Nelson Advisors: Ada Health is quietly building well governed and safe infrastructure layers and products for Healthcare AI
The loudest healthcare AI companies are not the ones building the foundations
In a market where every week brings a new announcement about a chatbot that can "act as your doctor", the most interesting healthcare AI company in Europe is one that has spent fifteen years refusing to say that. Ada Health, founded in Berlin in 2011, has never claimed to diagnose anyone. Its consumer app says so on the App Store listing, in capitals. Its enterprise documentation says so. Its clinical studies are designed around the assumption that Ada supports a clinician rather than replacing one.
That restraint is easy to mistake for a lack of ambition. It is the opposite. While the current generation of general-purpose language models has rushed into health questions at enormous scale, Ada has been doing something slower and harder: building the regulated, auditable, clinically validated layers that would let any AI, including a large language model, be used safely in front of a patient.
A certified medical device. A quality management system. A probabilistic reasoning engine that can show its working. A medical knowledge base curated by physicians rather than scraped from the web. Fifty-plus peer-reviewed publications, including one in NEJM AI this spring. And, as of March 2026, a European patent on the architecture that fuses all of that with an LLM.
Our article argues that Ada is quietly assembling exactly the kind of infrastructure the rest of the healthcare AI industry will need, and that the patent announced in April is less a product launch than a statement about where the safety boundary in clinical AI should sit.
It looks at what the patent actually protects, why the symptom checker underneath it is unusually well credentialed, how Ada functions as a first port of call inside real healthcare systems, where the company is being deployed as plumbing rather than as a consumer app, and what governance looks like when it is treated as a product feature rather than a compliance cost. It also takes the limits seriously, because a company that is honest about what it cannot do deserves to be assessed on the same terms.
The patent: a clinical layer that makes LLMs usable in healthcare
On 13 April 2026 Ada announced that the European Patent Office had granted patent EP4679451B1 for its hybrid clinical AI architecture. The grant took effect on 25 March 2026 with a priority date of 12 July 2024. The headline in Ada's own release was unusually direct for a company that normally avoids grand claims: it called the invention "the clinical layer that makes LLMs safe for healthcare".
What is being protected is an ensembling approach rather than a model. Ada has run a probabilistic graphical model (PGM) over an expert-curated medical knowledge base since 2016. The patent covers a way of wrapping a large language model around that engine so that the two reinforce each other instead of competing. The release describes three structural elements.
First, the LLM sits at the front. A user can describe symptoms, medical history, medications and preferences in their own words instead of tapping through a fixed questionnaire. Ada's argument is that free text captures the contextual detail that matters most for complex or rare presentations, which rigid forms tend to lose.
Second, the follow-up questions the system asks are constrained to Ada's validated knowledge base. The LLM can prompt dynamically, but it cannot invent a line of questioning that the clinical model does not recognise. Ada describes this as a structural safeguard against the ungrounded outputs that unconstrained LLMs produce in clinical settings.
Third, and most importantly, the two models assess the case independently and are then ensembled. The PGM is the decision-maker and acts as a guardrail; the LLM functions as a second opinion. Ada reports that the combination improves relative accuracy by more than 7% over either approach alone while remaining, in its words, a fully white-box, auditable system. CTO Graham French summarised the division of labour as the LLM capturing what the user wants to say, the probabilistic model determining what it means clinically, and Ada's reasoning engine always holding the final decision.
The framing matters because it inverts the usual pitch. Most healthcare LLM products present the language model as the intelligence and bolt on safety filters afterwards. Ada's patent treats the LLM as an interface and a cross-check, and keeps the regulated, explainable component in charge. That is why the company can say the hybrid is deployable in regulated environments: the part that makes the call is the part that already carries CE marking and a decade of validation data. Founder and CEO Daniel Nathrath put the competitive point bluntly, observing that LLM providers are learning that scale and medical safety are not the same thing.
A patent is not proof the architecture works, and EPO grants in software are narrower than the press releases that announce them. But the hiring behind it is telling. Ada's current job listings for senior AI engineers describe the core work as refining the interface between LLM-based reasoning and probabilistic logic, maintaining internal evaluation and validation frameworks, and producing technical documentation for regulatory compliance. That is the job description of a company building infrastructure, not one chasing a demo.
The most credentialed pure symptom checker on the market
Underneath the patent is a product that has been doing the unglamorous work for a decade. The Ada app is free on iOS and Android, available in seven languages (English, German, French, Portuguese, Spanish, Swahili and Romanian, according to the App Store listing), and runs a structured questionnaire against a medical dictionary of thousands of conditions. It returns a report of possible causes ranked by likelihood, an urgency level, and a recommended next step. It does not hold a conversation and it does not pretend to be a clinician. The same listing carries the warning, in capitals, that the services cannot give you a medical diagnosis.
What separates Ada from the dozens of symptom checkers that look similar on a phone screen is the regulatory and evidentiary stack behind it. Ada Assess, the enterprise version of the assessment engine, was certified in December 2022 as a Class IIa medical device under the EU Medical Device Regulation, among the first AI symptom assessment platformsto reach that class. Class IIa is the tier that covers software intended to inform clinical decisions, and the certification covers not just the assessment module but the care navigation, clinical handover, FHIR interoperability and authentication modules around it, according to Ada's enterprise help pages. Ada's quality management system is certified to ISO 13485, the medical device standard, and its information security to ISO 27001. In the UK the enterprise product is UKCA marked and registered with the MHRA. In Brazil it holds Class II registration. The consumer app is a Class I device with Class IIa pending.
The evidence base is equally unusual for a consumer-facing tool. Ada claims more than fifty independent peer-reviewed publications. The most cited is a BMJ Open study from December 2020 that benchmarked eight symptom checkers against a panel of seven GPs using 200 clinical vignettes. GPs reached 82% accuracy on their top suggestion; Ada reached 71%, and on top-three and top-five suggestions it was comparable to the physicians. The next-best app scored 43%. Ada, Babylon and Symptomate also had the highest performance on the safety of their urgency advice. The study's lead authors were Ada employees and its methodology was reviewed by academics at UCL and Brown; the paper itself called for independent replication. That caveat is worth holding onto, but the design, which included testing whether each app would even accept a case involving a child or a pregnant woman, became a template that regulators and later studies have borrowed.
There is a further detail that gets missed in product comparisons. Ada's engine is a probabilistic graphical model, not a neural network. Every suggestion it makes can be traced back to the symptoms, risk factors and prior probabilities that produced it. That is what Ada means when it calls its system white-box. In a sector where regulators increasingly ask vendors to explain why a tool reached a particular output, a model that can produce that explanation by construction is a far easier thing to certify, audit and defend than one that can only be probed from the outside.
Taken together, the certifications, the published evidence and the explainable architecture make Ada something closer to a regulated instrument than to a consumer app that happens to ask health questions. That is the asset the patent builds on.
A first port of call that hands you back to the system
The clearest way to understand what Ada is for is to watch it inside a real healthcare system. Consider the experience of someone who has recently moved to Germany and wakes up unwell. The German system is excellent and bewildering in roughly equal measure. There is the Kassenärztliche Vereinigung appointment service, a dense ecosystem of Fachärzte specialists, rules about when a referral from a Hausarzt is needed and when it is not, and a Notaufnahme that will see you but would much rather you had gone elsewhere. A newcomer does not know whether a persistent headache with a stiff neck is a Monday-morning GP appointment, an urgent call, or a night at home with paracetamol. Ada's job is to answer precisely that question. It asks a structured series of questions about symptoms, history and risk factors, returns a ranked set of possible causes, and tells the user whether this looks like home care, a GP visit within a few days, or something that needs attention now. The app is available in German and English, and many German insurers and doctors recommend it for exactly this purpose.
Two design choices make this work. The first is that Ada returns a structured report rather than a conversation. A user can send that report to a GP as a PDF, and the clinician gets a differential and a reasoning trail rather than a chat log. The second is that there is no human layer inside Ada. It does not run a telehealth service, employ doctors or sell consultations. It assesses, it navigates, and then it hands the person back to the healthcare system they already belong to. That self-limitation is what makes it attractive to health systems and insurers, who would otherwise be integrating a competitor.
Until this year the evidence that Ada actually changes what people do was mostly indirect. That changed with the ESSENCE study, published in the April 2026 issue of NEJM AI and announced by Ada and CUF, Portugal's largest private healthcare network. The study enrolled 1,470 adults through myCUF, CUF's patient app, where Ada's assessment is linked directly to care booking. It was prospective, used electronic health record linkage to see what patients actually did in the following weeks, and had an independent physician panel rate the appropriateness of both intended and observed care.
The headline results are striking. One in three patients revised their intended level of care immediately after the assessment, and the share who said they did not know what to do fell from 12.6% to 5%. When actual behaviour was tracked, nearly three in five attended a different care setting than they had originally planned. Most importantly, the proportion of patients receiving care that the physician panel judged appropriate more than doubled, from 29.8% to 64.4%. Among those who had planned to go to an emergency department, nearly 40% chose a lower-acuity setting that the panel rated appropriate. Subsequent emergency visits and hospitalisations were tracked, and no concerning safety signals were found.
The study is observational, and Ada's own release is careful to say that the effect reflects the combination of the assessment and the bookable services around it, not the assessment alone. But that is the point. Ada's value is not that it is a clever quiz; it is that it plugs into the front door of a health system and redirects demand to the right place without suppressing it. For a system like Germany's, or the NHS, where the question of who should be in the emergency department is a daily operational problem, a tool that can reduce inappropriate attendance while increasing appropriate care is close to the ideal intervention.
Ada as infrastructure: the app is the demo, the platform is the business
The free consumer app is how most people meet Ada, but it is not where the company's revenue or its strategic weight sits. Ada's business is licensing its assessment, navigation and handover engine to organisations that already own a patient relationship and need a safe, regulated front door. That is the clearest sign that the company sees itself as a layer rather than a destination.
The enterprise product is modular. Ada Assess is the symptom assessment engine. Ada Connect routes the user to the right service within a partner's own ecosystem. Ada Handover passes the structured report to a clinician or a partner system. FHIR services provide interoperability with electronic health records, and the platform has been available on Epic's app marketplace since 2021. All of these modules fall within the Class IIa CE certification, which means a health system integrating Ada is integrating a certified device rather than a widget it must validate itself.
The partner list reads like a map of the places where safety and explainability are non-negotiable. In health systems, Ada has been deployed at Sutter Health in California, where it was embedded in the My Health Online patient portal for more than three million patients, and at Jefferson Health in Philadelphia. It has been the digital front door at CUF in Portugal since 2021, which is what made the ESSENCE study possible. In insurance, Santéclair in France extended its partnershipin 2024 so that ten million members across more than sixty plans could continue to use Ada for care navigation, and Groupe Mutuel in Switzerland integrated it into its member experience the same year. AXA has been a partner since 2021. In life sciences, Ada has worked with Novartis on identifying patients with under-diagnosed immunological and rare diseases, with Pfizer on a nationwide COVID-19 care journey in the United States, and, as of October 2025, with Fullspan Health, an RVO Health offering that reaches millions of Americans each month, to shorten the time to treatment in areas such as autoimmune conditions and women's health.
The pharma work deserves a closer look because it is where the infrastructure framing is most obvious. A pharmaceutical company cannot ethically or legally run a chatbot that nudges people toward its drug. What it can do is fund a certified, explainable assessment that helps people with vague symptoms get to a correct diagnosis sooner, with the navigation layer pointing to appropriate specialists. For rare and autoimmune diseases, where the average time to diagnosis is measured in years, that is a defensible public-health intervention. Ada's early rare-disease work, including a 2019 retrospective study of its diagnostic tool, laid the groundwork, and the company now markets a dedicated patient-finding line alongside its care-utilisation line.
The company's funding history reflects the same long game. Ada raised a $90 million Series B led by Leaps by Bayer in 2021, extended it to $120 million, and added a further €30 million later. It has not chased the valuations of consumer telehealth, and it has not pivoted to selling consultations. Bayer, Samsung and the other backers have been patient because the thing being built is slow to build and hard to copy.
The patent announcement says explicitly where this goes next: Ada will extend its hybrid LLM approach across its pharma, health-system and payer partners. In practice that means partners who already trust the probabilistic engine will get the natural-language front end without changing the thing that makes the decision. That is the clearest articulation yet of the infrastructure strategy. The LLM is a new interface to an existing, certified reasoning layer, and the certified layer is what partners are paying for.
Governance as a product feature, not a compliance cost
Most technology companies treat regulation as a tax: something to be paid once, late, and as cheaply as possible. Ada's history suggests it has treated regulation as a moat. Each step up the regulatory ladder has arrived before the market demanded it, and each has become a thing the company can sell.
The pattern is consistent. The consumer app launched in late 2016 and was CE marked as a Class I device early. The enterprise engine went through the far more demanding Class IIa process under the new EU MDR and was certified in December 2022, at a point when many digital health vendors were still arguing that their software was not a device at all. The quality system was certified to ISO 13485, which is the standard for medical device manufacturers rather than software companies, and the security programme to ISO 27001. The engineering roles Ada advertises today list regulatory documentation as a core responsibility rather than an afterthought. When the EU AI Act began classifying medical AI as high-risk, Ada was already operating inside a regime that was, if anything, stricter.
Explainability is the second pillar, and here the architectural choice made a decade ago is doing enormous work. Because the reasoning engine is probabilistic rather than neural, Ada can show, for any assessment, which symptoms raised and lowered the probability of each condition and why a particular urgency level was assigned. That is not a feature that was retrofitted; it falls out of the model. It is also the thing that makes the LLM hybrid governable. In the patented design the language model is never the decision-maker, so an auditor asking why a user was told to go to the emergency department gets an answer from the white-box model, not from a probability distribution over tokens. Ada's own release describes the output as fully auditable, and for health systems and insurers facing liability questions that is worth more than a few points of accuracy.
The third pillar is data protection. Ada Health GmbH is a German company and therefore subject to GDPR as a matter of course, but its enterprise documentation goes further, covering anonymous use with no personal data collected, encryption standards, hosting location for client deployments, and the handover of reports only with user consent. That matters because the data a symptom checker collects is among the most sensitive a person generates, and because the consumer app is used by people in 130 or more countries with very different legal protections.
The fourth pillar is clinical oversight. Ada's knowledge base is curated by an in-house medical team and reviewed against clinical guidelines, with localisation for different regions so that the conditions and care pathways reflect the country the user is in. The company publishes its own research, including studies that it did not win outright. The 2020 BMJ Open paper was candid that GPs outperformed every app, including Ada, on top-one accuracy, and it called for independent replication. The NEJM AI paper was published with an explicit caveat about its observational design. A company trying to spin would not write those sentences.
Put together, these four things describe a governance posture that is unusual in consumer-facing AI: regulate early, explain by construction, protect data beyond the minimum, and publish the inconvenient results. The patent is best read as the fifth pillar. It takes a governance architecture that already works and gives it legal protection at the moment the industry most needs an answer to the question of how to use LLMs in front of patients without handing them the decision.
The limits, and the open questions
An article that praises a company for its honesty about its limits owes it the same honesty in return. There are several places where the infrastructure story could stall.
The first is the evidence base itself. Ada's strongest studies are Ada-led. The 2020 BMJ Open benchmark was funded and co-authored by the company, its employees helped write the vignettes, and the paper acknowledged a possible UK bias in case selection. The ESSENCE study was designed with CUF, and its first author is an Ada clinical studies advisor. None of that invalidates the results, and both papers were peer reviewed and both disclosed the relationships clearly. But the call for independent replication that Ada itself made in 2020 has not been fully answered, and a company that wants to be infrastructure should want to be benchmarked by people who do not work for it.
The second is the limit of what a symptom checker can do. Ada is explicit that it does not diagnose, and the 71% top-one accuracy figure, good as it is relative to competitors, means that roughly three in ten assessments put the wrong condition first. Top-three and top-five accuracy are much higher, which is why the structured report matters, but a user who reads only the first line gets a tool that is wrong more often than a GP. The company's answer is that Ada is a navigation aid rather than a diagnostic, and the NEJM AI data supports that framing. The marketing does not always keep the distinction as crisp as the product does.
The third is the hybrid architecture. A 7% relative improvement in accuracy from ensembling an LLM with the PGM is real but modest, and it comes at the cost of adding a component whose behaviour is harder to characterise. Ada's design keeps the LLM out of the final decision, which addresses the worst failure mode. It does not eliminate the risk that the free-text front end misunderstands what a user said, or that the dynamically generated follow-up questions, even when constrained to the knowledge base, steer the assessment somewhere the fixed questionnaire would not have gone. The natural-language input feature is currently limited to the United States with expansion planned, which suggests Ada is rolling it out cautiously. The company will need to publish on the hybrid system specifically, not just on the underlying PGM, before partners in regulated markets can treat it as proven.
The fourth is scale and the company's own position in the market. Ada's assessment count has grown from 9 million in the early days to more than 50 million, but the general-purpose LLMs it is positioning itself against handle health queries at a volume several orders of magnitude larger. If the large model providers decide to build their own clinical layers, with their own certifications, the question becomes whether a Berlin company with a few hundred employees can defend a patent and a brand against that. Ada's bet is that the clinical infrastructure takes a decade to build and cannot be shortcut. That is probably true, but it is still a bet.
The fifth is the consumer app itself. Ada has been simplifying it, retiring legacy trackers for BMI, medication and allergies and adding an assess-for-others feature for family members. The consumer product is a Class I device with Class IIa pending, a status that has been pending for some time. The free app is also, in practice, a loss leader for the enterprise business, and loss leaders can be cut. There is no sign of that, but a company whose public identity rests on a free tool should be watched for what happens to it.
None of these questions undermines the central argument. They are the questions one asks of infrastructure, not of a consumer app, and the fact that they are the right questions is itself evidence of what Ada has become.
The clinical layer, not the chatbot, is the durable asset
The healthcare AI market has spent the past three years arguing about which model is smartest. Ada Health has spent fifteen years answering a different question: what has to be true about the system around a model before it is safe to put in front of a sick person. Its answer is a certified reasoning engine, a curated knowledge base, an explainable architecture, a quality system, a body of published evidence, and a set of integrations into the places where care actually happens. The April 2026 patent is the first time all of that has been described as a single, protectable thing, and it is telling that Ada chose to describe it as a layer.
Layers are what durable technology companies are built on. The companies that own the payments layer, the identity layer or the mapping layer are rarely the ones with the loudest consumer brand, but they are the ones everyone else has to go through. Ada's hybrid architecture is a plausible candidate for the clinical layer in healthcare AI: the component that takes whatever a language model understood a patient to say, decides what it means clinically, and does so in a way that a regulator, an insurer and a doctor can all inspect.
Whether Ada ends up owning that layer, or whether a larger company builds its own, is not yet settled. But the strategy is coherent, the evidence is accumulating, and the company has been willing to be slow and boring in a sector that rewards neither. For anyone trying to work out which healthcare AI companies will still matter in a decade, that is a better signal than any demo.
Nelson Advisors > European Healthcare Technology Investment Banking
Nelson Advisors specialise in Mergers and Acquisitions for European HealthTech, MedTech, Digital Health, Healthcare IT, Healthcare AI companies in the Lower to Mid Market ranging from $25M to $250M EV. www.nelsonadvisors.co.uk
Healthcare.Digital is the Google News approved HealthTech and MedTech Thought Leadership platform for Nelson Advisors, positioning them as a specialised authority on European Healthcare Technology M&A and strategic corporate development. https://www.healthcare.digital
Nelson Advisors publish Europe's Leading Healthcare Technology Investment Banking Newsletter every week, join 5000+ HealthTech and MedTech subscribers today! https://lnkd.in/e5hTp_xb
Nelson Advisors LLP
Hale House, 76-78 Portland Place, Marylebone, London, W1B 1NT
Meet Nelson Advisors @ Events in 2026
Digital Health Rewired > March 2026 > Birmingham, UK
NHS ConfedExpo > June 2026 > Manchester, UK
HLTH Europe > June 2026, Amsterdam, Netherlands
HIMSS AI in Healthcare > July 2026, New York, USA
Bits & Pretzels > September 2026, Munich, Germany
HealthInvestor Healthcare Summit > September 2026, London, UK
World Health Summit 2026 > October 2026, Berlin, Germany
HLTH USA 2026 > October 2026, USA
Global Health Exhibition 2026 > October 2026, Riyadh, Saudi Arabia
Web Summit 2026 > November 2026, Lisbon, Portugal
MEDICA 2026 > November 2026, Düsseldorf, Germany
Leaders in Health Summit 2026 > November 2026, London, UK
Venture Capital World Summit > December 2026, Toronto, Canada
Meet Nelson Advisors @ Events in 2027
Digital Health Rewired > March 2027 > Birmingham, UK
Barclays Health Elevate > March 2027, London, UK
NHS ConfedExpo > June 2027 > Manchester, UK
HLTH Europe > June 2027, Amsterdam, Netherlands
Comments