How we protect your clients’ records

What we do today, what we haven’t done yet, and how to ask us anything.

Where we stand today

In place
You sign every note. Margent never signs, sends a note or sends a claim on its own.
In place
Data is encrypted when stored and when sent.
In place
Data is stored in the US, on Amazon Web Services (AWS).
In place
Recordings are deleted within 24 hours. Live audio isn't stored by Margent at all.
In place
We never train AI on your notes, transcripts or edits. Your style preferences stay inside your practice.
In place
A log records who did what, and the app has no way to edit or delete it.
Before you start
We sign a BAA with you before you use real client information. Until then, trials run on sample sessions.
Not yet
Not every vendor has signed a BAA. AWS (hosting) and AssemblyAI (speech-to-text) have. Our AI writing vendor and our insurance-eligibility vendor (Stedi) have not, so Margent runs on sample data only until they do.
Not yet
No SOC 2 report yet. It is planned: Type I, then Type II. No badge goes here before a report exists.
Not yet
No outside security test yet. A penetration test is planned before the first pilot with real client information.
Not yet
No live EHR connection yet. Copy or export works with any EHR today. Direct filing is tested in Epic and Cerner test systems only, not at any practice.

What happens to a session

Consent. You record the client's consent before you start. Margent checks the rules for the client's state.
Record. Audio goes to speech-to-text over an encrypted connection. Margent doesn't keep it.
Write. Margent writes the note from the transcript. Every sentence links to the moment it was said.
Sign. You review, fix anything flagged, and sign. Nothing leaves without your signature.
File. You copy or export the note into your EHR, which stays the official record.

Before the first pilot, your practice chooses how long transcripts are kept, from 24 hours to 90 days, and our servers delete them when that window ends. When your contract ends, we delete your data.

Extra care for sensitive records

Substance-use records (Part 2): Mark a program or a client as Part 2 once. Consent is recorded before each session and kept in the log.
School records (FERPA): Each program is set to HIPAA, FERPA or both, so school records stay separate.
State AI laws: Margent applies the consent rule for the state the client is in. Colorado, for example, needs written consent first.

Technical details

42 CFR Part 2 (2024 final rule, in force Feb 16, 2026)

One consent can cover treatment, payment and operations. Counseling notes need their own. Breach notification now applies.
Part 2 flag per program and per client, set once.
Consent, including Part 2 consent, is captured on the session before recording. Consent and changes go to the append-only audit log.
Signing is blocked until flagged sentences are resolved.
Before the first pilot: the Part 2 redisclosure notice and security label on every workspace export. Planned: the same on EHR filings, and an accounting of disclosures.

FERPA (school-based programs)

Records kept by school-employed clinicians are education records, not HIPAA records. Summaries placed in a school file become FERPA records too.
Consent is recorded per session, in English or Spanish.
Planned with school-based partners: separate school-facing summaries and guardian consents.

State laws on AI in mental health care

Colorado · HB 1195 (2026): Written client consent before AI records or transcribes a therapy session. Recording is blocked until a written e-consent is on file for Colorado clients. Declines switch the session to dictation.
Illinois · Wellness and Oversight for Psychological Resources Act (HB 1806): AI may help with administrative and documentation work but may not make therapeutic decisions or act as the therapist. Written informed consent before sessions are recorded or transcribed. Written consent workflow, and every AI output is a draft a licensed clinician reviews and signs. Nothing client-facing acts as a therapist.
Nevada · AB 406 (2025): AI may not present itself as providing mental or behavioral health care. Providers may use it for administrative support. Calls & referrals introduces itself as an automated assistant, never gives clinical advice, and hands off to staff and 988 when risk comes up.
Maine · LD 2082: Disclosure to clients when AI is used in their care, with the clinician responsible for the record. Plain-language disclosure in the consent form and client summaries, plus a per-client log of which AI features touched their record.

Our reading as of October 2026. Counsel has not yet reviewed our consent language. This is not legal advice; your counsel has the final word, and you can tighten any rule in your note rules.

We support clinical judgment. We never replace it.

AI writes drafts. You decide what is true, and you sign.
A second check compares every sentence with the session. Anything it can't trace is flagged, never hidden.
Letters, tasks and note edits that Ask suggests wait for your approval.
In the intake text assistant, a fixed word list (not AI) checks for crisis language first, and hands the person to staff.
If live AI is down, any sample output is clearly labelled. We never pass it off as real.
When the AI model changes, we test again and publish the results before calling it better.

See our test results, misses included

For your IT or security reviewer

The detail a security review asks for. We fill in your questionnaire too.

Technical details

Encryption

Encrypted at rest in AWS (S3 and DynamoDB).
TLS in transit, including to the AI and speech-to-text vendors.
Customer-managed encryption keys (KMS): being added before the first pilot.

Where data lives

Stored in AWS us-east-1.
US-only processing is available for the language model.
Every vendor is listed with its BAA status on our subprocessors page.
Each AI request carries only the chart context that task needs.

Access

Roles: owner, admin, supervisor, clinician, biller and front desk.
Your own admins manage seats and invites.
Before the first pilot: two-step verification with an authenticator app, required for owners and admins, and sign-out after 30 minutes without activity.
Single sign-on (SAML or OIDC) and SCIM provisioning: on the roadmap.

Audit log and retention

Append-only, per practice: signatures, edits, AI actions and approvals, with who did each.
Consent, retention, Part 2 and FERPA settings are set per organization and per program. Every change is logged.
Before the first pilot: your organization picks a transcript retention window of 1, 7, 30 or 90 days, enforced on our servers. Deletion at contract end is written into the customer agreement.
We will publish our AI vendor's retention terms before any client information is processed.

Compliance roadmap, in order

Before any client information: a BAA with our language-model vendor, a HIPAA risk analysis, and a BAA with you.
Then a third-party penetration test of the web app, APIs and browser extension.
Then SOC 2 Type I, then Type II.

Documents we can share

Security controls overview: what is built, and where each control is still being hardened. On request.
Business associate agreement: our template, or we review yours. Signed before a pilot touches real client information.
Security questionnaires (HECVAT, SIG Lite, CAIQ or your own), answered item by item with what is built and what is planned.

Questions, or found a problem?

Email security@getmargent.com for our security overview, our BAA template, or help with your security questionnaire. To report a vulnerability, send steps to reproduce. A person reads every report and tells you how serious we rate it.

Rules for security researchers

How to report

We haven't published a PGP key yet. If the details are sensitive, send a short first message without them and we'll set up a secure channel.
We send updates until it's fixed.
Safe harbor: good-faith research within these rules won't lead to legal action from us. We'll say so in writing if you ask.

In scope

getmargent.com, the web app, our APIs and the browser extension.

Please don't

Access real client data, degrade service, use social engineering on staff, or test against customer EHRs.