AI Medical Scribe for Psychiatry

The psychiatric documentation burden

Psychiatry is the specialty where the documentation burden is felt most acutely, and it is also the specialty where the visit and the chart most resist automation. An initial psychiatric evaluation is rarely a 15-minute affair. It runs 60 to 90 minutes in the room and another 30 to 60 minutes at the keyboard. The note has to capture a chief complaint, a developmental and psychiatric history that may span decades, a substance-use history, family psychiatric history, a medication history with prior trials and reasons for discontinuation, a full review of systems, a mental status examination, a risk assessment, a differential diagnosis with reasoning, and a plan that often includes a controlled substance and a safety contract. None of this lends itself to a templated dropdown.

Medication management visits are shorter on the clock and longer in the chart than they look. A 20-minute follow-up still has to document target symptoms by domain, response to each medication separately, side effects elicited and denied, adherence, sleep, appetite, sexual function, suicidality, substance use, and the rationale for any dose change. It has to attach to a CPT code that an auditor will accept, which generally means 99213, 99214, 99215, or 90833 added on for psychotherapy time. The fields are the same week after week and patient after patient, and they still take ten minutes to type.

The other cost is harder to quantify and easier to feel. Psychiatric interviews depend on attention. A patient describing a panic attack, a trauma history, or the early stages of a manic episode is reading the clinician's face and pauses for cues about whether it is safe to continue. Typing while they speak is not just slow, it is clinically counterproductive. Every psychiatrist who has tried to type a verbatim quote about suicidal ideation while maintaining eye contact knows the trade-off. The usual workaround is to write nothing during the visit and reconstruct the note from memory afterward, which produces a recall-biased note and pushes documentation into the evening.

Scriber was built so the keyboard never has to compete with the patient. The recorder runs in the background, the clinician listens, and the structured note is generated after the visit ends. None of this is novel framing on its own; what matters is whether the generated note is good enough that the clinician signs it with minor edits rather than rewriting it from scratch. That is the bar this page is going to walk through, section by section, with as little hype as we can manage.

What Scriber captures during a psychiatric visit

A psychiatric encounter has a recognizable shape. Scriber's note generation is structured around it, and the AI is given the section headers explicitly rather than asked to guess at organization.

Chief complaint. Captured in the patient's own words when possible, since direct quotes are clinically and medicolegally useful. Scriber preserves first-person language in the chief complaint field rather than paraphrasing into clinical shorthand.

History of present illness. Onset, duration, severity, course, triggers, modifying factors, associated symptoms, and impact on functioning. For a patient presenting with depression, that means quantifying sleep, appetite, energy, concentration, anhedonia, and suicidality. For an anxiety presentation, it means panic frequency, avoidance, and somatic features. The HPI is generated as prose, not bullet points, because that is what payers and auditors expect for a psychiatry note.

Past psychiatric history and medication history. Prior diagnoses, prior providers, prior hospitalizations, prior suicide attempts, and a medication trial table including drug, maximum dose, duration, response, and reason for discontinuation. This last column is the one that takes the longest to type by hand and the one that determines whether the next prescribing decision is defensible.

Substance use history. Current and lifetime use of alcohol, nicotine, cannabis, stimulants, opioids, benzodiazepines, and other substances, with quantities and last use dates. Critical for both clinical reasoning and 42 CFR Part 2 considerations, which we cover later.

Review of systems. Targeted to psychiatric and medication-relevant systems: sleep, appetite, weight, sexual function, GI, neurologic, endocrine. Scriber captures pertinent positives and negatives rather than emitting an empty templated ROS.

Mental status examination. Appearance, behavior, speech, mood, affect, thought process, thought content (including SI/HI), perceptions, cognition, insight, and judgment. Scriber generates the MSE from observed and reported content during the encounter and flags fields where it had insufficient evidence rather than fabricating a normal finding.

Assessment and plan. Differential diagnosis with reasoning, primary diagnosis, ICD-10 code, and a numbered plan covering medications (with rationale for any change), labs, referrals, psychotherapy modality and frequency, safety planning, and follow-up interval. The assessment is the section auditors read first and the section AI scribes most often get wrong, so Scriber's prompt requires explicit clinical reasoning rather than a list of orders.

Returnability. Follow-up interval and conditions for earlier return are captured as a discrete field so they can be acted on.

Psychiatric note formats Scriber generates

The same recorded encounter can be rendered as several different note types depending on what the visit was for and which payer or institution is reading it. Scriber generates each as a distinct format rather than a generic SOAP note with the words rearranged.

Initial psychiatric evaluation

The full intake note. Includes chief complaint, HPI, past psychiatric history, medication history with prior trials, substance use history, medical history, family psychiatric history, social and developmental history, ROS, full mental status exam, risk assessment, differential diagnosis with clinical reasoning, primary and secondary diagnoses with ICD-10 codes, and a treatment plan. Maps to CPT 90791 (psychiatric diagnostic evaluation) or 90792 (with medical services) depending on whether prescribing decisions are part of the visit.

Follow-up med-management visit

Shorter and structured around interval change. Documents target symptoms by domain, response to current medications, adherence, side effects elicited and denied, relevant ROS, focused mental status exam, current diagnosis with status (improved, stable, worsening), and the prescribing decision with reasoning. Maps cleanly to 99213, 99214, or 99215 with optional 90833 add-on if psychotherapy was incorporated. Scriber records the time-based elements and total visit time when those drive billing.

Mental status exam

Standalone MSE format for situations that require it as a discrete document, such as competency evaluations, capacity assessments, or pre-procedure psychiatric clearance. Covers the eleven canonical domains and refuses to populate fields it has no evidence for.

Suicide risk assessment

A structured risk note covering chronic and acute risk factors, protective factors, current ideation with detail on intensity, frequency, duration, plan, intent, access to means, prior attempts, and a documented risk stratification with the reasoning that supports it. Includes a safety plan if one was created in the visit. This is the document that gets pulled in any adverse event, so Scriber treats it as a separate generation rather than a checkbox in the main note.

Treatment plan

The longitudinal treatment plan rather than the per-visit plan. Includes problem list, measurable goals and objectives, interventions tied to each goal, frequency and duration of treatment, and review dates. Generated in the format payers and accreditors expect.

Discharge summary

Episode-of-care summary for patients leaving the practice or completing a course of treatment. Reason for discharge, course of treatment, response, current medications, current diagnoses, residual symptoms, recommendations for ongoing care, and provider follow-up arrangements.

Court-ready documentation

For visits where documentation is likely to be subpoenaed (custody evaluations, disability determinations, fitness-for-duty), Scriber generates a more formal note with explicit sourcing language ("patient reports," "collateral indicates," "observed during interview") and avoids speculative language. This is a configuration choice on the visit type rather than a separate product.

Telehealth psychiatry with Doxy.me + Scriber

Telehealth is now a majority of outpatient psychiatry visits in many practices. Most AI scribes treat the telehealth platform as a separate concern and rely on the clinician to share their browser tab audio manually, which works but is a small failure point at the start of every visit. Scriber ships a real Doxy.me integration.

The workflow is:

1. Open Scriber. Click your Doxy.me room. The Doxy waiting room loads inside Scriber, signed in to your existing Doxy account.
2. The patient joins their session through the same Doxy link they have always used. Nothing changes for the patient.
3. When the call connects, Scriber captures audio from the call tab automatically. The recording indicator goes green.
4. End the visit when the conversation is finished. Scriber stops the recording and begins transcription and note generation. A draft note is in your queue before you have logged the visit in your EHR.

The differentiation is that this is a real integration, not a copy-paste landing page or a Zoom-share workaround. The Doxy room runs inside Scriber, audio capture is automatic via tab share, and there is no separate recorder to start. That matters for psychiatric visits because the most common documentation failure mode for telehealth is forgetting to start the recorder during the first minute of intake, which is exactly when the chief complaint and mood are being established. With direct integration, the recorder is on by the time the patient appears.

There is also a smaller benefit that adds up over a full clinic day. With separate-tab telehealth, the clinician has Doxy open in one window, Scriber in another, the EHR in a third, and possibly a PDMP lookup tool in a fourth. Window-switching during a session is itself a clinical distraction. Running the call inside Scriber removes one window from that stack and keeps the recorder visible without taking attention away from the patient's video.

Doxy is the first telehealth platform we shipped a real integration for. It will not be the last, but everything claimed on this page is something we tested in production this quarter. Competitor scribes that list "AI scribe for Doxy.me" pages without a real integration will, in our experience, ask the clinician to share a tab and start a separate recorder; that is fine, it just is not the same product.

Accuracy on psychiatric terminology

Generic speech-to-text gets around 95% of words right, which is excellent for podcasts and miserable for psychiatry. The 5% it gets wrong is concentrated in exactly the words that matter: medication names, diagnostic terminology, and dosing. "Lamotrigine" becomes "lamentation," "clozapine" becomes "closer pin," "100 milligrams BID" becomes "100 milligrams beat" — and the reviewing clinician now has to either fix it or trust a number they did not say.

Scriber routes audio through one of seven transcription engines, two of which are tuned for clinical vocabulary: Microsoft AI (Azure Speech with the medical conversation model) and Speechmatics (with its medical language pack). On our internal benchmark of psychiatric encounters, both engines come within 1% word error rate of human transcription on medication names and dosing. The remaining engines are kept available because they trade accuracy against speed or cost differently and because some clinicians prefer the cadence of one transcript style over another.

The fastest engine in the rotation is xAI, which returns transcripts in roughly three seconds for a typical visit. That speed is useful for quick-turnaround review during back-to-back visits. The medical-vocabulary engines are slightly slower (typically 5 to 8 seconds for a 20-minute visit) but more accurate on terminology. You do not have to choose every time. Set a default engine in account settings; override per visit when you want to.

Diagnostic terminology and ICD-10 F-series codes are handled at the note-generation layer rather than the transcription layer. The model is given the F-series code list and the DSM-5 criteria as part of the prompt context for psychiatric notes, so when the transcript says "patient meets criteria for major depression, recurrent, moderate," the generated note attaches F33.1 rather than guessing at F32.1 or F33.9. Where the transcript is ambiguous between two adjacent codes, Scriber emits the code it has the most evidence for and flags the ambiguity in a reviewer comment rather than committing silently.

Word-error-rate numbers are the floor, not the ceiling. The harder accuracy problem in psychiatry is structural: getting the right symptom into the right field, not getting the right syllables out of the right speaker. We benchmark structural accuracy separately by sampling generated notes and having a clinician score field-level correctness. Numbers from the most recent run are available on request.

HIPAA + state mental-health-specific privacy

Psychiatric records are subject to HIPAA like any other health record, and to additional protections that other specialties do not have to think about as often.

Federal HIPAA. Scriber signs a Business Associate Agreement at signup, before any audio is processed. Not enterprise-gated. The BAA covers Scriber as the BA and the clinician's practice as the covered entity. Audio is captured for transcription and discarded; we do not retain audio after note generation completes. Transcripts and notes are encrypted at rest with AES-256 and in transit with TLS 1.3. Infrastructure is hosted on Microsoft Azure under our enterprise BAA with Microsoft. We do not train models on patient data.

42 CFR Part 2. Records that identify a patient as having a substance use disorder and that are created by a federally assisted SUD program are subject to additional protections beyond HIPAA, including stricter consent requirements for disclosure. If your practice falls under Part 2, the same Scriber-generated notes can be tagged at the visit level so they carry a Part 2 marker into your EHR, but the legal compliance work — patient consent forms, disclosure logging, qualified service organization agreements — happens in your practice, not in Scriber. We are not a Part 2 program ourselves. We are a business associate that handles the records you create.

State minor consent laws. Many states grant minors the right to consent to mental health treatment without parental involvement starting at 12, 14, or 16. Notes for those visits cannot be released to parents under the state consent statute even if HIPAA would allow it. Scriber does not enforce these rules on your behalf, but the visit-level tagging that supports Part 2 also supports flagging minor-consent records so they are visible in your release-of-information workflow.

The short version: BAA at signup, audio not retained, AES-256, Azure with BAA, no training on PHI. The longer version is your security review document, which is available on request before signup.

Common psychiatrist questions

Controlled-substance documentation. Scriber generates the visit note that supports the prescribing decision. It captures the indication, the prior trials and responses, the rationale for the controlled substance, the discussion of risks and side effects, the PDMP check if you reference one in the visit, and the agreed-upon follow-up. The actual prescription is entered in your EHR or e-prescribing system per DEA, EPCS, and state requirements. Scriber does not e-prescribe and does not transmit prescriptions. The boundary is intentional.

Audit-readiness. A psychiatric chart is auditable when the note supports the level of service billed and the medical decision making is documented in prose, not abbreviations. Scriber generates notes that meet AMA E/M guidelines for the time-based or MDM-based code you select, and the prose structure of the assessment-and-plan section is the part that auditors look at hardest. The note is your note — review it before signing — but the structural defaults are tuned for audit defensibility rather than brevity.

Integrating with my EHR. The honest answer for most EHRs in 2026 is copy-paste. Scriber generates the note in plain text and rich text, and you paste it into your EHR's progress note field. This is how most psychiatric practices using AI scribes actually work, and it is not a workaround — it is the lowest-friction path that keeps the EHR as the system of record. For Doxy.me telehealth visits we have a real direct integration as described above. For SimplePractice, TherapyNotes, and Valant, integrations are on the near-term roadmap; we will list them as direct only when they are shipped and tested.

What about visits with collateral informants? Initial evaluations frequently include a parent, spouse, or other collateral. Scriber does not currently distinguish speakers in the generated note text, but the underlying transcript preserves speaker turns, and the note generation prompt is given hints about who is in the room. The result is reasonable on most multi-speaker visits and imperfect on a few. Where speaker attribution matters medicolegally — for example a custody evaluation — we recommend reviewing the transcript alongside the note before signing.

What about visits where the patient asks not to be recorded? Document the request, do not record, and use Scriber for the visits where consent is given. There is nothing in the product that requires every visit to be recorded, and the practice's consent process should be the gating mechanism. We do not see consent as a feature.

Try Scriber for psychiatry

Scriber is built for outpatient psychiatry, not pitched at it. The note formats, the medical-vocabulary transcription, the Doxy.me integration, and the BAA-at-signup all exist because the alternative — generic AI scribes pointed at psychiatric encounters — produces notes that need so much editing they save no time at all.

The 7-day trial lets you test it on real visits before committing. The BAA is in place from the moment you sign up. If it does not save you an hour a day in the first week, do not subscribe.

Start a free trial — set up takes about three minutes, including your first telehealth visit.