Are you like the AI vendors that talk to patients?

No. They are the AI. We are the layer their output passes through, and yours, and the record both leave behind.

Do we have to change our model or our vendor?

No. The gate sits between your AI and the patient. Your model, your vendor, your policies.

What happens when the gate holds something?

It waits for a person you name, with the reason attached. Until they act, the patient hears only what your clinic wrote in advance, nothing model-written.

Does the AI get a second chance?

Once. It is told why the first draft stopped, in your rule's own words, and the second draft is checked exactly like the first. Both are kept, and a sample of turns allowed on retry is set aside for review.

Does this make us compliant with AB 3030 or AB 489?

It gives you the controls those statutes describe, disclosure on every generative message, no implied license, a human path, and a record that shows they ran. Whether you are compliant is a determination your counsel makes; we make it answerable.

Can our governance committee see it?

Yes. Read access through your identity provider, the monthly report, and the audit export.

Do you work inside Epic?

Our write path is proven against Epic's developer sandbox; a live instance waits on a clinic that grants access.

PATIENT-FACING AI

Do you put AI in front of patients?

If you already run patient-facing AI, we govern it and give your reviewers the evidence. If you are about to, we run the program with the governance built in.

SOC 2 Type II across all five trust services criteria · HIPAA · model providers under BAA with zero data retention

SOC 2 Type II

HIPAA

Patent Pending

THE OBLIGATION IS ARRIVING

The rules for AI that speaks to patients are no longer hypothetical.

Jan 2025 · California AB 3030
generative-AI patient communications need a disclosure and a way to reach a human.
May to Aug 2025 · Utah, Nevada, Illinois
mental-health chatbot disclosure; AI barred from delivering mental or behavioral health care outright.
Sep 2025 · Texas SB 1188
disclosure when AI is used in diagnosis or treatment.
Sep 2025 · Joint Commission and CHAI guidance
voluntary guidance: formal AI governance, local validation, ongoing monitoring, patient disclosure.
Jan 2026 · California AB 489
AI may not imply it holds a clinical license.
Jun 2026 · Joint Commission certification
a voluntary Responsible Use of AI certification opens, built on that guidance.
2026 to 2027 · more states, more committees
the pattern is set.

Jan 2025 · California AB 3030
generative-AI patient communications need a disclosure and a way to reach a human.
May to Aug 2025 · Utah, Nevada, Illinois
mental-health chatbot disclosure; AI barred from delivering mental or behavioral health care outright.
Sep 2025 · Texas SB 1188
disclosure when AI is used in diagnosis or treatment.
Sep 2025 · Joint Commission and CHAI guidance
voluntary guidance: formal AI governance, local validation, ongoing monitoring, patient disclosure.
Jan 2026 · California AB 489
AI may not imply it holds a clinical license.
Jun 2026 · Joint Commission certification
a voluntary Responsible Use of AI certification opens, built on that guidance.
2026 to 2027 · more states, more committees
the pattern is set.

Most of the statutory duties land on the provider or the company that deploys the AI; a few bind whoever builds it. The guidance is addressed to the same people. Your governance committee is now the gate.

THE GOVERNED MESSAGE

Nothing reaches a patient that your policy would not allow. Every decision is on the record.

AI draft

Your latest reading is high and could mean a serious heart problem. You should change your medication dose today.

The gate

Allow

Hold for a person

Block

Delivered to the patient: the clinic's holding line

Thanks for sharing today's reading. Your care team is reviewing it and will follow up with you today.

DisclosureThis message was generated by AI and checked against your clinic's rules before it was sent. Reply HELP to reach your care team.

The record

Policy and version: Patient messaging · v4.2
Model and version: BAA-covered model · v2
Draft: Stopped before sending
Outcome: Held · implied dose change
Reviewer: R. Chen, RN · refused the draft, phoned the patient · 14:21 PT
Timestamp: 2026-09-15 14:09 PT

Your rules, not ours.

Policies are written by your clinical or compliance owner, versioned, and every block traces to one.

Held, then a person.

A stopped draft does not go. The clinic's holding line goes out at once and a person decides, with the original draft and the reason on the record.

A record your committee can read.

What was proposed, what was delivered, who held it and who released it, line by line.

MODEL ROUTING

PHI never leaves the safe zone.

PHI REQUEST

Heart Rate: 106 bpm

Patient ID: 4287

HIPAA Tier

BAA-Covered Model Providers

HIPAA-compliant

NON PHI REQUEST

Summarize ECG Trends

General Tier

Broader AI Models

No Patient Data

PHI requests

PHI REQUEST

Heart Rate: 106 bpm

Patient ID: 4287

HIPAA Tier

BAA-Covered Model Providers

HIPAA-compliant

Non-PHI requests

NON PHI REQUEST

Summarize ECG Trends

General Tier

Broader AI Models

No Patient Data

THE GATE

Three outcomes on every message.

AI OUTPUT

↓

Governance layer

PHI-Safe Model Routing

Human-in-the-Loop Review

LLM Cost and Usage Transparency

Compliance Gates

Agent Types with Scoped Authority

Full Audit Trail

↓

Allow

Meets your policy. Sent as written.

Hold for a person

The draft is written again once, with the gate's reasons, and checked the same way. If that fails, the message is surfaced for someone you name; until they act the patient hears only what your clinic wrote in advance, nothing model-written.

Block

The patient opted out or never consented on this channel. Nothing is sent, and the attempt is recorded.

OBSERVE BEFORE ENFORCE

Two weeks in observe mode, then a report your reviewers can sign off on.

01

Connect.

Your AI's patient-facing output flows through the gate in observe mode. Nothing changes for your patients.

02

Watch.

Two weeks of what the gate would have allowed or held, by policy, over the turns it answered. Unanswered turns are shown separately, with the conversation traces behind each count.

03

Decide.

You turn enforcement on policy by policy, and the observe report goes to your committee as the evidence for the deployment decision.

We show what the gate did; your reviewers judge.

WHAT YOUR REVIEWERS GET

What your compliance and IT reviewers get.

The governance report

monthly, per application, incidents with dispositions, the active policy set. A report your committee can put on the record.

The audit record

exportable; every decision traceable to inputs.

Access their way

single sign-on through your identity provider; roles; patient text visible only by explicit grant.

The paperwork

BAA, SOC 2 Type II report, subprocessor list, data-flow diagram, on request.

THE AUDIT RECORD

Every decision traceable to its inputs.

What was proposed, what was sent, which policy and model ran, who held it, and when. Exportable, on request.

Audit trail · message to pt_0000

Verified

1.Draft received

AI draft in, checked against your policy

policy: Patient messaging · v4.2

model: BAA-covered model · v2

channel: SMS · pt_0000 · PHI tagged

hash: 0000x00 · verified

2.Policy checked

disclosure, license, scope · v4.2