Medwise.ai

The model is the engine of the core module, and the platform carries other value around it. The product is now an answer engine with selectable AI modes: a reasoning mode that searches trusted sources, a basic mode, a web mode that expands scope beyond them, a drug mode that reads UK summaries of product characteristics and the drug tariff, a writing mode that polishes drafts, a local services mode and a support mode, with file and photo upload and audio recording. The app store listing calls it a next generation search platform for healthcare education, and the terms state that the platform involves an AI interface, a recommender system and automated decision making when returning results and generating responses.

Around that engine sit things that stand without it: continuing professional development logging, a document folder, clinician authored articles in the Clinical Evidence Hub, and search across UK national guidance.

The company's own peer reviewed pilot (JMIR Human Factors, 2025) tested an earlier AI supported, context sensitive search engine and concluded that adding large language models might improve accuracy and speed; the generative modes now on the platform postdate that engine.

The oversight structure is described and one part is missing. The user chooses how far the engine ranges: a reasoning mode that searches trusted sources, a basic mode, and a web mode that expands scope and sources beyond them, with a drug mode confined to UK summaries of product characteristics and the drug tariff. Answers link to their sources, and the interface tells users to always visit the source to verify information and use their clinical judgment. The terms set the rest: the platform is for information retrieval, guideline summaries and education, AI can make mistakes, every output should be verified independently, it is not intended to replace human decision making, and use by the public or by anyone under 18 is prohibited. File uploads are marked experimental.

What is missing is the threshold at which the system stops or defers: nothing says what an answer does when UK national guidance and a web source disagree, whether the engine declines when its trusted sources do not cover a question, or how a superseded guideline is detected. No absolute accuracy claim is made by the company; app store reviews calling it accurate are users' words.

The vocabulary and the modes are published and nothing under them is identified. The terms name an AI interface, a recommender system and automated decision making in returning results and generating responses; the home page names a reasoning mode, a web mode, a drug mode reading UK summaries of product characteristics and the drug tariff, a writing mode and others; the app store listing calls the product a next generation search platform. No model class, foundation model, model provider or version is named, no retrieval or ranking method is described, and no release notes say when the answer engine changed; the app's version history describes interface fixes only.

The company's own peer reviewed pilot (JMIR Human Factors, 2025) described an earlier AI supported, context sensitive search engine and said large language models might improve it; which models now generate the answers is not published. Nothing says how UK national guidance, web sources and drug references are ranked against each other, or how a superseded guideline is detected and retired.

The architecture is described and no model provider is named. The platform generates answers with AI across several modes, so a clinician's questions and uploads reach a model, and neither the model nor its provider is identified in the terms, the privacy notice, the home page or the app store listing.

The privacy notice names recipients only by category: IT and website hosting, communications and analytics providers, CRM, email and marketing automation providers, professional advisors, insurers and sub contractors engaged to deliver a service. It says third parties may not use personal information for their own purposes and process it only on the company's instructions, and that transfers outside the UK and EEA rest on adequacy decisions or approved standard contractual terms. It also discloses a second, outward flow: anonymized analytics, reports and insight products derived from platform use may be shared with or sold to analytics providers, research partners and commercial counterparties.

No subprocessor list is published, and no retention or training terms are stated for whichever model provider receives prompts.

One peer reviewed study of the product is published, it tested an earlier version, and it found the product slower. Ejaz, Tsui, Patel, Ulloa Paredes, Knights, Aftab and Subbe, JMIR Human Factors 2025;12:e52358 (doi 10.2196/52358): a prospective pre and post observational time and motion study in acute medicine at Ysbyty Gwynedd, Bangor, in which searches by 10 doctors using the hospital intranet over 10 working days were compared with searches by 10 doctors using an AI supported, context sensitive search engine over a further 10 days, after a focus group of 14 clinicians. Doctors using the engine ran fewer searches, user satisfaction and query resolution were similar, searches took 43 seconds longer on average, and the Net Promoter Score was 20. The cohorts differed in experience (median 23 against 54 months). Three authors, including cofounder Keith Tsui, are company employees and the study was led from the hospital. The authors call it a feasibility pilot, say incorporating large language models might improve accuracy and speed, and say more research is needed to establish clinical impact, so it describes the engine before the current generative modes.

Nothing published measures whether the current platform's answers are right. The company's scale claim, thousands of clinicians across more than 2,000 NHS organizations, records use rather than benefit, and the named practices on the home page are customers without outcome figures.

The data practices are stated precisely, and what they state is a use of conversation data rather than a protection of it. The privacy notice (last updated 8 April 2025) lists conversation content and interaction data (prompts, messages, uploaded content, outputs, timestamps and metadata) among what the company collects. A conversation removed from the visible history may be retained for up to 90 days for security, fraud prevention, compliance, debugging and service administration, and for up to six years where needed for legal claims, regulatory compliance or investigating misuse. Where conversation content is used for analytics or insight products it is anonymized within 90 days where reasonably possible, and anonymized information may be used for research, product development and commercial insight products, including sharing it with or selling it to selected third parties. The terms grant the company a worldwide, perpetual, transferable license to content users upload, including to promote the platform.

The upload screen tells users not to upload images of people or medical images, or documents containing personally identifiable information, and says detected personal information is removed automatically and uploaded files are not saved. Whether conversation content trains any model is not stated.

Retention and deidentification are stated specifically enough to be contradicted. However precisely it is worded, the stated practice sells derived data onward. No safety engineering is published beyond the automatic removal of detected personal information: no guardrails, no hallucination mitigation and no process for handling a safety event.

The United States rule does not govern this company and it does not invoke it; the UK regime is documented to depth. Medwise AI Ltd, registered in England and Wales with an office in Leeds, states in its privacy notice (last updated 8 April 2025) that it is the data controller, gives its Information Commissioner's Office registration (ZB034856), sets out what it collects including conversation content, the lawful basis for each use, named retention periods (conversation content up to 90 days after removal from history, up to six years for legal claims or misuse), safeguards for transfers outside the UK and EEA, data subject rights, and the route to complain to the ICO. The notice reaches the platform and the app, not only the website.

The product is scoped away from patient data: the upload screen prohibits documents containing personally identifiable information and medical images, and the terms prohibit patient specific use. No data processing agreement, Data Security and Protection Toolkit status or data protection impact assessment is published for NHS organizations, which are the instruments a buyer in this market would ask for.

Controls are asserted with nothing independent behind them. The privacy notice says the company applies appropriate technical and organizational measures, including pseudonymization, encryption, access and retention policies, and that no system is completely secure; the terms say the company does not guarantee the platform will be secure or free from bugs or viruses. No ISO 27001 certificate, Cyber Essentials or Cyber Essentials Plus certificate, SOC 2 report, Data Security and Protection Toolkit status, penetration testing statement or trust center is published on the home page, terms, privacy notice, cookie policy, Clinical Evidence Hub or app store listing.

Suppliers to NHS organizations are routinely asked for a defined compliance pack during procurement, and the company reports use across more than 2,000 NHS organizations; none of that material is published.

A regulatory position and a statement of intended purpose are published, with no reasoning behind the position. The terms of use state that the platform is intended for medical information retrieval, summaries of guidelines and educational purposes only, that it is not a regulated medical device, and that it is not intended to diagnose or treat any health condition, provide patient specific diagnostic or treatment recommendations, or substitute for a healthcare professional's clinical assessment and judgment. Use by the public in a private capacity is prohibited. The app store listing repeats that results are for educational and informational purposes.

That positively describes what the product is for and who may use it, under the UK regime that governs the company. What is missing is the reasoning. The company publishes no analysis of why the platform sits outside medical device regulation in the UK, and claims no MHRA registration. The product's modes (a reasoning mode, file and photo upload, a drug mode reading summaries of product characteristics) invite questions close to a patient, which is where a stated intended purpose is tested; the terms' prohibition on patient specific recommendations is the published line.

Responsible AI appears in policy language with no evaluation behind it. The terms open with an AI notice stating that the platform involves an AI interface and recommender system, that good AI governance means telling users so expressly, that returning results and generating responses involves automated decision making, and that it makes no decisions about people with legal or similarly significant effects. The same terms say the platform uses experimental and novel technologies, that AI can make mistakes, and that every output should be verified independently. The Clinical Evidence Hub carries clinician authored articles under a published editorial policy, written around changes in guidance and MHRA safety updates with the source trail preserved, which governs that content rather than the model's answers.

Nothing is published on how the answer engine is tested: no accuracy evaluation, no error taxonomy, no bias or subgroup analysis, no model card, and no description of how the web mode's wider sources are weighed against national guidance.

The privacy notice adds a governance fact of its own. Conversation content is used to create anonymized analytics and insight products that may be shared with or sold to selected third parties.

Mechanisms exist that let a clinician challenge an answer, and nothing stands behind the output. Every answer links to its sources, the interface tells users to always visit the source to verify information and use clinical judgment, the Clinical Evidence Hub preserves the source trail for its articles, and a feedback form is linked from the app menu.

The terms of use state that AI can make mistakes, that all outputs should be verified independently, and that the company makes no representation or warranty that platform content is accurate, complete or up to date; the platform is provided as is. For open access free of charge users, liability is excluded for indirect losses and capped at 100 pounds in any rolling twelve months, with claims to be notified within twelve months; subscribers' limits sit in a subscription agreement that is not published. Users, by contrast, indemnify the company without cap for breaches of the terms.

No accuracy figure, error rate or correction commitment is published, and nothing says what happens when a clinician reports a wrong answer.

No integration evidence is published. The platform is a standalone web application and mobile app reached by an individual clinician's own login, with NHS mail sign in as one route; that is authentication, not integration. No record system, primary care system, marketplace listing, application programming interface or health data exchange standard is named on the home page, terms, privacy notice or app store listing, and the upload screen tells users not to upload documents containing personally identifiable information, so no patient record is expected to reach the product.

That is consistent with a clinician reference used outside the record, and it leaves a point of care tool that the clinician must switch to rather than one embedded in the systems NHS clinicians work in all day.

A single hosted option with location implied rather than committed. The platform is a web application with iOS and Android apps, reached by individual clinicians with NHS mail sign in among other routes, from Medwise AI Ltd, registered in England and Wales with its office in Leeds and registered with the Information Commissioner's Office (ZB034856).

The privacy notice names hosting providers only as a category and says that whenever personal information is transferred out of the UK or the EEA, the company relies on adequacy decisions or on standard contractual terms approved in the UK and the EU. That tells a buyer data may leave the UK and does not say where it rests or which processing crosses a border. No hosting provider, cloud region, tenant isolation or on premises option is published.

A price is published without full tiers. The terms of use set out two routes: open access free of charge for an individual healthcare professional or student, subject to use limits, and a paid subscription the company may require above certain thresholds or on other criteria. The company's App Store listing, a field the developer controls, shows the app as free with in app purchases and names one: Medwise AI subscription at 69.99 pounds. The home page offers a Plus account and sign up through NHS mail. A clinician can start without a sales conversation.

What is not published is the subscription's term (monthly or annual), what it adds over the free tier, where the free use limits sit, and whether any organization level license exists. The terms refer to a formal subscription agreement for subscribers, with its own liability limits, that is not published.

Coverage is stated clearly, for one national health system, with nothing validating the current product yet. The platform serves UK clinicians and students: sign in through NHS mail, a drug mode reading UK summaries of product characteristics and the drug tariff, clinician authored Clinical Evidence Hub articles organized by clinical category (acute and critical care, cancer, contraception, diabetes and endocrine, immunization, infections, medicines management, skin, urology), and named general practices on the home page. The terms prohibit use by the public and by anyone under 18.

The one published evaluation (JMIR Human Factors, 2025) covers acute medicine at one Welsh hospital and an earlier version of the engine, so it does not validate the current product in any setting. No coverage outside the UK is claimed, and nothing measures performance by specialty or question type.