Limbic

The model performs the clinical work. Limbic Access conducts the assessment conversation, forms hypotheses about the patient's presentation as it goes, and classifies across eight disorder categories; there is no referral portal or case management system underneath that would function without it, because the assessment is the product.

The regulator's view supports it: software is classified as a Class IIa medical device because its output carries clinical consequence rather than administrative convenience, and a certification body assessed this product on that basis. The newer caseload assistant adds an ambient scribe under a separate Class I classification, a lower risk administrative function that sits beside the core rather than diluting it.

The oversight structure is described and one part is missing. Limbic states a classification accuracy of 93 percent across the eight disorder categories the talking therapies program treats, positions the tool as supporting rather than replacing clinician led assessment, and operates under device certification that mandates audited clinical, safety, usability and security risk processes.

The Limbic Layer is itself an oversight mechanism: the language model is constrained by clinically validated guidance and its output is checked against the instruction it was given. What is missing matters more here than an aggregate figure suggests, because the errors are not symmetric.

No breakdown of error by disorder or severity is published, and nothing published describes what happens when a patient discloses acute risk mid conversation, how that is escalated, or how quickly a person sees it. Limbic's own research list includes work on crisis detection models, which suggests the capability exists; the protocol around it is not published.

A specific description of the architecture is published. The Limbic Layer is a cognitive layer of probabilistic machine learning models trained on domain specific data, sitting between the user and a large language model: it constrains what the model is asked, supplies clinically relevant context and instructions, and checks the output against that instruction, which Limbic says makes the interaction predictable.

The approach appears in its own published paper on a cognitive layer architecture for psychotherapy interactions, so the method is set out in research rather than on a product page. Output scope is enumerated precisely, the eight disorder categories the national talking therapies program treats, and performance is stated as a figure, 93 percent classification accuracy.

Nothing is published on the foundation model provider or version beneath the layer, the training data beyond domain specific, accuracy by disorder, or versioning and update practice, so there is no way to tell when the system changed.

Limbic concedes the shape of its chain and names no party in it. Its own description says a large language model sits beneath the Limbic Layer, which constrains and checks what that model produces, so a buyer knows an external general purpose model is involved. Which one, in what version, and on whose infrastructure is not published, and no subprocessor register appears.

The website privacy notice names Heroku as a hosting provider and Amazon Web Services as the servers behind it, but that notice covers website visitors rather than patient assessments, so it does not answer where clinical conversations are processed.

One structural assurance exists without publication: Class IIa certification requires supplier controls and a traceable development process inside an audited quality system, so a documented chain has been reviewed by a certification body even though its contents are not public.

Two published studies carry the evidence, and both are on Limbic's research pages with authors named. A Nature Medicine study (Habicht, Viswanathan, Carrington, Hauser, Harper, Rollwage) is a multi-site real-world observational study of 129,400 patients across 28 NHS Talking Therapies services, comparing referral data before and after Limbic Access against matched control services: a 9 percentage point stronger rise in referrals (15 percent against 6 percent), with larger increases among minority groups, 179 percent for non-binary patients and 29 percent for ethnic minority patients.

A randomized controlled trial (McFadyen, Habicht, Dina, Harper, Hauser, Rollwage) ran six weeks among 540 adults in the United States, comparing the Limbic Care app with a static PDF of the same CBT exercises: 2.4 times more frequent sessions and 3.8 times the usage duration, with a clinical effect Limbic describes as comparable to the workbook rather than better. In short, the trial measured engagement and found parity on outcome, not superiority. Deployment corroborates rather than substitutes: assessments for more than 650,000 patients across a large share of English services.

Deidentification is described as an operation rather than a policy: incoming referral records are logged and patient identifiable information is stripped from the record. The company states that it is not the end destination for patient data and works under the data policies of the commissioning provider organization, which describes a processor relationship accurately and shows where the controlling rules come from.

A data protection officer is named with a contact route, patient rights are enumerated individually, and breach notification is committed to. Device certification at Class IIa brings audited handling of security risk. The central question is unanswered: nothing states whether patient conversations contribute to model development, in either direction, and the Limbic Layer is described as trained on large sets of domain specific data without saying what that data is. For transcripts of people describing depression, anxiety and in some cases thoughts of self harm, that is the gap that matters most. No retention schedule with periods is published either.

Compliance is stated for both regimes Limbic works in, the United States health privacy framework and European and UK data protection law, and published detail backs the claim: a named data protection officer with a contact route, patient rights enumerated individually across access, rectification, erasure, restriction, portability and objection, a route to the national data protection regulator, and a breach notification commitment.

The company also cites the national clinical risk management standard for health software manufacturers, which is a procurement gate in the system it primarily serves. One statement stands out: Limbic says it does not currently use automated decision making that affects a patient's care or legal rights, and commits to telling patients first and explaining how it works if that changes.

What is missing for the United States market it is expanding into is the business associate posture: no agreement, template or execution requirement is published, and no retention schedule with periods appears. The privacy notice on its website covers website visitors rather than the clinical service.

Limbic names two credentials in the same sentence and uses a different word for each: it states that it is ISO 27001 compliant and that it holds a Cyber Essentials certification. Those are not equivalent. Compliant with a standard is a self description; a certification means an accredited body audited and issued it. Take the second at face value and treat the first as a claim to test.

Unambiguously external are the Class IIa device certification, whose audited scope includes security risk alongside clinical and usability risk, and the national clinical risk management standard for health software manufacturers that Limbic cites for NHS procurement. No trust center, report request process, audit period, penetration testing statement or vulnerability disclosure policy is published.

Limbic Access is a regulated medical device, which sets it apart from documentation tools that correctly sit outside device regulation. It holds Class IIa certification under the United Kingdom conformity regime, obtained in January 2023, and Limbic states it was the first AI mental health chatbot to gain that status.

Certification requires documented risk management across the software development process, clinical risk, safety risk, usability risk and security risk, with a traceable and controlled process for design, development and testing, all audited by a certification body.

The classification boundary is applied within the portfolio rather than claimed across it: the ambient scribe in Limbic Care carries its own lower Class I classification, which is the right treatment for an administrative function. Limbic also describes Limbic Access as a clinical decision support tool regulated for adult mental health services. The gap is jurisdictional.

Limbic sells into the United States, where the same functionality meets a different regulator, and no clearance, submission, designation or enforcement discretion position is published for that market.

Published equity research exists, and it answers a different question from how well the classifier performs. The company's Nature Medicine study is an access finding: services using Limbic Access saw proportionally larger increases in referrals from minority groups, 179 percent for non-binary patients and 29 percent for ethnic minority patients, against matched controls. That is evidence that more people from those groups reached the service.

It is not evidence that the classifier performs equally well for them once they engage, and conversational assessment is exactly where idiom and expression of distress could produce differential accuracy. Limbic's research list includes a conference item on validating a diagnostic model across cultural and linguistic differences, but the listing does not show whether it reports accuracy by group, with method and numbers. Device certification at Class IIa brings externally audited risk management, which is process rather than a bias evaluation.

The recourse framework here is statutory rather than contractual. Limbic Access is a certified Class IIa medical device, which brings obligations enforced by a regulator rather than negotiated with a customer: a stated intended use the product must not exceed, post market surveillance, vigilance reporting of incidents, and the possibility of enforcement or withdrawal.

A buyer therefore has a route that does not depend on their own contracting leverage, and a stated performance level to hold the product to, 93 percent classification accuracy tied to a certified intended use rather than floated as marketing. The commercial layer is bare.

No service level agreement, warranty, indemnity or remediation commitment is published, no accuracy breakdown by disorder appears, and nothing addresses what happens when an assessment misses an acute risk presentation, which is the failure mode with the gravest consequence and the one a buyer most needs allocated.

The integration is real and nowhere specified. A self referral and assessment tool feeding a talking therapies service has to reach that service's patient management system, and Limbic's FAQ says it is interoperable with a number of record systems and can operate alongside patient management systems, so staff no longer update patient information from intakes and referrals by hand. That is the claim; the specification is missing.

No record or case management system is named, no interface standard is described, no connection mechanism is given, and nothing states whether assessments write back automatically or are transferred by staff. Nothing addresses what integration exists in the United States, where the systems are different.

Nothing published establishes where the clinical service runs. No hosting provider, region, tenancy model or residency commitment appears for patient assessments; the website privacy notice names Heroku and Amazon Web Services, but it covers website visitors rather than the clinical product, and it commits only to keeping personal data inside the United Kingdom or the European Economic Area unless an adequacy decision or equivalent safeguard applies.

The question is live rather than formal. Health data from a national health system carries expectations about processing location and offshore access, and Limbic sells into the United States at the same time, so one organization handles mental health assessment conversations under two regimes; whether those estates are separated is not addressed.

The partial answer Limbic gives is contractual: it is not the end destination for patient data and works under the commissioning organization's data policies, which implies residency follows the customer rather than a vendor default.

Nothing about cost is published: no pricing page, unit of charge, range, implementation fee position, minimum commitment or pilot terms, and no proxy either, no cost per assessment or per referral and no clinician time released figure, though capacity relief is the commercial argument Limbic makes to a health system.

The procurement context explains part of it: national health system buying runs through framework agreements and local commissioning rather than list prices, so a public rate card would be unusual. It does not prevent publishing a unit of charge. The United States expansion sharpens the question, because a different market with different buyers needs a commercial model that is not described anywhere.

Coverage is narrow by specialty and deep within it. The setting is psychological therapy services, and the claim is precise rather than general: the eight disorder categories the national talking therapies program treats, across self referral triage, assessment, support between sessions and caseload management, with an ambient scribe added to the caseload assistant.

Penetration is the evidence behind it: assessments for more than 650,000 patients, roughly 45 percent of talking therapies services and 66 percent of integrated care boards in England, and the published referral study ran across 28 services.

Coverage outside that one national program is not evidenced: United States expansion is stated with no deployment scale, and American referral pathways, payer structures and assessment conventions differ from the program this product was designed around. Nor does anything address child and adolescent services, severe mental illness or substance use, which sit outside the talking therapies remit.