Nudge AI

Nothing sits underneath the models; they are the product. Every module sold is model output: ambient capture and note generation, template conversion from a pasted example, CPT and add on code detection with generated justifications, ICD-10 diagnostic coding with supporting rationale, an auditor that reads existing documentation for missed billing and compliance gaps, a patient context feature that references prior notes, and a voice receptionist that answers calls, schedules and triages. Strip the models out and what remains is a template gallery and a browser extension that fills form fields with nothing to put in them.

No services business, no practice management system, no billing engine and no record system sits underneath. The browser extension and the published integration API are delivery surfaces for model output rather than products a buyer would purchase on their own.

The coding modules are the clearest case. Identifying an evaluation and management level from session complexity and authoring the justification that supports it is not a rules engine with a model attached; there is nothing else in that feature.

A review step is described, and two things run without one.

The clinician reviews and refines the draft through instruction based editing and can regenerate against a different template. That is the documentation half and it is conventional.

What happens automatically is the larger issue. Diagnostic codes are generated with supporting rationale, and the billing module generates add on justifications on its own. Writing the justification for a billing code is a materially different act from suggesting the code, because it produces the documentary evidence a payer would examine rather than a recommendation a clinician evaluates. Combined with one click population of structured fields, checklists and radio buttons, a clinician can move from recording to a fully coded and justified chart entry with very little friction. No confidence threshold, accuracy figure or abstention behavior is published for any of it.

The second autonomous surface is the artificial intelligence receptionist, and here oversight appears as a phrase rather than a control. It answers patient calls around the clock, schedules appointments, answers questions about the practice, gathers information and connects callers to the team when needed. When needed is the whole of what is published: no escalation criterion, no threshold, no described behavior for a caller in distress, and nothing stating when a person is reachable. Answering every call at every hour means taking on the caller in crisis at two in the morning, and the practice list this is sold into includes crisis intervention, addiction counseling and partial hospitalization programs. That is the population where the gap bites hardest.

Nothing is published about what produces the output, and the positioning fills the space with superlatives.

No model or model family is named, no foundation model provider is identified, no model card exists, no version or update cadence is stated and no evaluation methodology is published. Product behavior is described at feature level, ambient capture producing a note in thirty seconds, context awareness that references prior notes, complexity analysis that selects evaluation and management levels, but nothing describes what performs any of it.

The headline claim is that this is the best artificial intelligence scribe and clinical platform, which appears in the page title and the main heading. A superlative is not a disclosure and cannot be tested.

One inconsistency in the vendor's own figures has been resolved. The supported record system count was stated at more than 50 in some vendor materials and more than 140 in others; the current site says 25 or more throughout. That is a correction in the right direction, and it matters, because inconsistent figures on a vendor's own material are a poor basis for reliance on any other figure it publishes.

The case study figures are the closest thing to measurement the vendor publishes: 90 percent coding accuracy against a stated 54 percent national average, a 19.9 percent lift in relative value units per visit, and documentation time falling from 20 minutes to 7. Those measure commercial and workflow outcomes rather than model behavior, and the national average they are compared against carries no source.

The terms governing the chain are now published and the parties in it are still entirely unnamed.

What changed is real. The company states plainly that it does not train on customer data, describes a zero retention architecture, and says protected health information is redacted during session capture. Those are terms about what happens to content once it leaves the clinician, and a buyer can hold the vendor to them.

What has not changed is that nothing establishes who else sits between a patient record and an answer. No model or model family is named, no foundation model provider is named, no hosting arrangement is described, and no sub processor list is published. A published integration API and browser software development kit describe the vendor's own surface, not what is underneath it. Zero retention is the standard phrasing for an arrangement with a third party model service, so the term itself implies a party the disclosure does not identify.

The asymmetry with the consent feature is worth a buyer's attention. This vendor can tell a patient that their consent was recorded, and cannot tell them which companies processed what they said. Consent to be recorded is not consent to any particular chain, and a consent record that does not enumerate recipients documents the agreement without describing what was agreed to.

Named encryption and audit logging protect and record access without saying who holds it.

Cohort sized figures with denominators, attached to organizations that are not named.

Three performance studies are published. A multistate provider network of more than 200 clinicians is reported at 90 percent coding accuracy against a stated 54 percent national average, a 75 percent reduction in post visit coding queries and a 19.9 percent lift in relative value units per visit. A multispecialty clinic of more than 100 clinicians is reported to have had missed billable add on codes in 20 percent of medication management sessions, with a stated $9,600 increase in revenue per clinician per year from detecting them. A behavioral health enterprise of more than 250 clinicians is reported to have cut documentation time from 20 minutes to 7 per session, a 65 percent reduction, producing 299.5 hours of clinician time reclaimed over 10 weeks.

Those are better than unsupported percentages. Cohort sizes are given, the measurement window is stated in one case, and the mechanism behind each figure is stated plainly enough to argue with.

What is missing is attribution: not one organization is named. A multistate provider network and a behavioral health enterprise are descriptions, not deployments a buyer can call. Nothing is independent, every study is the vendor's own, and the 54 percent national coding accuracy figure the headline result is measured against carries no source at all, which makes the most striking number on the page the least checkable. The full studies sit in separate downloadable documents rather than on the page, so the figures above are the summaries the site itself displays.

Retention, training use and deidentification are each answered specifically enough to be held to.

Retention: audio is stated to be automatically deleted within seven days. Training use: no training on customer data, stated flatly with no carve out, no permission gate and no third party qualifier. Deidentification: protected health information is stated to be redacted during session capture, which places the redaction at the point of capture rather than at some later processing step. Each of those is a sentence a customer could later show to be untrue, which is the test.

The supporting controls are named rather than gestured at. AES-256 encryption, comprehensive audit logging, and consent tracking, which makes patient consent a recorded and auditable object rather than a policy paragraph or a script the clinician is trusted to remember. For a product whose input is recorded psychotherapy and substance use treatment sessions, an auditable record of having asked is the control that matters most a year later when someone disputes whether a session should have been captured at all.

One tension is unresolved. The same paragraph describes a zero retention architecture and audio deleted within seven days. Those can be consistent, if zero retention describes the terms governing whatever model service processes the session while seven days describes the vendor's own audio store, but nothing published says which the seven days applies to, or whether any copy persists in backups beyond it.

No safety engineering is published: nothing describes guardrails, hallucination mitigation or how a safety event is handled.

Status is stated in two regimes and supported by real controls, and the one thing an entry tier buyer needs is contradicted across the vendor's own pages.

HIPAA compliance is claimed throughout, and Canadian PIPEDA compliance alongside it, which adds a second regime. The claim sits on substance rather than a badge: AES-256 encryption, redaction of protected health information at capture, audio deleted within seven days, no training on customer data, and comprehensive audit logging.

The contradiction is about who gets an agreement. Earlier company material described business associate agreement support as being for enterprise use. The current pricing page marks the free tier as 100 percent HIPAA compliant while attaching complete healthcare privacy compliance for enterprises to the HIPAA claim beside it. Those are not the same statement. Compliance is a posture a vendor asserts about its own practices; an executed business associate agreement is the contract that makes it accountable to the covered entity, defines permitted uses and creates breach obligations. A product can be engineered to HIPAA standards and still leave an individual subscriber without one.

The exposed buyer is specific: a solo therapist or psychiatrist on the $99 plan, recording psychotherapy sessions, who reads 100 percent HIPAA compliant and assumes an agreement is in place. A clinician recording sessions without an executed agreement has committed a violation whatever the software does. What that entry buyer can rely on is not established.

A business associate agreement page exists on the site. Whatever it holds, it would not settle a contradiction that sits on two other pages.

No subcontractor flow down position is published. Separately, the product ships templates for partial hospitalization and substance use disorder programs, which places those records under a federal confidentiality regime that a standard agreement does not by itself satisfy.

Nothing published says whether an agreement is executed at each tier, what template it follows, or how the vendor treats substance use disorder records and redisclosure.

No attestation of any kind is published, and the button that promises one leads nowhere.

There is no SOC 2 report of either type, no HITRUST certification, no ISO 27001 and no trust center. The security section names real controls, AES-256 encryption, automatic redaction of protected health information at capture, a zero retention architecture and comprehensive audit logging, and HIPAA and PIPEDA compliance are both claimed. Controls are not independent examinations of controls. A Read Security Protocol control appears in that section and carries no destination, so the one document a buyer is invited to read cannot be reached.

The delivery mechanism creates a further security question, and it is the most important finding here. The product is delivered substantially through a browser extension that reads the record system screen and writes into its fields. An extension with permission to do that has access to whatever is rendered in that browser, which is not limited to the encounter being documented: it extends to any record the clinician opens in that session and potentially to other systems accessed in the same browser. That is a materially different access surface from a server to server integration, and it sits on the clinician's own device rather than inside a controlled environment. No store listing is linked from the site, so the permission manifest cannot be read either.

The company does not say what page permissions the extension requests, whether access is scoped to particular record system domains or granted broadly, what it transmits from pages other than the active encounter, or whether an organization can restrict or audit it centrally rather than relying on each clinician's browser.

The data involved raises the stakes. This is a behavioral health product handling psychiatry, therapy and medication management sessions, with templates for substance use disorder programs.

None of this means the controls are absent; it means they cannot be verified before contracting.

No clearance claimed and none required for documentation. No United States device pathway attaches. What governs this product is payment integrity, and the structural question it raises is unusual enough to state carefully.

The product generates diagnostic codes with supporting rationale, identifies procedural and add on codes with justifications, analyzes session complexity to set the evaluation and management level, and flags what it calls missed revenue opportunities and leakage from undercoding. The vendor quantifies the effect in its own case studies: a 19.9 percent lift in relative value units per visit, $9,600 of additional revenue per clinician per year, and a finding that a fifth of medication management sessions carried uncoded billable add ons, naming the psychotherapy and crisis code families.

The structural point is this. The system does not only propose a code. It generates the written justification supporting that code. Documentation is supposed to support a code because it records what happened; here the same system selects the code and authors the text that defends it, so the justification cannot function as an independent check on the selection. Nothing published says what the justification is derived from, or whether it can ever fail to support the code the system chose.

The audit framing deserves the same scrutiny. Notes are described as audit proof, and the platform is sold as shielding a practice from audit risk while eliminating revenue leakage. Those are opposite ends of a trade off that regulators treat as opposed, and no documentation product delivers both by construction.

One further regime applies and is unaddressed. The product supports substance use disorder program templates, including partial hospitalization and intensive outpatient settings, and those records fall under a separate federal confidentiality regime with its own consent and redisclosure rules that general privacy compliance does not satisfy.

Aggressive on coding, and the language is the vendor's own.

The product describes identifying billing opportunities such as psychotherapy add ons, naming the 90833 to 90838 family and the G2211 complexity code specifically, assessing session complexity to recommend evaluation and management levels, flagging missed revenue opportunities, and eliminating leakage from undercoding. A separate auditor module is sold as turning existing documentation into a financial guardrail. The case studies quantify the effect: a 19.9 percent lift in relative value units per visit, $9,600 of additional revenue per clinician per year, and a finding that a fifth of medication management sessions had uncoded billable add ons.

Generating the justification for an add on code, rather than only proposing the code, is the sharpest version of this. The system produces the very documentation that would be examined if the claim were questioned.

Real counterweights exist and belong in the same paragraph. Notes are framed as audit proof and as demonstrating medical necessity, consent is tracked, and actions are audit logged, all of which point at defensibility rather than only at revenue. Shielding a practice from audit risk and maximizing reimbursement are still presented together, and regulators treat those as opposite ends of a trade off.

Fairness disclosure is absent entirely. No fairness statement, no subgroup analysis, no performance disclosure by any patient or clinician characteristic, no calibration data, no drift monitoring and no external audit of model behavior. No language coverage beyond English is published, while the product is marketed across more than 50 specialties.

The coding behavior makes the absence sharper rather than incidental. A system that selects evaluation and management levels from session content is making a judgment whose inputs include how a patient presents and how their symptoms are described, and documented disparities exist in exactly that. Nothing published examines whether the coding output varies by patient group.

One shipped feature matters here, and nothing stands behind the output itself.

Consent tracking is a named control, which makes patient consent a recorded and auditable object rather than a policy paragraph or a script the clinician is trusted to remember. That matters specifically, because recourse begins with the affected person having been asked, and an auditable record of asking is what makes that checkable a year later when someone disputes whether a session should have been captured. Audit logging and named encryption sit alongside it, and the product publishes a retention position, a no training commitment and capture time redaction.

Those are mechanisms for establishing what happened. None of them is a commitment about whether the output is right.

No accuracy figure, error rate or evaluation methodology is published for the note, the code selection or the justification text. No service level agreement, warranty, indemnity or remediation obligation exists, and no route is described for a patient or a payer affected by a wrong output. Positioning rests instead on superlatives, the page title and headline describing the product as the best artificial intelligence scribe and clinical platform, which is not a falsifiable claim.

One figure has improved. The record system count, previously stated inconsistently at more than 50 in some materials and more than 140 in others, now reads 25 or more consistently across the site. The earlier inconsistency was a poor basis for reliance and it has been resolved.

The exposure here is larger because the product authors billing justifications. If a justification is wrong, the consequence lands on a claim rather than only on a note, and the vendor has published nothing about who answers for it.

Breadth is claimed and not one record system is named.

The published position is 25 or more supported systems, stated on the product pages and beside the browser extension. No list appears anywhere on the vendor's site, no system is named in any material, no marketplace or partner listing is published, and no link to the extension's store listing is offered, so nothing checkable sits behind the number.

The figure has at least settled. Earlier vendor material put the count at more than 50 in some places and more than 140 in others, and the current site says 25 or more consistently. The change reads as a correction rather than a reduction in capability.

The transfer mechanism itself goes beyond a paste workflow. The Chrome extension operates in the same tab as the record system and populates fields with one click, including checklists, radio buttons and structured fields rather than dropping a block of narrative, which is the difference between filling a form and pasting into one.

A sanctioned channel also exists alongside it. The company publishes an authenticated REST integration API and a browser software development kit, documented publicly, covering providers, clients and notes including SOAP and psychiatric assessment formats. That is a permissioned route rather than screen manipulation, and no record system is named as connected through it.

Browser level field population leaves the usual questions open: whose credentials the write runs under, how it appears in the record system's audit log, whether the permission model that constrains a human user constrains it equally, and what happens when the host system changes its interface.

A single hosted option with nothing published about where it runs.

No hosting region, no residency option, no tenancy model and no subprocessor list is published, and nothing establishes whether a third party model service processes the session or what it retains. A zero retention architecture is claimed, which is a term about retention rather than a statement of location, and PIPEDA compliance is claimed alongside HIPAA without any statement about whether Canadian customer data rests in Canada.

Delivery runs through three surfaces: a web application, an iOS application, and a Chrome extension that operates in the same tab as the record system. A public integration API and browser software development kit are documented separately.

The transfer mechanism reframes the compatibility claim. The extension reaches the record system by operating on the rendered page. Supporting 25 or more record systems is achievable that way precisely because the extension works on what is displayed rather than through each system's own interface, so the breadth claim and the absence of permissioned integrations are the same fact stated two ways rather than evidence of unusually wide engineering. The published API is the exception, and no system is named as connected through it.

Reaching the chart through what a user sees raises questions nothing answers: whose credentials the write runs under, how it appears in the record system's audit log, whether the permission model constraining a human user constrains it equally, and what happens when the host system changes its interface.

Three published tiers with figures, a stated unit of charge and a free entry point that needs no card.

The free tier is $0 and allows five notes with no credit card, which lets a clinician see the output on their own sessions before deciding. Pro is $99 a month, described as being for solo providers, and carries unlimited session notes, the coding module, record system transfer, custom templates and the clinical assistant, with cancellation stated as available at any time. A monthly and yearly toggle is published, so the billing period is a choice the buyer can see. Enterprise is quoted, adding volume discounts, a dedicated account manager, enterprise analytics and premium support. A small clinic program offering ten percent off is advertised separately.

The unit of charge is per provider per month, which the tier description makes explicit rather than leaving to inference, and the note allowance is published at both levels, which is the number that decides affordability.

What is not published is the Enterprise figure, the volume discount structure, and whether the yearly option carries a stated discount rather than only a toggle. None of that prevents the solo and small practice buyers this product is aimed at from sizing the cost before contacting anyone.

Behavioral health depth sits inside a product now marketed far more broadly, and the depth is the part with support behind it.

The behavioral coverage is specific and unusual where it goes deepest. Clinician types are enumerated as psychiatrists, clinical psychologists, psychiatric mental health nurse practitioners, therapists, licensed clinical social workers and behavioral health consultants. The distinguishing depth is by level of care rather than by note format: intensive outpatient programs, partial hospitalization programs, transcranial magnetic stimulation, addiction counseling, crisis intervention and group therapy are each listed as served settings, and the product ships templates for partial hospitalization daily progress notes in substance use disorder programs written to Medicare and insurance documentation requirements. Those are documentation regimes with their own regulatory expectations rather than simply another template.

Support sits behind that core. A behavioral health enterprise of more than 250 clinicians is reported as a deployment, and more than 5,000 clinicians are claimed overall.

The scope has widened well past behavioral health. The site now claims more than 50 specialties and leads with psychiatry, psychotherapy, family medicine, internal medicine and pediatrics, with a medicine and surgery category alongside the behavioral one. Nothing validates the general medical coverage, and a coding module tuned on psychiatric session complexity is not obviously the same instrument when applied to a surgical or pediatric encounter.

Templates are created by pasting an existing note for the system to replicate, which means coverage in practice is whatever a customer brings rather than what the vendor has validated. English only, with no position published on sessions conducted in another language.