Adopted

  It has no adopters, customers, pilots or completed audits. Shadow-evaluation partners are being sought; none is claimed.

• NOT

  A clinical safety system

  It does not govern models, bias, or clinical reasoning.

NHID-Clinical was submitted as a public comment to NIST (NIST-2025-0035-0026). A public comment is a submission to a docket. It is not adoption, endorsement, or approval by NIST, and should not be read as any of those.

The project's own boundary on what may and may not be claimed is kept in the repository as claim-boundaries.md, and a test fails if the documentation and that file ever disagree.

Code is Apache-2.0. The specification and documentation are CC BY 4.0. Both are published openly, and criticism of the framework is more useful to it than adoption of it.

Read the specification (v1.3) · Technical reference · What is still open