Limbic

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

The regulatory status is the strongest available confirmation. Software is classified as a Class IIa medical device precisely because its output carries clinical consequence rather than administrative convenience, and a regulator assessed it on that basis. Nothing else in this index has had its centrality externally adjudicated.

The newer caseload assistant adds an ambient scribe under a separate Class I classification, which is a lower risk administrative function, and does not dilute the core.

A published accuracy figure, an externally audited risk management framework and a stated human boundary, held below A by one specific and consequential gap.

What is disclosed is substantial. Classification accuracy is stated at 93 percent across eight named disorder categories. The tool is positioned to augment rather than replace human led clinical assessment, with a therapist retaining the assessment decision. Device certification at Class IIa mandates documented processes covering clinical risk, safety risk, usability risk and security risk, externally audited, which is an oversight framework no competitor in this index operates under.

What is missing matters more here than the aggregate accuracy suggests. No breakdown of error by disorder or severity is published, and in mental health triage the errors are not symmetric: a missed presentation carrying suicide or self harm risk is categorically different from a misclassification between two anxiety presentations, and a single aggregate figure describes both together. Nothing located describes the risk escalation protocol, what the system does when a patient discloses acute risk mid conversation, or how urgent presentations are routed to a human immediately.

Ask for accuracy by disorder, the false negative rate on risk presentations, and the crisis escalation pathway.

The first A on this axis in the index, and it rests on peer review rather than marketing.

The architecture is named and, per the company, described in the peer reviewed literature: a cognitive layer the company calls the Limbic Layer, sitting above foundation models and specialising them for clinical delivery. Publishing a method in a journal subjects it to a standard of description no product page reaches, and makes it checkable by a buyer's own clinicians rather than only by their procurement team.

Output scope is enumerated precisely, the eight disorder categories treated by the national talking therapies programme, and performance is stated as a number rather than an adjective at 93 percent. The Class IIa technical documentation underpinning certification is itself a body of specification that a regulator has reviewed.

Two qualifications keep this honest. No foundation model provider or version is named beneath the cognitive layer, and no accuracy breakdown by disorder is published. And the publications were not retrieved and read in this pass, so the grade reflects that a peer reviewed description is reported to exist rather than verification of its contents. Ask for the papers and the underlying model provenance.

The existence of the chain is acknowledged more clearly than by most, and no party in it is named.

The company describes its cognitive layer as enabling foundation models to deliver therapy, which concedes plainly that external general purpose models sit beneath its own specialisation. That is more candour than the proprietary label most vendors use, and it tells a buyer the right question to ask.

It does not answer it. No model provider, family or version is identified. No cloud platform is named. No sub processor list was located.

One mitigating factor is structural rather than disclosed. Class IIa certification mandates a traceable and controlled process for design, development and testing, and supplier controls form part of the quality system a certification body audits, so a documented chain exists and has been externally reviewed even though its contents are not public. That is genuinely more assurance than an uncertified vendor's silence, and it is not the same as publication.

The training data question is unanswered in either direction, which matters given the corpus is mental health assessment conversations. Ask for the model provider, the sub processor register, and whether patient conversations contribute to model development.

The strongest evidence base in this index by a clear margin, and the only record supported by a randomised controlled trial.

The company reports two publications in a leading peer reviewed medical journal: a 2024 study demonstrating improved accessibility for minority groups, and a randomised double blind clinical trial reporting that specialist trained clinical agents delivered cognitive behavioural therapy outperforming both human clinicians and general purpose large language models. Every other record in this index rests on satisfaction surveys, analyst rankings, vendor run studies or customer testimony. This is the only one claiming the design that clinical medicine treats as dispositive.

Deployment corroborates rather than substitutes for it: assessments delivered for more than 650,000 patients, coverage of 66 percent of integrated care boards in England and roughly 45 percent of talking therapies services. Class IIa certification required clinical evidence to be reviewed by an external certification body, which is independent scrutiny of the evidence itself rather than of customer satisfaction.

The caveat is stated plainly and does not lower the grade so much as bound it. Neither paper was retrieved and read in this pass. A claim that an artificial intelligence outperformed human clinicians in a double blind trial is extraordinary, and a buyer should read the paper, its endpoints and its comparator arm before relying on it.

Substantially better than most records here, with one central question unanswered.

Published practice includes de identification as a described operation rather than a policy: incoming referral records are logged and patient identifiable information is stripped from the record. The company also states it is not the end destination for patient data and operates under the data policies of the commissioning provider organisation, which is an accurate description of a processor relationship and sets expectations correctly. Data subject rights are enumerated individually, a data protection officer is named, and breach notification is committed to. Device certification at Class IIa mandates documented handling of security risk as part of the audited quality system.

The unanswered question is whether patient conversations contribute to model development. Nothing located states a position either way. That gap is more consequential here than in most records because of what the data is: transcripts of people describing depression, anxiety and, in some proportion of cases, thoughts of self harm. Vendors elsewhere in this index handling far less sensitive material commit explicitly that customer data never trains their models.

No retention schedule with periods was located either. Ask for the training position on patient conversations and the retention periods by data type.

Compliance is claimed across both regimes the company operates in and, unusually, the detail behind it is published rather than asserted.

What is stated: compliance with both the United States health privacy framework and European data protection law, a named data protection officer with a published contact route, individually enumerated patient rights covering access, rectification, erasure, restriction, portability and objection, a route to the national data protection regulator, and a breach notification commitment. The national clinical risk management standard for health software manufacturers is also cited, which is a procurement gate for the health system it primarily serves.

One statement is notable and rare. The company states 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. That is a direct engagement with the automated decision provision of European data protection law, and a forward commitment rather than a present tense claim.

Held at B because no business associate agreement posture, template or execution requirement was located for the United States market it is expanding into, and no retention schedule with stated periods was found. Ask for the agreement position and retention periods.

Several credentials are named and one important distinction is being blurred, which is why this sits at B rather than higher.

The company states it is compliant with the principal international information security standard and separately that it holds a certification under the national cyber baseline scheme. Those are not equivalent claims. Compliant with a standard is a self description requiring no external assessor; holding a certification means an accredited body audited and issued it. The company uses the precise word for one and the looser word for the other in the same sentence, and this index records the difference rather than reading them as the same. This is the third instance in this session of a credential stated in a form weaker than it appears, after one vendor presenting alignment with a framework as a badge and another presenting a compliance framework with no certification scheme as a certification.

What is unambiguous is stronger than most. Class IIa device certification mandates audited processes covering security risk alongside clinical and usability risk, and the national clinical risk management standard for health software manufacturers is separately cited. Both are externally assessed.

No trust centre, report availability process, audit period, penetration testing disclosure or vulnerability disclosure policy was located. Ask whether the information security standard is certified or self assessed, and by whom.

The first A on this axis anywhere in this index, and the reason every other record reads differently is that this is the first product that is actually a regulated medical device.

Every other vendor built in this sweep carries a note that no device pathway applies and none is claimed, because coding, scheduling and documentation are administrative acts. This product makes a clinical determination, classifying a patient's presenting disorder, and is regulated accordingly. Limbic Access holds Class IIa certification under the United Kingdom conformity regime, obtained January 2023, audited by a named certification body against clinical effectiveness, safety and risk management, and was the first artificial intelligence mental health chatbot in the world to achieve it. The company states the certification mandates documented processes for clinical, safety, usability and security risk and a traceable controlled process for design, development and testing.

A second product carries its own separate and lower classification at Class I for the ambient scribe component, which is the correct treatment for an administrative function and shows the classification boundary being applied within the portfolio rather than claimed across it.

The gap is jurisdictional. The company is expanding into the United States and no clearance, submission, breakthrough designation or enforcement discretion position was located for that market, where the same functionality would face a different regulator. Ask for the United States regulatory pathway and its current status.

Genuine published equity research, and a distinction that keeps this at B rather than A.

The company reports a 2024 peer reviewed publication demonstrating improved accessibility for minority groups, which is real published evidence bearing on equity and the only such publication associated with any record in this index. It also participates in national work with a health system membership body on safe adoption and implementation of clinical artificial intelligence, and operates under device certification requiring externally audited risk management.

The distinction matters and should not be blurred. A study showing that more patients from minority groups self referred through the tool is an access finding. It is not a demonstration that the classifier performs equally well across those groups once they engage. Those are different questions, and the second is the one this axis asks: whether a patient from one background is as likely to be correctly classified as a patient from another. Conversational assessment is precisely where differences in language, idiom and how distress is expressed could produce differential accuracy.

Nothing located reports performance parity across demographic groups, and no ongoing monitoring output, drift reporting or independent bias audit was found. Ask for classification accuracy broken down by ethnicity, first language and age.

The recourse framework here is statutory rather than contractual, which is unique in this index and is worth setting out precisely.

Every other record graded on this axis has been assessed against what the vendor will commit to in a contract, and nearly all publish nothing. This vendor operates under medical device regulation, which imposes 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 an avenue that does not depend on their own contracting leverage.

A stated performance level exists to hold the product to, at 93 percent classification accuracy, and it is tied to a certified intended use rather than floated as marketing.

Held at B rather than A because the commercial layer is bare. No service level agreement, warranty, indemnity or remediation commitment was located, no accuracy breakdown by disorder is published, and most importantly no disclosure 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. Ask how liability is apportioned on a missed risk presentation.

Integration is necessarily real and nowhere specified. A self referral triage tool that feeds assessments into a talking therapies service must write into that service's case management system, and doing so across roughly 45 percent of those services means the integration works at scale in production.

No detail is published. No case management or record system is named, no interface standard is described, no connection mechanism is specified, and nothing states whether assessments write back automatically or are transferred by staff. The national procurement framework the company must satisfy includes an interoperability section, so evidence exists in submissions that is not public.

The United States expansion makes the omission more consequential going forward, since the systems in that market are entirely different from those the product currently connects to, and nothing indicates whether integration there is built or planned.

Graded C for capability evidenced by deployment scale with no published specification, consistent with the treatment of the same absence across this index. Ask which case management systems are supported, through what standards, and what United States integration exists.

Nothing was located on hosting, region, tenancy or residency.

The question is live rather than formal for two reasons. Health data from a national health system carries specific expectations about processing location and offshore access, and the company is simultaneously expanding into the United States, which means a single organisation is handling mental health assessment conversations under two regimes with different rules. Whether those are segregated, and where each is processed, is unaddressed.

The partial answer available is contractual rather than technical. The company states it is not the end destination for patient data and operates under the data policies of the commissioning provider organisation, which implies residency follows the customer's requirements rather than a vendor default. That is a reasonable posture and it is not a statement of where anything actually runs.

Graded C consistent with how silence has been treated across this index. Ask for processing and storage regions by jurisdiction, the tenancy model, and whether the two markets are segregated.

Nothing about cost is published. A dedicated pass located no pricing page, no unit of charge, no range, no implementation fee position, no minimum commitment, no pilot terms and no return calculator.

Nor is there a proxy. No cost per assessment, no clinician time released figure, and no cost avoidance estimate against the alternative of human assessment, despite capacity relief being the entire commercial argument the company makes to a health system.

The procurement context explains part of it without excusing it. National health system purchasing runs through framework agreements and local commissioning rather than list pricing, so a published rate card would be unusual. That does not preclude publishing a unit of charge or a cost per referral, and competitors selling into the same system manage more.

The United States expansion sharpens the question, since a different market with different buyers will need a commercial model that has not been described anywhere. Ask for the pricing basis, whether charge is per assessment, per referral or per service, and how the two products price relative to one another.

Deliberately narrow by specialty and unusually deep within it.

The setting is psychological therapy services, and the coverage claim is precise rather than general: the eight disorder categories the national talking therapies programme treats, spanning self referral triage, assessment, support between sessions and caseload management. Depth of that kind in one specialty is more useful to a buyer in that specialty than breadth would be.

Penetration is the evidence. Roughly 45 percent of talking therapies services and 66 percent of integrated care boards in England represents national scale within a single system, which few records in this index approach.

Held at B rather than A on two counts. Coverage outside that one national programme is not evidenced: United States expansion is stated as intent with no deployment scale disclosed, and the American mental health system has entirely different referral pathways, payer structures and assessment conventions from the programme this product was co designed around. And nothing addresses child and adolescent services, severe mental illness, or substance use, which sit outside the talking therapies remit. Ask for United States deployments and coverage beyond common mental health disorders.