Lyssn

The measurement is the engine and the platform around it carries value without it.

What the models do cannot be done another way. Automatically estimating fidelity to an evidence based practice from a session recording, identifying whether a clinician used specific techniques competently across more than 75 named metrics, is not achievable by workflow software. The only alternative is a trained human rater listening to the session, which is exactly the bottleneck the company exists to remove, and the founders are clinical researchers and data scientists rather than operators who added a model to an existing service.

What would remain is more than nothing. Two solutions are sold. Quality Improvement is a secure platform for recording, uploading, storing, playing back and annotating therapy sessions, with time linked commenting and asynchronous supervisor review, retention controls and role management across clinicians, supervisors and sessions. Learn and Practice is a library of training modules for named evidence based practices, including a suicide prevention fundamentals program. Strip every model out and a supervision video platform and a curriculum remain, and both were sold as products before automated fidelity scoring existed.

The distinction that matters is what a buyer is choosing between. An organization picks Lyssn over a generic session review tool because of the automated fidelity scoring, and picks it over a training vendor because the modules evaluate skill growth automatically. The intelligence is the reason, and it is the engine of the core modules rather than the whole of what is delivered.

One further component is entirely model dependent: a patient like conversational agent used for practicing basic counseling skills, which has no non model form at all.

The boundary is drawn clearly and the thing that would make it enforceable is not published.

What is stated is consistent. Fidelity scores feed supervision rather than replacing supervisors: the platform supports time linked comments inside session playback and asynchronous review between therapist and supervisor, so the software produces the evidence and the supervisor produces the judgment. The company states it keeps the human element front and center and positions the product as quality improvement and training rather than as an arbiter. That is a well drawn boundary for a measurement tool.

What is missing is where the score stops. No threshold is published at which a rating is treated as actionable, no confidence interval or margin accompanies a score, and nothing describes what happens when the system and a human supervisor disagree.

The gap that matters most is contestability, and it is specific to this product rather than generic. The output is an evaluation of a named clinician's practice, scored against fidelity frameworks, displayed on dashboards summarizing performance across a caseload. Nothing published states what those scores are permitted to be used for, whether they can enter a supervision record, a credentialing decision or a personnel process, or what recourse a clinician has to contest a rating they believe is wrong. A measurement that attaches to someone's professional standing needs an appeal route, and none is described.

The company's own research includes published work on clinician knowledge and attitudes toward artificial intelligence based fidelity measurement in supervision, so the acceptability question has been studied. The governance answer it implies is not published as product policy.

The approach and the training corpus are described in unusual detail, and nothing describes the deployed artifact.

What is published goes well past a category description. The method is supervised measurement against established gold standard fidelity frameworks used in academia and clinical practice, with the software trained by licensed clinicians and artificial intelligence specialists to evaluate sessions using the same metrics a human rater would. The training corpus is characterized concretely: more than 35,000 expert evaluated behavioral health conversations, 4.3 million analyzed real world sessions, and 6.6 billion labeled words drawn from clinical, social services and crisis conversations. More than 75 output metrics are enumerated by name, spanning treatment proficiency across motivational interviewing, cognitive behavioral therapy, dialectical behavior therapy, acceptance and commitment therapy, trauma informed care, multicultural orientation, safety planning and suicide prevention, and clinical behaviors including empathy, collaboration, talk time, engagement, reflections and question types. The underlying methods sit in peer reviewed venues, which is a durable form of disclosure because a published architecture cannot be quietly revised the way a product page can.

What is absent is the version discipline. No model class, family or provider is named, no version is given for any fidelity model, no release or retraining cadence is stated, and no model card exists. Nothing tells a customer which version scored their sessions or how they would learn it had changed, which matters for a product whose output attaches to a named clinician's record and is compared across a caseload over time.

The deployed against studied question follows from the same gap and is the one to put in writing. Methods published across 17 years of research are not necessarily the models running today, and nothing connects the two.

One inconsistency between the company's own pages: publications are stated as more than 60 on one and more than 65 on another, analyzed sessions as 4.3 million and 4.5 million.

The system is described component by component in peer reviewed literature, and no party that supplies or operates any of it is named.

Publication is a real and durable form of disclosure. A published architecture cannot be quietly revised the way a product page can, and it lets a technical reviewer understand what the platform is composed of rather than accepting a category description. The company answers much of the composition question through its literature.

What it does not do is name anyone. No foundation model provider, model class or version is identified, no hosting arrangement is published, and no sub processor register exists, so a buyer can describe the architecture and cannot name a single party that operates or supplies any part of it. Data is stated to be housed in the cloud rather than on devices, without saying whose cloud.

One half of the question is now answered and answered well, on the security page rather than in the literature. Recordings are retained for a default 90 days and then de identified, or deleted entirely at the customer's election. Before any use for algorithm calibration, identifying information is removed by automated means and by targeted independent human review. Customer sessions are used for calibration, with the permission of consenting customers, reviewed by the company's own in house clinical team on de identified data. So what happens to the content inside the company is described; who else touches it on the way is not.

The content raises the stakes of the remaining gap. This platform holds full recordings of psychotherapy, crisis line calls, child welfare and justice involved conversations, and whether any external transcription or model service ever receives that audio is unstated in either direction.

The claim is about agreement with expert human raters, and that is precisely what the published work measures.

What the company claims, and it is disciplined about it, is measurement accuracy rather than patient benefit. Its headline figures are stated as the proportion of human performance the system matches when evaluating for evidence based practices, given separately for motivational interviewing, cognitive behavioral therapy fundamentals, multicultural orientation, safety planning and research use. Those are claims about whether the software rates a session the way a trained expert rater would.

The evidence answers that question directly. The company reports more than 60 peer reviewed publications built on 17 years of academic research, with proprietary data comprising more than 35,000 expert evaluated behavioral health conversations, 4.3 million analyzed sessions and 6.6 billion labeled words drawn from clinical, social services and crisis conversations. Metrics are coded against established gold standard fidelity frameworks used in the field rather than against a scale the company invented, which is what makes agreement with human raters a meaningful test at all. More than 75 metrics are enumerated by name across treatment proficiency and clinical behaviors.

Independence is real rather than asserted. Project AFFECT is funded by the National Institute of Mental Health, conducted with the Penn Collaborative for CBT and Implementation Science and registered as NCT05340738, comparing AI generated fidelity feedback against usual care. Separate University of Washington work with King County's healthcare system evaluated detection of cognitive behavioral therapy for psychosis skills. The founders' equity stakes are disclosed as competing interests inside the papers, a transparency standard imposed by the venue rather than chosen by the vendor.

The causal chain from better measured fidelity to better patient outcomes is not established, and Project AFFECT is the trial that would settle it and has not reported. The company's unquantified taglines about outcomes and improvement run ahead of that. Every quantified claim it makes is a measurement claim, and those are answered.

Two figures disagree between the company's own pages: publications at 60 or more against 65 or more, and analyzed sessions at 4.3 million against 4.5 million.

Retention, deidentification and training use are all answered specifically, on the most sensitive material any behavioral health product handles.

Retention is published as a schedule with a customer control. Recorded sessions are available for playback and annotation for a default of 90 days from recording, customers can change that window, and afterward recordings are deidentified and no longer accessible. A customer may instead choose that sessions are deleted entirely and retained in no form. That is a figure and a default a customer could hold the company to, not a posture.

Deidentification is described as a practice rather than named as a principle. Before any use for algorithm calibration, personally identifiable information associated with a recording is removed, and both automated algorithms and targeted independent human review are used to find and remove identifying information spoken inside the session. Naming the human review step matters here, because automated redaction alone does not catch a client naming their street or their employer mid sentence.

Training use is stated plainly and in a form that can be contradicted. With permission from consenting customers, an in house team of clinicians reviews a small sample of sessions to assess empathy and other quality metrics, working on privacy protected deidentified data, to validate and calibrate the algorithms. So customer sessions do contribute to model calibration, the consent basis is stated, the data is deidentified first, and the reviewing party is the company's own staff rather than an unnamed third party.

Supporting controls: encrypted transfer and storage, two factor authentication, continuous monitoring, access restricted to approved users who can review and delete content, cloud storage rather than local devices, and education records compliance for public sector customers.

One responsibility is explicitly placed on the customer rather than absorbed, and the company says so: users are responsible for obtaining appropriate consent to record and for knowing the recording laws where they operate, with a template offered as a starting point. That is honest rather than evasive, and it leaves consent as work the customer owns.

No safety engineering is published.

Status is stated and supported by a substantive security document, with the agreement itself unpublished.

The supporting material is real rather than a badge. The security page describes encrypted transfer and storage, two factor authentication, continuous system monitoring, access restricted to approved users, a published retention default of 90 days with customer configuration and a delete entirely option, a described deidentification practice, and regular external health privacy and SOC 2 audits with attestation reports available on request. Education records compliance is stated separately for public sector customers, which is the right additional regime for a company selling into child welfare, justice involved services and school linked programs.

The platform's health privacy posture is also documented in peer reviewed literature describing it, which is stronger corroboration than a marketing claim.

The instrument is the gap. No business associate agreement template, execution requirement, subcontractor flow down position or statement of which entity contracts is published, so neither the agreement nor its scope can be read before contracting.

One feature of this product sharpens the agreement question. The subject of the recording is a client receiving therapy, the person being measured is the clinician, and the buyer is a training program, health system or public agency. Responsibility for obtaining consent to record is explicitly placed on the customer, with a template offered. Nothing published says where the company sits in that chain for each deployment, or, in the human services lines, whether the counterparty is a covered entity at all.

The attestation position is real, and the report type matters.

The company states regular external health privacy and SOC 2 audits, with attestation reports available on request, alongside encrypted transfer and storage, two factor authentication, continuous system monitoring, access restricted to approved users, and education records compliance for public sector customers. Its own announcements have described a combined health privacy and SOC 2 attestation completed annually over several consecutive years, with the auditor named.

One distinction runs in an unusual direction. The company's own announcements have consistently said Type 1. A third party product directory describes it as holding SOC 2 Type II. The directory has very likely inflated it, since the company's own account has not changed across those years. The security page itself names no type at all, so that page alone does not say.

What the distinction means substantively: a Type 1 report assesses whether controls were suitably designed at a point in time. A Type II tests whether they operated effectively across a period. Several annual Type 1 reports are not equivalent to one Type II, because repeating a point in time assessment does not aggregate into period coverage. Annual repetition does give a recency benefit a single report would not, and it demonstrates sustained investment.

Retrievability and scope are the gaps. Reports are available on request rather than published, no audit period or current validity date is stated, and no scope statement says which parts of the platform are covered.

For a platform holding complete therapy session recordings, crisis line calls in which suicidality is discussed, and child welfare conversations, the absence of a stated Type II report matters.

Not a device, and the oversight that does apply is nowhere addressed.

The device question resolves quickly and correctly. The system measures clinician performance and supports training and supervision. It does not diagnose patients or direct treatment, so it sits outside software as a medical device.

What is absent is a stated position on the regimes that do govern it. The relevant surfaces are professional training and supervision standards, state licensure board expectations about how supervision is conducted and documented, and for public sector deployments the procurement, quality assurance and records requirements those buyers carry. Nothing published sets out which standards the company believes govern its assessments, how its scoring relates to a supervision requirement, or whether a fidelity report can satisfy one.

That matters more here than for a documentation tool, because the output is an evaluation of a named clinician's practice held by their employer. Nothing published says what the scores are permitted to be used for, whether they can enter a personnel or credentialing process, or what recourse a clinician has to contest one.

The sector expansion raises further regimes. Child welfare and justice involved services carry their own confidentiality frameworks, court oversight and records rules, distinct from health privacy law, and recordings of those conversations may be discoverable in proceedings involving the families concerned. Crisis line recordings raise the same question where suicidality is discussed. None of it is addressed.

Education records compliance is claimed for public sector customers, which is the one adjacent regime the company does name.

Cultural responsiveness is treated as a measurable dimension rather than a principle, and no subgroup performance is published.

What exists is better than a policy page. The published research program includes work on automating the assessment of multicultural orientation through machine learning and natural language processing, and multicultural orientation appears as a named, shipped metric in the platform rather than only as a paper. Treating cultural responsiveness as something a fidelity model measures, and selling it as such, is a substantive position. A separate published study examined clinician knowledge and attitudes toward artificial intelligence based fidelity measurement in supervision, which takes the acceptability question seriously from the side of the person being measured. The company also publishes de identification and consent practice, and works to a stated position that people should be told a session is being recorded and how the recording may be used.

What is absent is the output. No performance breakdown by clinician or client characteristic is published, no calibration analysis, no fairness audit, no drift monitoring result and no external audit of model behavior.

The stakes are specific. A fidelity score that systematically underrates clinicians with particular accents, dialects, first languages or therapeutic styles becomes an unfair performance evaluation attached to a named professional's standing, and it would be invisible to that person. The workforce being measured here spans community mental health, crisis lines, child welfare and justice involved services, which are among the most linguistically and demographically varied workforces in health and human services. The same question runs on the client side, since what counts as empathy or engagement in a transcript is culturally patterned.

Agreement rates broken out by clinician demographic and by client population are not published, even though the company already scores against human raters at scale.

Published methods that outsiders can contradict, and no figure or commitment attached to the product a buyer receives.

The strength is publication, and it carries more weight than a vendor white paper for a structural reason. The methods appear in academic venues, including work on automated quality assessment of cognitive behavioral therapy sessions, so the approach is described where other researchers can examine and contest it. More significant than the publication itself is what the venue required: the founders' equity stakes in the company are disclosed as competing interests inside the papers. That is a transparency standard imposed by a third party rather than chosen by the vendor. A reader of a vendor white paper has to assess the author's interest for themselves; a reader of these papers is told.

What the company publishes about accuracy is agreement with expert human raters, stated per practice, and that is a falsifiable performance claim rather than a slogan.

What is absent is anything attached to the shipping product. No accuracy figure for the deployed platform as distinct from the published research, no error rate with a denominator for a named model version, no warranty, indemnity or remediation commitment, and no stated route for a clinician to contest a rating.

The consequence of an error is unusual. Most products here get a patient's care wrong. This one gets a clinician's competence wrong. A fidelity score that understates a therapist's practice attaches to a named professional's record, is displayed on a dashboard comparing them with colleagues, and may inform supervision. Nothing published describes what that person can do about it, and the absence of a contest route is the gap that would matter most to the people being measured.

A standalone platform by design, with no named record system integration, and that is a coherent choice rather than only a gap.

Sessions are recorded directly in the browser, captured from telehealth, or uploaded from elsewhere, and the platform operates as its own environment for supervision and quality assurance. User management is organized around clinicians, supervisors and sessions rather than around patients and encounters. No electronic health record is named, no interface standard is described, no marketplace or partner listing exists, and no application programming interface documentation is published.

The design logic is real. The buyer is a training program, supervisor or quality function rather than a documentation workflow, and the object being measured is the clinician's practice rather than the patient's care. Fidelity scores are not clinical findings about a patient and arguably should not enter a patient's chart at all.

What follows from that is a workforce data question rather than a clinical one, and it is unaddressed. Fidelity data lives apart from every other system an organization runs: apart from the learning management system that records training completion, apart from the supervision record, apart from whatever holds credentialing and performance information. For a buyer whose purpose is proving fidelity to evidence based practices across a workforce and reporting it to a payer or a state agency, getting the data out matters more than getting it into a chart, and no export path, reporting interface or data extract specification is published.

The company does publish dashboards summarizing performance across a caseload, so reporting exists inside the product. Whether it leaves is the question.

A browser delivered cloud platform with one concrete commitment published and the infrastructure undescribed.

What is stated: sessions are recorded in person or from telehealth or uploaded from elsewhere, all data is housed in the cloud rather than on individual devices, access is restricted to approved users who can review and delete content, and transfer and storage are encrypted. The retention position is specific, a default 90 days for playback and annotation with customer configuration, then de identification, or deletion entirely at the customer's election. A retention default published as a number is a concrete residency adjacent commitment.

What is absent is everything about location. No hosting provider, no region, no residency commitment, no tenancy model, no segregation description for a platform serving many organizations, and no customer controlled or private deployment option.

That matters more here than the omission alone suggests, because of who the customers are. Public sector buyers, state and local agencies running child welfare and justice involved services, routinely carry their own hosting, authorization and data location requirements, and the company markets to exactly those buyers and offers education records compliance for them. Nothing indicates whether a state agency deployment can be isolated or located to order.

The content raises it again. Full recordings of psychotherapy, crisis calls and child welfare conversations are held for a default 90 days somewhere, and where is unstated.

No pricing published and no pricing basis disclosed. The buyer is an organization, a training program, health system, payer or public agency, so pricing is presumably contracted, and nothing about the structure is public. No unit of charge, no indication of whether the basis is per clinician, per session analyzed, per site or an annual license, no minimum, no term and no pilot position.

Two solutions with different economics make the silence more costly. Quality Improvement scales with sessions analyzed; Learn and Practice is curriculum and scales with learners. A state agency buying workforce training and a health system buying fidelity measurement across a caseload are not the same purchase, and nothing describes whether they are priced the same way.

One relationship bears on independence. Federal research grants supported development of software the company intends to commercialize, a relationship the academic papers disclose openly. That disclosure counts in the company's favor, but the research base and the commercial product share a lineage, so a published finding about the method is not an arm's length assessment of the product.

Coverage is wide across sectors and bounded by which fidelity models exist, and the second constraint is the one that matters.

The sector reach has grown well beyond behavioral health and is now the company's own framing: health and human services including child welfare, justice involved services and prevention and community based services, alongside behavioral health with dedicated lines for certified community behavioral health clinics and for health systems and payors, and crisis care. More than 100 organizations are claimed. Public sector reach is real, evidenced by a King County healthcare system partnership and by education records compliance offered for public agency customers.

The clinical coverage is enumerated rather than claimed. More than 75 metrics ship across named evidence based practices: motivational interviewing, cognitive behavioral therapy, cognitive behavioral therapy for psychosis, dialectical behavior therapy, acceptance and commitment therapy, trauma informed care, multicultural orientation, safety planning and suicide prevention, with a dedicated suicide prevention fundamentals training program. Clinical behavior metrics cover empathy, collaboration, talk time, engagement, feedback and question and reflection types.

What bounds it is the structure of the product rather than ambition. This measures fidelity to specific named practices, so coverage extends exactly as far as the fidelity models that have been built and validated and no further. A practice without a model is not partially covered; it is not covered. That is an honest constraint.

The sector expansion raises a validation question nothing addresses. Fidelity models trained on clinical psychotherapy conversations are being applied to child welfare and justice involved conversations, which differ in purpose, power dynamic and language, and no per sector validation is published.

No language coverage is published, and nothing addresses markets outside the United States.