You bought an AI documentation tool to get your evenings back. Now you face a harder question than any vendor demo prepared you for: when you use AI for notes, what exactly do you tell your clients, and how do you do it without losing their trust?
It feels like a small disclosure, but it carries real weight. In late 2025, a proposed class action against a major California health system put the stakes in writing. How you answer depends almost entirely on one technical detail most clinicians never think to ask about: does your tool record the room, or does it just help you write?
Below is what the law actually requires, what the major professional bodies now expect, why open disclosure can quietly improve your consent rates, and how to write a disclosure you'd be comfortable defending. This is clinician and practice education, and it isn't legal advice. When in doubt about your state's recording statutes, talk to a lawyer licensed where you practice.
The key line: recording vs. drafting
A vendor can mean two very different things by "AI notes," and they carry very different legal weight.
The first is an ambient scribe. It listens to the live session through a microphone, transcribes the audio, and drafts a note from that transcript. SimplePractice describes its own Note Taker this way: a tool that "transcribes our sessions and then prepares a draft progress note," with audio deleted shortly after and transcripts retained for a limited window (SimplePractice, July 2025). Abridge, the tool at the center of the lawsuit below, similarly records the conversation and uses that audio to draft the note.
The second is a drafting assistant that never hears the room. The clinician types a short scratchpad of what happened, the AI turns that text (plus structured chart fields like diagnosis, goals, and interventions) into a polished draft, and the clinician edits and signs it. There's no microphone, no transcript, and no audio recording in the loop. CoralEHR supports this path: you can type or dictate your notes with no session recording required. Its optional notes assistant on Pro is the first kind of tool, so treat it like one: it listens live, asks the client for consent in the portal, and never stores the audio.
For recording law, that architectural difference decides the question, and the reason comes from the statutes themselves. Wiretapping and eavesdropping laws are triggered by recording a private conversation, meaning capturing the audio. Using software to write up notes afterward is a different act. The McDermott Will & Schulte AI Scribes FAQ frames its consent guidance specifically around tools that "record live doctor-patient conversations" (McDermott Will & Schulte, March 2026). If no audio is captured, the conduct those statutes regulate never happens.
One caveat: this distinction is sharpest under recording law, and it does not get you off the hook ethically. Professional bodies, as we'll see, expect disclosure of AI use whether or not a microphone is involved. And in practice, many compliance teams choose to treat all documentation tools the same way instead of parsing the line case by case. A non-recording tool still needs disclosure. A recording tool needs more, and carries legal exposure a drafting assistant simply doesn't.
What the law actually requires for recording tools
If your tool records audio, all-party consent law is the part to take seriously.
Most U.S. states are "one-party consent" jurisdictions, where one person in the conversation can lawfully record it. But roughly a dozen are all-party (also called two-party) consent states, where every participant in a private conversation must agree before it can be recorded, and recording without that consent can be a criminal offense. One commonly cited list names twelve: California, Connecticut, Delaware, Florida, Illinois, Maryland, Massachusetts, Montana, New Hampshire, Oregon, Pennsylvania, and Washington (Recording Law, 2026). A few more states (Michigan and Nevada among them) are borderline, written as all-party but read narrower by their courts. Sources disagree on the exact count. In practice, if you practice in a strict state, a recording scribe almost certainly requires consent.
The McDermott FAQ goes a step further than minimum compliance. As a best practice, it advises getting "any required consent from patients and caregivers each time an AI scribe tool is used," which rules out a one-time form buried in intake paperwork. It also notes you should get consent from family members or anyone else in the room during a recording (McDermott Will & Schulte, March 2026). For a telehealth session with a partner sitting just off-camera, that is a real operational problem.
What happens when you skip this
In November 2025, a patient filed a proposed class action against Sharp HealthCare in San Diego County Superior Court over its use of Abridge's ambient documentation tool. The complaint alleges Sharp recorded confidential exam-room conversations without obtaining all-party consent, in violation of California's wiretapping law (CIPA) and its Confidentiality of Medical Information Act (CMIA) (Fisher Phillips, December 9, 2025).
The detail every clinician should notice: according to the analysis, patient charts allegedly contained documentation stating patients "were advised" and "consented" to the recording, when, the suit claims, they were not and did not. The proposed class "may exceed 100,000 patients." These are allegations at an early stage, and Sharp had not yet responded when the analysis was published. The lesson holds however the case resolves: a consent box that auto-fills in the chart is different from a consent conversation that actually happened, and a record that says otherwise is its own liability.
What the professional bodies expect beyond the law
Recording statutes set a legal floor. Your licensing ethics set a higher one, and they apply even when no audio is captured.
The American Psychological Association issued ethical guidance on AI in the professional practice of health service psychology in June 2025 (updated December 2025). It is unambiguous that AI use should be disclosed: psychologists are encouraged to disclose AI use to "individuals who are receiving direct care services," and to "obtain informed consent by clearly communicating the purpose, application, and potential benefits and risks" of the tools they use (APA, June 2025). The guidance ties this to Principle E (respect for people's rights and dignity) and frames transparency as a way of letting clients make an informed decision about engaging with your practice at all.
This obligation turns on whether AI touches the client's care, whether or not the tool records. A drafting assistant that never hears the session still uses AI to shape the clinical record, and the APA's logic (tell people what's being used in their care, and let them decide) reaches it. So think of it as two stacked obligations: recording law governs whether you may capture audio, and professional ethics governs whether you must disclose AI at all. A non-recording tool clears the first hurdle automatically but still has to clear the second.
Why honesty works in your favor
One study changes how to think about this conversation.
A JAMA Network Open study published July 22, 2025 looked at how much information you give patients changes whether they say yes to ambient AI documentation. When patients received only basic information about the technology, 81.6% consented. When they received fuller detail about the AI's features, where the data is stored, and the corporate involvement behind it, consent dropped to 55.3% (JAMA Network Open, July 22, 2025).
It's tempting to read that as an argument for saying less. Read it the other way. Patients react to specifics more than to the word "AI": third-party data storage, corporate access, audio leaving the room. The more there is to disclose, the more disclosure costs you. So the cleanest path to a confident "yes" is having less to disclose in the first place.
That is the practical case for choosing a tool whose data story is simple. When the true answer to "where does my session go?" is nowhere; no audio is recorded, and the AI drafts from notes my therapist already typed, the conversation gets shorter and the trust gets deeper. Transparency becomes something you can lead with instead of a risk to manage.
What CoralEHR's AI does and deliberately doesn't do
We hold our own documentation AI to the standard this article argues for, so here are the boundaries in concrete terms.
- It never keeps a recording. You can type or dictate with no AI listening. The optional notes assistant on Pro listens live and drafts the note, but the audio is never stored, and the transcript is deleted when you sign the note, or after 7 days if you don't. Clients signed in to the portal agree before it listens.
- Drafts are preliminary until you sign. An AI draft persists in the record as a preliminary document (FHIR
docStatus=preliminary) until a licensed clinician reviews and signs it. The AI never diagnoses, never decides, and never finalizes a note on its own. - Treatment-plan AI will suggest and never decide. AI treatment-plan suggestions are coming to Pro, and they will be options for a clinician to accept, edit, or reject. The software never adopts a plan on its own. Guardrails forbid the AI from issuing diagnoses, treatment decisions, medication advice, or prognoses on every surface.
- A single, controlled AI path. Requests run over Anthropic's first-party Claude API on one enforced path. Inputs are not used to train models, operational logs carry hashed identifiers instead of session text, and assistant chats are deleted after 30 days.
For client trust, this part matters most: CoralEHR signs a Business Associate Agreement before any protected health information is handled, governing PHI under 45 CFR Parts 160 and 164 (CoralEHR BAA). We're HIPAA-compliant and BAA-backed today. On formal attestation, we don't hold SOC 2 Type II or HITRUST today, and we'd rather tell you that plainly than imply otherwise.
We ask you to give your clients the same framing we use: the AI helps you write, and a human stays accountable for what gets signed. For the longer argument on why that signature can never move to the software, see Why a Clinician Must Sign Every AI-Drafted Note.
How to write your disclosure
You don't need a lawyer to draft a good first version. You need to cover four things truthfully in plain language. (Do have counsel review it against your state's law before you rely on it.)
- What the tool does. Name it. Say whether it records audio or works from notes you type. "I use an AI tool that helps me turn my own session notes into a draft. It does not record or listen to our conversation" reads very differently from "an AI transcribes our sessions," because it is different.
- Why you use it. Clients extend more grace when the reason is about them. Usually the real reason is: less time on paperwork, more attention in the room.
- How their information is handled. Is there a BAA in place? Is their data used to train anything? How long is anything retained? If the tool doesn't record, say so. It's the most reassuring sentence you have.
- Who's accountable. State that you, a licensed clinician, review and sign every note, and that the AI never makes clinical decisions.
Then give a genuine choice, and document the answer through real consent. A chart entry that fills itself in doesn't count. SimplePractice offers a free sample form structured roughly this way (SimplePractice, July 2025), and the APA guidance (APA, June 2025) backs the disclose-and-choose structure. Adapt either to your tool and your state. If your tool does less, your form gets shorter and your conversation gets easier.
For the broader compliance picture around private practice, our HIPAA guide for private-practice therapists covers the surrounding obligations; if you're weighing tools, our CoralEHR vs. SimplePractice comparison lays out the documentation differences side by side.
What to tell clients, in short
"How do I tell clients I use AI?" is much easier to answer when the tool is easy to talk about. If it records the room, you're in all-party-consent territory, you owe a real consent conversation every time, and a class action in California shows what happens when the chart says "consented" but the client never did. If it only helps you write from notes you already took, there's no audio to consent to. You still owe your clients the truth under professional ethics, and telling it with specifics is what earns their trust.
Either way, the steps are the same: tell the truth up front, keep a human accountable for every signed note, and pick a tool whose data story you'd be glad to read aloud to the person across from you.
Want to see how notes work in CoralEHR, with or without the optional assistant? Book a 20-minute demo to walk through it with us, then try it free for 2 weeks.
This article is clinician and practice education, not legal or clinical advice. Recording and consent laws vary by state and change over time; consult an attorney licensed in your jurisdiction and your own licensing board's current ethics guidance before relying on any disclosure or consent process.
Frequently Asked Questions
Do I legally have to tell my clients I use AI for my notes?
It depends on what the AI actually does. If your tool records the audio of the session, you almost certainly have a legal obligation to get consent. Roughly a dozen states are all-party (two-party) consent states where recording a private conversation without everyone's agreement is a crime, and the McDermott Will & Schulte AI Scribes FAQ (March 2026) recommends getting consent every time a recording-based scribe is used. If your tool only helps draft a note from text you already typed, no audio is captured, so wiretapping statutes aren't triggered. Separately from recording law, the APA's June 2025 ethical guidance treats AI disclosure as an informed-consent obligation regardless of the mechanism, so disclosing is both the ethical floor and good practice even when recording law doesn't apply.
What's the difference between an AI scribe that records and one that doesn't?
A recording scribe captures the live audio of your session, transcribes it, and drafts a note from that transcript. That audio capture is what triggers wiretapping and eavesdropping statutes. A non-recording assistant never listens to the room: the clinician types a scratchpad of what happened, and the AI turns that text plus structured chart fields into a draft. There is no microphone, no transcript, and no audio to consent to. CoralEHR offers both: you can type or dictate with no AI listening, or use an optional notes assistant on Pro that listens live, asks clients for consent in the portal, and never stores the audio. Both still warrant open disclosure under professional ethics, but the legal exposure is meaningfully different.
Does being upfront about AI lower the chance my client agrees?
More disclosure does change consent rates, and you should know that going in. A JAMA Network Open study published July 22, 2025 found that 81.6% of patients consented to ambient AI documentation when given basic information, versus 55.3% when given fuller detail about the AI features, data storage, and corporate involvement. Clients respond to specifics, so the cleaner your data practices actually are, the better transparency works for you. A tool that doesn't record, doesn't train on inputs, and signs a BAA gives you a stronger story to tell.
What should an AI-use disclosure to clients actually say?
Cover four things in plain language: what the tool does (drafts notes, and whether it records audio or works from your typed notes), why you use it (to spend less time on paperwork and more time present with them), how their information is handled and protected (BAA in place, no training on their data, retention limits), and that a licensed clinician reviews and signs every note while the AI never diagnoses or decides. Then give a real choice. SimplePractice's free sample form (July 2025) and the APA's June 2025 guidance both model this disclose-and-choose structure; the specifics will be cleaner if your tool does less.
Arhaan Gupta
Co-founder and CEO, CoralEHR
I co-founded CoralEHR, an EHR for private-pay therapy practices, and I set up every new practice with them myself.
LinkedIn