Wysa

The conversational agent is the product. A user opens the application and talks to a model that delivers structured therapeutic technique, and there is no platform, portal or workflow layer that would retain value without it. The company reports a billion artificial intelligence conversations, which is the product being used rather than a platform being administered.

One qualification applies. The premium consumer tier and parts of the provider offering include human coaches and therapists delivered through scheduled sessions, so the company sells labor alongside software. That is a stepped care design rather than a services business: the free and default experience is entirely automated, the human layer is an escalation above it, and the scale claims plainly describe the automated product rather than a coaching roster. The hybrid provider product places the model between therapy sessions with a clinician retaining the sessions themselves, which again positions the model as the continuous component and the human as the periodic one.

A second qualification is newer. The company has acquired a primary care behavioral health business and a physical therapy business, both delivering care through licensed clinicians. Those are care delivery operations that would exist without any model. The conversational product is where the company's description leads and where its evidence, its regulatory designation and its reach all sit.

One architectural fact matters too: some institutional deployments run primarily rule based chat rather than language model powered chat. A rule based dialogue tree is still the whole product in those deployments, so the product is the dialogue either way, but the learned component is not uniformly present across every configuration sold.

The boundary is published in both directions and the threshold inside it is not.

What the product is not for is stated plainly and in more than one place, which is the harder half to publish. An intended use statement describes the application as standalone software for self help and self monitoring of emotional wellbeing by individuals, says it is not a substitute for professional medical advice and provides no medical advice or diagnoses, and states that it is not recommended for crisis situations nor suitable for people with severe and enduring mental health problems. The user facing terms repeat it: this is not a crisis or emergency assistance service, and a person in crisis is told to call the relevant emergency services. Users are told unambiguously that they are talking to software and not a person, and that artificial intelligence does not replace human judgment.

What the system does is also now described rather than only asserted. The company states that its artificial intelligence never makes decisions on its own and that all care pathways include human oversight. On the children and young people line it goes further and names a mechanism: worrying triggers are stated to alert local mental health support services. The provider hybrid product places the model between sessions with a clinician holding the sessions themselves, which is a described division of labor.

What is missing is the operating point, and for a product of this reach it is the disclosure that matters most. Nothing states what counts as a worrying trigger, how sensitive that detection is, how quickly an alert reaches anyone, or whether the alerting path exists outside institutional deployments. The free consumer product is the default experience, is available to users from 13, and nothing published says whether a disclosure of intent there reaches a human at all or only returns a signpost. Across more than a hundred countries, crisis services differ entirely, and nothing describes how local protocols are engaged.

A scope exclusion tells a buyer where responsibility sits; it does not tell a user in distress what happens next.

The approach is named, the therapeutic content is specified, and nothing underneath either is identified.

The architecture is described at the level that matters most for a therapeutic chatbot. The company states that its artificial intelligence uses a combination of rule based algorithms and large language modeling, and, unusually, that the mix varies by deployment: some institutional partners offer primarily rule based chat, others a combination of rule based and language model powered chat. That is a material fact for a buyer, because it means two organizations can buy Wysa and deploy meaningfully different software, and the generative half carries risks the rule based half does not.

The clinical basis is specified too: cognitive behavioral therapy, behavioral activation, dialectical behavior techniques and mindfulness, all named modalities with established evidence bases, which lets a clinician judge whether the intervention is appropriate before judging whether the model delivers it well. A published intended use statement sets out what the software is and is not for.

What is absent is everything beneath the approach. No model provider, model family or version is named, no update or release cadence is stated, no model card exists, and nothing says which deployments run which configuration or how a customer learns that theirs has changed. For a product where the difference between rule based and generative is the difference between two safety profiles, version discipline is not a formality.

No performance figure is published outside the trial literature. Efficacy in a trial answers whether the intervention works; it does not answer how often the model misreads a message, misroutes a conversation or responds inappropriately, and no such measure exists publicly.

No model party is named. No foundation model provider, model family or version is identified anywhere, and no sub processor register is published, on a product the company itself describes as combining rule based algorithms with large language modeling.

That description makes the omission sharper rather than softer. A language model powered therapeutic conversation either runs on a model the company built or on one it licenses, and the answer decides whether a user's disclosure of distress is processed by a third party the customer has never been told about. Nothing says which, and the company states the configuration varies by institutional partner, so the answer may differ between two deployments of the same product.

One link is disclosed in a useful form. Cloud hosting is described as managed by third party providers operating under stringent privacy and health data programs, with service level and data protection agreements maintained with those hosting partners, and data processing agreements with service providers described as reviewed and maintained as an ongoing control. That is a supplier governance process rather than a list, which is better than silence and short of enumeration.

The training position is the remaining gap. Nothing states whether user conversations contribute to model development. The architecture makes that question less acute than it would otherwise be, since the application generally does not collect personal data and identifiers shared in conversation are hashed within 24 hours, so any corpus is substantially de identified by design rather than by policy. That is a real mitigation and it is not a statement.

Real independent trial evidence, and a deployment far wider than the indication it covers.

An independent peer reviewed clinical trial published in JMIR found the tool effective for chronic pain and associated depression and anxiety, more effective than standard orthopedic care and comparable to in person psychological counseling. Separate peer reviewed work reports that therapeutic alliance formed with the agent is comparable to that formed with human therapists, which matters because alliance is one of the strongest predictors of therapy outcome. A more recent peer reviewed study of 1,213 users found that access to the agent between coaching sessions made users nearly three times more likely to complete a coaching program, averaging 6.57 sessions against 0.55 for coaching alone, which is an engagement finding with a stated denominator.

Independent endorsement sits around it. The American regulator reviewed the trial evidence in granting Breakthrough Device Designation. Wellcome awarded a competitive grant of 5.3 million pounds in 2026 to fund a scale up study adapting the intervention for adolescent girls in rural India. The company reports more than 45 peer reviewed publications overall.

Deployment corroborates scale: more than six million users across more than a hundred countries, 11 million lives covered, a reported 185,000 patients through a national talking therapies gateway, a national platform in Singapore, and named institutional customers spanning a national health system, a health ministry, a pediatric academic center and several multinational employers and insurers.

The gap is between the evidence and the deployment. The trial addresses one narrow indication, adults with chronic musculoskeletal pain and coexisting depression and anxiety. The product is used by a general population many orders of magnitude larger, from age 13 upward, in more than a hundred countries and multiple languages. Efficacy in that cohort does not establish it for everyone using it, and the headline figures, 91 percent finding it helpful and a 33 percent reduction in lost time days, carry no citation, period or denominator.

Anonymity built into the architecture rather than promised, on a certified privacy management system, with one key question still unanswered.

The certification is the uncommon part. The company holds third party audited certification to the international information security standard and to its privacy extension, the latter being a privacy information management system specifically covering the handling of personally identifiable information rather than security alone.

The practice detail is specific and operational. The application generally does not ask for personal data at all, and users are told plainly not to share anything that identifies them; identifiers shared inadvertently are hashed within 24 hours. Random identifiers are used for transactions between application and servers. Data is encrypted in transit and at rest under a named cipher, with storage in the United Kingdom for the clinical program. Access is role based with two step verification, endpoint security is applied across staff systems, background verification and regular training are described, third party compliance audits run annually, and vulnerability scans and penetration tests are described as regular. Conversations with the agent and with human coaches are stated to be private and not shared.

The company also states plainly that conversations are never processed to form a diagnostic opinion or for medical purposes, a clear boundary on secondary use.

What is absent is an explicit position on whether conversations train models, and that is the gap that matters most. The company reports a billion artificial intelligence conversations. The de identification architecture substantially mitigates the question, and a corpus de identified by design is a materially better position than one de identified by policy, but mitigation is not a statement. Of de identification, retention and training use, only de identification is answered.

Four jurisdictions are named individually and the position is stated with candor about its limits.

Published compliance covers the American health privacy framework where applicable, European and British data protection law, and Indian information technology and sensitive personal data rules. The company is registered with the British data protection regulator, meets the national health system data security toolkit standards, and complies with the clinical risk management standard for health software manufacturers. Pseudonymization is defined and described as practiced rather than asserted, and the application's default position is to collect no personal data at all.

The candor is specific. The company qualifies the American framework with the words where applicable, and does not present the consumer application as covered by that framework for individual users, while enterprise and insurer deployments carry stronger obligations. That distinction is correct. A consumer downloading the application is not a patient of a covered entity, and saying so plainly is better practice than implying blanket coverage.

The instrument and the mapping are the gaps. No business associate agreement template, posture or execution requirement is published for the American provider market, and nothing maps which regime governs a given deployment, though the answer differs by product line and country.

The under 16 population adds a dimension nothing addresses. The application is intended for users from 13, or 11 where an institution agrees, and no consent model, parental notification position or safeguarding data flow is published for those users.

Several certifications, described with precision, and the artifacts are not reachable.

What is stated: a third party audited information security management system and privacy information management system, certified to the international information security standard at its 2022 revision and to its privacy extension at the 2019 revision, plus certification under the national cyber baseline scheme. The company complies with the clinical risk management standard for health software manufacturers, meets the national health system data security toolkit standards, and is registered with the British data protection regulator.

The language is disciplined. The company writes certified where it is certified and compliant where it is compliant, and gives the standard revisions. Naming the revision year of a standard makes the claim datable.

Supporting practice is published rather than implied: annual third party compliance audits, and regular vulnerability scans and penetration tests of applications and infrastructure.

Retrievability is the gap. No trust center is published as a consolidated page, no certificate or report is published, and no availability or request process is described, so the claim can be read and the artifact behind it cannot. The scope of each certification is also unstated, which matters for a company running a consumer application across more than a hundred countries, a provider line inside a national health service, and now acquired care delivery businesses: a certificate covering one of those is not a certificate covering all of them.

A real regulatory engagement, accurately described by the company and misdescribed by others.

Wysa holds Breakthrough Device Designation granted in May 2022 for an agent serving adults with chronic musculoskeletal pain and associated depression and anxiety. The designation is meaningful: the regulator reviewed the underlying peer reviewed trial evidence before granting it, and it commits the agency to expedited engagement on a device addressing an unmet need. It is not clearance and it is not approval, and a designated device is not authorized for marketing on the strength of the designation. That is a pathway stated and in progress.

A published intended use statement sits alongside it and strengthens the position. The application is described as standalone software for self help and self monitoring of emotional wellbeing by individuals, stated not to provide medical advice or diagnoses, not recommended for crisis situations, and not suitable for people with severe and enduring mental health problems, with an age floor of 13 or 11 where an institution agrees. Publishing an intended use statement with explicit exclusions is the artifact a regulator works from.

One outside description is wrong. At least one independent directory describes the company as holding an approved and cleared Class II device specifically cleared for clinical use. That is incorrect and would mislead materially.

The company's own conduct is the opposite. It appended an editorial note to its own 2022 announcement stating that the designation applies to that specific indication and does not represent blanket regulatory approval across its products. That voluntarily narrows its own regulatory claim.

A designation is not a certification or a clearance, and the indication it covers is far narrower than the deployment.

Governance is asserted and nothing is measured, against a fairness question this product's reach makes unusually pointed.

What the company publishes is a posture: artificial intelligence that never makes decisions on its own, human oversight stated across all care pathways, explainable and defensible support, an intended use statement with explicit exclusions, and a clinical risk management standard for health software that requires hazard identification and a safety case. Those are real commitments about how the product is bounded.

What is absent is evaluation. No bias or fairness testing, no performance breakdown by any population, no calibration data, no drift monitoring output and no independent audit of model behavior is published. An ethics board and a set of artificial intelligence ethics principles have been described in third party summaries; neither appears in the company's own material.

Fairness is the central question for this product. A conversational therapeutic agent operating in more than a hundred countries with multilingual support is encountering expressions of distress that differ by language, culture and idiom, and the same words carry different clinical weight in different contexts. The company is actively extending into exactly the populations where this bites hardest: a funded scale up study adapting the intervention for adolescent girls in rural India delivered in Marathi, and a national platform in Singapore serving teenagers and young adults. Whether the model recognizes distress equally well outside English, or in dialects and registers under represented in training, is unaddressed, and a person the system failed to understand leaves no complaint and no record.

The detection question compounds it. Recognition of risk is what the safety case depends on, and recognition is exactly the capability that varies by language.

Trial evidence establishes that the intervention can work; nothing establishes what happens when the product fails an individual user.

What exists on the positive side is real. Peer reviewed efficacy in a defined indication, regulatory review of that evidence through the designation process, compliance with a clinical risk management standard that requires hazard identification and a safety case for health software, and a published intended use statement that tells a user what the product is not for. The company also states that its artificial intelligence never makes decisions on its own and that support is explainable and defensible, which is a claim about auditability.

What is absent is any commitment or measure. No service level agreement, warranty, indemnity or remediation commitment is published. No conversational failure rate, no risk detection sensitivity, and no escalation performance figure exists. The explainable and defensible claim carries no artifact showing what a reviewer would see.

The allocation question is sharper here than for any administrative product. A user in acute distress interacting with an automated agent that does not recognize risk is the failure mode, and the consequence is not a denied claim. The free consumer product places no clinician in that path, is available from age 13, and runs in more than a hundred countries where crisis services differ entirely. The company's terms disclaim complete security and direct a user in crisis to emergency services, which is honest and is not a performance commitment.

Further efficacy or user scale figures do not touch this gap, and neither a published risk detection measure nor a stated commitment on escalation exists.

Integration is necessarily real for one product line and nowhere specified.

The consumer application needs no integration at all, which is a legitimate architectural position and part of why it reaches the scale it does. The provider products are different: a gateway feeding a national talking therapies service, a navigator, a clinician facing copilot, and a collaborative care line all require connection to the service's case management system to be useful, and a reported figure of more than 185,000 patients through that gateway indicates it works 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 assessment or engagement data flows back into the clinical record. The company does meet the national procurement assessment criteria, which include an interoperability section, so evidence exists in submissions that is not public.

The multi country footprint compounds it, since each health system runs different infrastructure and nothing indicates what integration exists outside the one national program. The children and young people line raises it again against school and local service systems rather than clinical ones, and the stated mechanism by which worrying triggers alert local mental health support services implies an outbound connection that nothing describes.

Graded for capability evidenced by deployment with no published specification.

A good deal is answered, without amounting to a full residency position.

What is published: storage servers for the clinical program are stated to be United Kingdom based, data is encrypted at rest under a named cipher and in transit under transport layer security, hosting is provided by third party cloud operators running privacy and health data compliant control programs, and service level and data protection agreements are maintained with those hosting partners. Naming the storage jurisdiction for a specific product line is a concrete residency statement.

What is not published is the rest of the picture. The cloud provider is not named, no region list exists for the consumer product operating across more than a hundred countries, no tenancy model is described, and nothing states whether data from one jurisdiction can be accessed from another. The company operates under American, British, European and Indian regimes with substantial Indian engineering presence, so personnel access location is a live question that a United Kingdom storage statement does not answer. A funded program in rural India and a national platform in Singapore each raise it again in a different form.

The acquired care delivery businesses add a further unknown, since delivered care generates records of a different kind and nothing states where those rest.

The clinical line carries a specific jurisdictional commitment, and the rest is open.

One product line has a published commercial model and the other has none.

The consumer offering is transparent: the basic conversational product is free, and a premium subscription adds unlimited use and access to human coaches, described as available in some locations at additional cost. A prospective user knows what costs money and what does not before engaging.

The provider and enterprise line publishes nothing. No pricing, unit of charge, per patient or per employee rate, minimum commitment or contract term appears for the health system, payer, employer or children and young people products that the company describes as driving its growth. The named buyer segments span a national health system, a health ministry, insurers, multinational employers and schools, which are five different commercial models, and none is described.

The two tier structure raises a question the company does not address: whether the free consumer product and the commissioned provider product are the same software, and if so what a health system is paying for that a member of the public receives at no cost. The company's own disclosure gives the beginning of an answer, since institutional deployments differ in whether they run rule based or language model powered chat, but nothing connects that to price.

Very wide reach across geography, channel and buyer type, with depth of deployment undisclosed.

Geographic coverage spans more than a hundred countries with multilingual support and access through a widely used messaging platform as well as the application, which is a deliberate accessibility design rather than a distribution accident. Buyer coverage runs across direct to consumer, employers, insurers, providers, national health bodies and a national health ministry, with named customers in each, and dedicated lines for primary care, collaborative care, a national talking therapies gateway and children and young people.

The age scope is stated precisely: intended for individuals aged 13 and over, or 11 and over where a children and young people institution agrees. A pediatric line is sold, a national platform in Singapore serves teenagers and young adults, and a funded study is adapting the intervention for adolescent girls in rural India. That is a real pediatric position rather than silence, and it raises its own questions about consent and safeguarding.

Clinical coverage is common mental health presentations, principally low mood, stress and anxiety, with one specifically evidenced indication in chronic pain with coexisting depression and anxiety. Severe and enduring mental illness and crisis care sit outside the stated scope, correctly and explicitly.

The company has also acquired adjacent care delivery businesses, one in primary care behavioral health and one in physical therapy, which extends the footprint beyond conversational support into delivered care.

Breadth of reach is not depth of deployment, and usage by market is not disclosed, so the difference between a country with a commissioned service and one with app downloads cannot be established. Language coverage is claimed without the list.