Nextvisit AI

A multi agent documentation core inside what is now a practice platform.

The models are the engine. AriaMD drafts from the current encounter, references prior records, tracks clinical change and assembles a patient timeline, and a separate agent handles pre visit intake, scoring PHQ-9, GAD-7 and C-SSRS before the clinician sits down. The model was built and tuned on psychiatry, therapy and addiction medicine encounters rather than adapted from a general scribe. The company is physician founded, with a practicing psychiatrist co founder who still runs his own outpatient practice on the product.

What has changed is everything around it. The platform now sells e-prescribing through Surescripts with controlled substance support, an insurance and billing module covering eligibility, claims, denials and a ledger, built in telehealth video, scheduling, analytics and reporting, and an automations engine. Remove every model and a behavioral health practice system remains: a prescriber can still write scripts, a biller can still work claims, a clinician can still run a video visit. That is a real product a buyer would pay for.

The documentation and intelligence capabilities do not exist without the models, and they are the reason the platform is bought. Both things are true.

The boundary is drawn firmly in both directions, and the threshold that would complete it is missing.

What the system may do and may not do are both published. AriaMD drafts, suggests and surfaces patterns. A clinician reviews, edits and signs every note, and the company states the system cannot sign, file or finalize a chart. The mobile companion supports recording, editing and review before anything reaches the main workspace.

One published sentence could read as a promise of error free output, and the mechanism published beside it keeps it testable: the company states the assistant will not invent medications, diagnoses or clinical events that were not in the encounter. That is a claim about what gets added rather than what gets left out, and a reader can check it. Each draft section displays the encounter context that produced it, drafts are stated to come from the current visit audio and the existing chart rather than from general training, a pre signature pass flags omissions and inconsistencies, and nothing enters the chart without a clinician signing. A clinician can therefore test the claim line by line against the session. The same sentence with none of that behind it would read as a promise the output is error free, and a reader who believed it would skip the review step. It is still a claim to test rather than a guarantee, because no figure supports it.

What is missing is the operating point. No confidence threshold is published, nothing states when the system declines to draft or routes a case for a second look, and no acceptance rate or edit burden figure is published. The marketing emphasis on notes arriving ready to sign sits in tension with wanting a clinician to read carefully, and there is no published measure of how often that tension bites.

How the system is built is described in useful detail; what it is built from is not identified.

What is published is the shape of the system rather than its components. Multiple models collaborate, agents reference prior records and track clinical change, the assistant is described as carrying more than 37 tools across chart edits, cited research, prior authorizations and eligibility, and drafts are stated to be grounded in retrieved encounter context with the source displayed on each section. That tells a buyer how the system is meant to behave, which is more than a proprietary label.

Nothing is named. No model, model family or foundation model provider appears anywhere, no model card is published, no version is given for any component, and no release or update cadence is stated. The company says it built and tuned its own model on specialty encounters, which leaves open whether a commercial foundation model sits beneath that work.

The one number attached to the technology is note quality above 98 percent, stated as internal scoring across active clinicians. The internal label is honest, but no methodology, reference standard or denominator accompanies it, so it cannot be tested.

Nothing in the chain is identified, and the one disclosure that exists points the other way.

No model or model family is named, no foundation model provider is named, no hosting provider is named beyond a statement that infrastructure is United States based and carries its own business associate agreements, and no sub processor register is published. The company states it built and tuned its own model on psychiatry, therapy and addiction medicine encounters, which suggests proprietary work sitting on top of or instead of a commercial foundation model, and does not say which.

On the training boundary, the disclosure does not favor the vendor. The published position is that the company does not train on protected health information, and that it trains on other customer data by default with an opt out. The first half is a real commitment. The second is a carve out, and it bears directly on what reaches a model and what happens to it there. Which party performs that training, and whether any third party model service is involved in it, is unstated.

Parties in the data path are named elsewhere on the site without being presented as a supply chain disclosure. Surescripts handles prescribing, Change Healthcare and Availity handle claims, Stripe handles payment, Twilio handles messaging, and Zoom and Doxy.me appear on the telehealth capture path. That is a partial map a buyer can start from, but it is an integration list rather than a sub processor register, and it says nothing about which service sees session audio or note text.

A named deployment with dated figures, and no controlled evaluation anywhere.

The case study is the substantive item. Dr Varinder Rathore, a solo outpatient psychiatrist at Restore Wellness Psychiatry, is named, the study is dated 1 September 2025, and the measurement windows are stated: primary metrics inside thirty days, retention over the following year. The figures are 3.4 hours a day returned, revenue up 30 percent attributed specifically to same day billing rather than to volume, and patient retention up 19 percent over the year. The mechanism is stated plainly enough to test, which is that charts code at departure and reach billing the same day instead of lagging 48 to 72 hours.

Platform figures now carry denominators where they previously did not. Time saved is given as 2.5 hours a day average, measured across active clinicians over 2024 to 2025. Claim first pass acceptance is 98 percent across 2.3 million claims. Draft turnaround is a median under 60 seconds for visits under 45 minutes. Total volume is 2.3 million encounters across private practice, groups and academic sites, at more than 300 practices.

What is absent is any control. The case study is a single provider before and after with no baseline figures published, no comparison practice and no independent involvement, so the 30 percent revenue change cannot be separated from anything else that happened in that practice that quarter. The platform averages are internally measured and internally reported. No study, no peer reviewed work and no third party evaluation exists.

A categorical commitment on protected health information, with a training carve out beside it that needs a closer look.

The published sentence has two halves and both matter. The company does not train artificial intelligence on protected health information. It also trains on other customer data by default, and a customer can opt out. That wording appears identically on the security page, the responsible AI page and the home page FAQ, so it is a deliberate position rather than a slip, and stating it plainly is better than hiding it. But default on with an opt out puts the burden on the customer to find the setting, and the buyers here are solo clinicians and small practices with no procurement function. What counts as other customer data, once protected health information is excluded, is not defined: whether it means templates, prompts, usage telemetry, deidentified transcripts or something broader is the question the sentence leaves open, and the answer decides whether this is a minor telemetry provision or a substantial one.

Retention is partly specified. Recordings are deleted on signature, which is a schedule rather than a control, and retention elsewhere is described as configurable without the ranges or defaults being published. No retention period is stated for notes as distinct from recordings.

No deidentification practice is described anywhere, and that is what would settle what the training carve out actually covers.

One question follows from the product's own history. The company states it built and tuned its model on psychiatry, therapy and addiction medicine encounters from the start, so a corpus of exactly this content already existed before any customer commitment applied. What that corpus was and on what consent basis is unaddressed.

Clarity on business associate status is the strength here; the agreement itself is not published.

The company says it signs a business associate agreement with every customer, describes the agreement as available by default on the security page, lists it as included on the $49 entry tier, and repeats it in the footer and on the demo call to action. That settles whether an agreement is on offer, and settles it at the tier that matters: the buyers here are solo psychiatrists and small practices with no procurement function, so whether the agreement arrives by default or only on request determines whether it exists at all. By the company's account, it arrives by default at every tier.

It sits on a substantive security document rather than a bare claim: encryption in transit and at rest, configurable retention with recordings deleted on signature, single sign on, provisioning and audit logs, and United States hosting on infrastructure carrying its own business associate agreements.

The agreement cannot be seen before contracting, and its subcontractor terms are not published.

One scope question is specific to this product rather than generic. The company sells a purpose built medication assisted treatment note type and a whole solution page for addiction medicine, which places substance use disorder treatment records in scope. That content sits under a separate federal confidentiality regime whose consent and redisclosure rules differ from the general privacy rule, and a standard business associate agreement does not by itself satisfy it. Nothing published addresses it.

The attestation is now named as the vendor's own, with the report itself still behind a request.

SOC 2 Type II is stated on the security page and in the footer, and the company says in its own FAQ that an independent third party audits those controls annually. That is presented as the company's own attestation rather than attached to a hosting provider. HITRUST r2 is stated as in progress with a target of the third quarter of 2026, a dated commitment rather than an aspiration.

What is missing is the scope and the clock. The audit period and the assessor are not named. The trust portal is referenced from both the security page and the home page, and the home page describes it as open on request, so the report is not reachable without contact. Nothing states whether the scope covers the iOS companion application and the Chrome extension as well as the main platform, which matters here because both capture or handle session content outside the browser workspace.

Supporting controls are published: TLS 1.3 and AES-256 in transit and at rest, single sign on, SCIM provisioning and audit logs.

The sensitivity of the content raises the bar. This product records psychiatric, therapy and addiction medicine sessions, and a co founder has published buyer guidance naming SOC 2 Type II specifically as the standard mental health data warrants.

No device claim and no pathway attaches to documentation, with two exposures.

The positioning is appropriate and consistently stated. The system drafts, the clinician reviews, edits and signs, and the company states the assistant cannot sign, file or finalize a chart. Documentation of a session the clinician conducted is support whose basis the clinician witnessed, so no device position is required.

The risk language exposure is the first. The platform tracks risk factors across visits, ships a Columbia Suicide Severity Rating Scale screener inside its pre visit intake agent, and the company states that risk language usually lands in the first draft. In behavioral health, risk language means suicide and self harm, the highest stakes inference in the specialty, and a generated risk sentence is the sentence a later reviewer will treat as the clinician's own assessment. Source display on each draft section goes some way to answering it. What is not addressed is what the system does when a session contains ambiguous risk content, whether generated risk language is visually distinguished for confirmation rather than presented as drafted text, and what happens when the intake agent scores a screener positive before any clinician has seen the patient.

The second is regulatory rather than clinical. The company sells a purpose built medication assisted treatment note type and a dedicated addiction medicine solution covering buprenorphine, methadone and naltrexone. Documentation of substance use disorder treatment falls under a separate federal confidentiality regime with its own consent requirements and specific restrictions on redisclosure, stricter than the general privacy rule. Nothing published addresses it, and a vendor shipping a purpose built note for that setting sits squarely inside its scope.

A responsible AI page with no measurement on it, against a specialty where the measurement would be cheap to produce.

What is published is a governance posture in policy language: model behavior is monitored, risks are reviewed, controls improve over time, and clinicians review every output before it enters the chart. Enterprise contracts are said to include AI governance controls. No named body, no described evaluation process, no monitored metric and no certification sits behind any of it, and no external audit of model behavior is published. That is commitment without evaluation.

No fairness statement, subgroup analysis, accent or dialect performance disclosure or language coverage claim is published.

The gap is sharpened by a figure the company does publish. Note quality is scored above 98 percent, and the company states plainly that this is internal scoring measured across active clinicians over a stated period. Labeling it as internal is honest. It is also a single aggregate, and an overall score is dominated by ordinary sessions with clear audio and familiar speech patterns and conceals whatever variation sits underneath. The scoring apparatus already runs continuously across active clinicians. Breaking that existing measurement out by speaker characteristics would require no new infrastructure.

The specialty makes the gap more pressing. Behavioral health assessment carries documented disparities in how symptoms are recorded across patient populations, particularly in the language used to describe affect, insight and cooperation. A system tuned on psychiatric encounters will have learned the conventions in its training material, including any patterned differences, and a platform that carries data points forward across visits propagates them rather than leaving them in one note.

Real mechanisms for challenging an output, with nothing standing behind the output itself.

What exists is a chain that can be traced. Each draft section displays the encounter context that produced it, so a clinician can see what the sentence was built from. Drafts are stated to come from the current visit audio and the existing chart rather than from general training. A pre signature review pass flags omissions and inconsistencies. Nothing enters the chart without a clinician reviewing and signing, and the company states the system cannot sign, file or finalize a chart. Audit logs are available. Any encounter or output can be flagged in one click, and the company says it uses those flags to refine prompts, evaluations and tooling.

That is traceability, review before commit and a feedback channel.

What is absent is a commitment. No service level agreement, accuracy warranty, indemnity or remediation obligation is published, and no route is described for a patient harmed by a wrong note. The company does state that the assistant will not invent medications, diagnoses or clinical events that were not in the encounter, which is falsifiable in principle, but nothing stands behind it: no measurement, no threshold and no consequence if it fails. A claim without a figure or a remedy is not a commitment. The one published quality figure, internal scoring above 98 percent on note quality across active clinicians, is openly labeled as internal but carries no method, reference standard or denominator, so it cannot be tested either. The claim first pass acceptance figure measures billing outcomes, not clinical output.

One risk is specific to this architecture. An agent that references prior records can carry content forward into a new note with the authority of fresh clinical observation. Source display goes some way to addressing it, but nothing states whether the draft distinguishes what was heard in the current visit from what was carried over from past visits.

The segment raises the stakes. A psychotherapy or addiction medicine note supports a clinical formulation, is submitted for reimbursement and is discoverable in custody, criminal and civil proceedings.

Named systems and named standards, with nothing yet verifiable outside the vendor's own pages.

Nine record systems are named: Epic, Cerner, athenahealth, eClinicalWorks, NextGen, Osmind, Elation, DrChrono and Kareo. The standards are named too, FHIR R4 and HL7 v2, and the scope of what moves is stated rather than implied: schedule, problems, medications and notes, described as bidirectional. Three systems are singled out as having a direct integration, Osmind, SimplePractice and Kalix, which are the behavioral health native systems this vendor's buyers actually run. A Chrome extension docks a sidebar beside the record system for practices without an interface, and finished sections are dragged into the chart.

Separately the platform connects to Surescripts for prescribing, which moves part of the clinical record through a named national network, and to Change Healthcare and Availity on the claims side. Clearinghouses are commercial rather than clinical plumbing.

Verification is missing. Every one of these claims sits on the vendor's own marketing pages. No marketplace listing, partner directory entry or integration document is published for any of the named record systems, and nothing establishes that a given connection runs in production at a named site. Bidirectional is asserted at the level of the platform, not demonstrated per system, and the depth very likely varies between a native Osmind connection and an Epic interface.

Residency is stated, the path the data takes to get there is not.

What is published is a United States hosting commitment on infrastructure carrying its own business associate agreements, encryption in transit and at rest at TLS 1.3 and AES-256, configurable retention, and recordings deleted on signature. Delivery is cloud across desktop and browser with an iOS companion and a Chrome extension. Single sign on, SCIM provisioning and audit logs are available.

What is open is everything underneath. No cloud provider is named, no region is specified within the United States, no tenancy model is described, and no sub processor list is published. Nothing states where model inference happens or whether any processing crosses a border, which matters here because the company states it tuned its own model on specialty encounters, leaving open whether a commercial foundation model sits underneath and where that service runs.

Two product facts sharpen the question. The iOS companion is built for capture away from the desk, including home visits and offline queues, and the product screens show sessions queued for sync. What is held on the handset before it syncs, whether it is encrypted at rest there, and what happens if the device is lost with unsynced sessions on it are not addressed anywhere. Separately, deleting recordings on signature narrows the residency question usefully: if audio does not come to rest, the surface is the processing window and the note store rather than an audio archive.

Published tiers with figures, a stated unit of charge, and a way to start without talking to anyone.

Three tiers are published. Nextvisit Go is $49 a month billed monthly for non prescribing therapists and counselors, covering 100 notes a month, SOAP, DAP and BIRP formats, interactive note editing, one click copy out to any record system, and a signed business associate agreement. Provider+ is $199 a month billed monthly for prescribers, adding the AI assistant, custom templates and cohort tags, AI assisted peer review, the patient timeline, and unlimited notes. Enterprise is quoted, for groups, networks and academic sites, and lists what it adds: single sign on, advanced reporting, supervision workflows, templates at scale, a dedicated customer success manager, and business associate, SOC 2 and AI governance controls.

The unit of charge is stated plainly and the company says so in its own words: per month, with no per note fees and no seats. The note ceiling on the entry tier and its absence above are both published, which is the figure that usually decides whether a documentation product is affordable at volume.

The self serve route is real rather than nominal. A fourteen day trial runs without a card, the company states solo practices need no onboarding call, and a first usable note is claimed within ten minutes.

What is not published is the Enterprise figure and whether prescribing tier pricing changes by seat count at group scale. Neither is unusual at that tier, and neither prevents a solo or small group buyer from sizing the cost before contact.

Coverage is enumerated properly and the validation sits behind only part of it.

What is named is specific on every dimension. Clinician types: psychiatrists, psychiatric mental health nurse practitioners, psychologists, therapists, counselors, licensed clinical social workers and physician assistants, with a separate non prescribing workspace for those who do not write scripts. Settings: outpatient psychiatry, therapy clinics, community mental health, group practices, academic sites, telehealth through a built in video visit, and home visits through the iOS companion. Specialties listed on the product's own page: psychiatry, adult therapy, adolescent therapy, psychiatric nursing, addiction medicine, couples and family, neuropsychology and group therapy.

The distinguishing depth is treatment specific rather than categorical. Templates ship for SOAP, DAP, BIRP, intake, Spravato, TMS, medication assisted treatment and lithium monitoring. Lithium is the most specialized of them: it has a narrow therapeutic index and requires serial serum monitoring on a schedule that a general purpose tool would not know to track.

Support sits behind the core and not the edges. Volume is stated at 2.3 million encounters across private practice, groups and academic sites, with more than 300 practices, and the one named deployment is a solo outpatient psychiatry practice. Nothing validates the adolescent therapy, neuropsychology or group therapy coverage specifically, and group therapy in particular raises a distinct technical problem, separating individual voices in a room, that the coverage claim does not address.

No language coverage is published, and no position is stated on non English sessions, which matters for a product whose output is a clinical record.