
Gemini for Healthcare and France's reliance on US tech
Gemini for Healthcare targets US hospitals. In France, the Health Data Hub already sits on Microsoft Azure and Doctolib's AI uses Google and Anthropic models.
On November 10, 2026, Google Cloud launched Gemini for Healthcare, an AI agent for hospitals and health insurers. It reads the patient record, summarizes years of clinical notes and prepares prior authorization requests. The service is in preview in the United States only, ten months after OpenAI and Anthropic started selling the same thing in January.
The three big US AI labs are now competing for the hospital. France is not on their list, for now. Yet the dependence question is already there, with no help from Google. France's national health data platform, which pools the data of the public health insurance system, has been hosted by Microsoft since 2019. Doctolib, the French medical appointment platform, stores all its data with Amazon Web Services, and its AI features run on models from Microsoft, Google and Anthropic.
Gemini for Healthcare will not be running French hospitals tomorrow. But French health data already flows through US companies, and the next layer, AI agents wired into the patient record, is being decided right now, without any rules of its own.
What is Gemini for Healthcare?
Gemini for Healthcare is a Google Cloud AI agent for healthcare providers and insurers. An agent is a program that chains several steps on its own: it queries software, cross-checks what it finds and hands back finished work. This one connects to the electronic health record (the software where a hospital keeps notes, prescriptions and results), to billing systems and to the reference data of US public insurance. According to Google's announcement, it summarizes a patient's history across several records, prepares a visit, checks a member's benefits and assembles a prior authorization file, the approval a US insurer requires before some treatments.
It is reserved for Gemini Enterprise customers, with no published price and no general availability date. Google lists HCA Healthcare, Northwell Health, Optum and the Stanford School of Medicine among its first customers. Customers pick the model that does the work: Google's, but also Anthropic's and OpenAI's, with a fast, cheap model for administrative tasks and a reasoning model for clinical analysis.
Google promises that patient data is never used to train its models and that each agent has its own identity, with the permissions of the user who runs it. The product is "built to support HIPAA compliance". That is a design claim, not a certification, and Google Cloud's HIPAA page does not list it by name yet. None of the launch texts mentions Europe, GDPR or hosting outside the United States.
An agent that reads the record without writing to it
The most interesting detail is a limitation. The connectors to Epic and Oracle Health, the two big US electronic health record vendors, are read-only: the agent never changes the record. Brendan Keeler, who writes about health APIs, took the Epic connector apart: every request queries the record live, nothing is copied to Google, and clinical notes are cut off at around 20,000 characters.
For the research firm Futurum, read-only access caps what the agent can finish: Gemini assembles the packet, a human or another system files it. We see it rather as the right way to bring an agent into a critical system. The agent prepares, the human signs off, and the record vendor keeps control of what gets written. That rule deserves to survive the move to production, in France and elsewhere.
Why Google, OpenAI and Anthropic all want the hospital
Google is the last to arrive. OpenAI launched OpenAI for Healthcare on January 8, 2026, naming Cedars-Sinai, Boston Children's Hospital and HCA among its customers. Anthropic followed on January 11 with Claude for Healthcare. The three offerings look alike: access to the patient record, Medicare coverage data, the ICD-10 disease classification, the US provider registry, and ready-made tasks such as prior authorization. What these labs sell to hospitals is first and foremost connectors.
The sales pitch is administrative time. Google says US clinicians lose nearly 28 hours a week to paperwork, a figure from a survey Google Cloud commissioned from The Harris Poll in 2024, based on hours reported by respondents. Read it as a sales argument. The need, though, is real, and so is the market: OpenAI says more than 230 million people ask ChatGPT health questions every week.
The labs are not alone. Epic, the health record vendor, is building its own agents and claims more than 85% of its customers use its AI in some form. In October it introduced connectors that let its agents read and act inside other vendors' applications. The fight is over whose agent sits at the center of clinicians' work.
In France, only Microsoft is already in the hospital
None of these offerings targets France. ChatGPT Health, OpenAI's consumer version that connects to medical records, excludes the European Economic Area, Switzerland and the United Kingdom. Claude's features that read personal health data are reserved for US subscribers.
The exception is Microsoft. Its Dragon Copilot assistant, which listens to the consultation and drafts the note, officially launched in France in October 2025, after a trial at CHRU de Nancy, Hôpital Foch and the American Hospital of Paris: more than 220 doctors, 5,700 consultations. The announcement says nothing about where it is hosted.
Does French health data already flow through US companies?
Yes, through two doors: the public platform that pools national health insurance data, and Doctolib, where people book their doctor's appointments.
The Health Data Hub, seven years on Microsoft
France's national health data platform, known as the Health Data Hub, makes public health insurance data available to researchers. It has been hosted on Microsoft Azure since it was created in 2019, a choice made without a call for tenders. Legal challenges followed one another. The latest was rejected on March 20, 2026 by the Council of State, which nonetheless wrote that the risk of US authorities asking Microsoft for access to health data "cannot be entirely ruled out". It found the safeguards sufficient: pseudonymized data, limited retention, servers in France.
Microsoft does not deny it. On June 10, 2025, before a French Senate committee of inquiry, its head of public and legal affairs in France was asked under oath whether he could guarantee that French citizens' data would never be handed over on an order from the US government. His answer: "No, I cannot guarantee it, but, again, it has never happened."
The government finally made a call. On April 23, 2026, Scaleway was chosen to host the platform, with a standalone copy of the main database targeted between late 2026 and early 2027. The contract requires the SecNumCloud qualification, the label from ANSSI, France's cybersecurity agency, that shields a provider from foreign law. Scaleway is still going through qualification. As far as we know, no public announcement says the migration is complete. Seven years after choosing Microsoft, France has only started to leave.
Doctolib: hosted on AWS, with AI from Microsoft, Google and Anthropic
Doctolib claims more than 50 million patients in France. Its privacy policy, in its July 2026 version, is explicit. Doctolib stores with Amazon Web Services "all Personal Data, whether or not it qualifies as Health Data" (our translation). The servers are in Europe, in Frankfurt and Paris. The same document states that AWS "does not hold the SecNumCloud 3.2 qualification", then adds:
"AWS's parent company is established in the United States. In specific cases, AWS may therefore be required to comply with the law or with a valid and binding order from a government authority (USA Patriot Act, FISA Section 702, Cloud Act, EO 12333), in which case personal data may be accessed from outside the EEA, in particular from the United States." (our translation)
That sentence comes from Doctolib itself. The risk of US access is spelled out in its own terms. The company answers with encryption: the master key is held in France by Evidian, an Atos subsidiary, so a US authority would only obtain unusable encrypted files, according to Doctolib. Two caveats. Encryption is server-side: the application decrypts the data to process it, on machines rented from AWS. And in 2022, after an investigation by Radio France, Doctolib acknowledged that appointment data was not end-to-end encrypted.
AI adds a second dependence. The same policy lists Microsoft Azure, Google and Anthropic as processors that "provide the AI model" for its health assistant. The data processing agreement signed with doctors names Anthropic and Google's Gemini for "providing the LLM model", and places Gemini's servers in both the European Union and the United States. In practice, when a doctor uses the consultation assistant to transcribe and summarize an appointment, part of the work is done by US models, as the French outlet Next found: some notes "do pass through Anthropic's and Google's servers for inference".
Then there is what is disputed. In June 2026, the weekly Le Canard enchaîné ran a story saying Doctolib "hands its users' information to US AI giants" and that their models were training on consultation notes. Doctolib denies any training: these providers act on its instructions, and their contracts forbid it. Nothing public proves otherwise. As for the research program announced this summer, which will use pseudonymized patient data by default to train in-house models, both of its studies are still listed as "Not started" on Doctolib's portal.
The only court to have ruled, as far as we know, did so on a narrow question. In March 2021, the Council of State refused, in summary proceedings, to suspend the vaccination partnership between the French state and Doctolib: the case involved appointment data with no medical reason attached, deleted within three months and encrypted by a trusted French third party. The underlying question has never been settled. Doctolib remains a French company, but Nvidia took a stake on October 13, 2026, alongside US funds already on board, as detailed in our article on who owns France's tech champions.
What HDS certification protects, and what it does not
In France, hosting health data on behalf of healthcare providers requires HDS certification (for "hébergeur de données de santé", health data host). Microsoft, Google and AWS all have it. Since version 2 of the standard, in force since 2024, data must be stored in the European Economic Area and customers must be told about possible access from outside it. A decree of March 24, 2026 also requires listing in the contract the foreign laws that could force access. The passage from Doctolib's policy quoted above is that requirement at work.
HDS makes the risk visible without protecting against it. The CLOUD Act requires a US company to hand over data it controls, wherever it is stored. Section 702 of FISA, the US foreign intelligence law, lapsed in June 2026 after Congress failed to renew it, but collection continues under existing certifications until March 2027. For its part, AWS says it has never disclosed enterprise content stored outside the United States since 2020. The risk remains theoretical. It is not zero.
The only legal barrier is the SecNumCloud qualification, which excludes providers subject to non-European law. Article 31 of France's SREN law has required it since 2024 for sensitive data held by the state, its agencies and the national health data platform. Its implementing decree, due within six months, only came out in April 2026. It adds the national digital health agency, but hospitals, clinics, Doctolib and healthcare software vendors remain out of scope. France's Court of Audit estimates that SecNumCloud costs 25 to 40% more, an extra cost it considers acceptable.
Will US companies run French healthcare tomorrow?
Not through Gemini for Healthcare as such: it is cut for the US healthcare system, its insurers and its forms. But it shows where tomorrow's dependence will form. We see three layers.
The first is hosting. It is the most visible one, the one the law is starting to regulate, and the one that can be moved, slowly, as the Health Data Hub shows. The second is models. Doctolib and Dragon Copilot show that this layer is already American in the tools clinicians use every day. Swapping models is still easy when the architecture allows it: an LLM gateway lets you move from Gemini to a European model by changing one line of configuration.
The third layer is the one coming next: agents that carry the procedures. An agent that knows how to assemble a coverage file, code a hospital stay or prepare a visit encodes how a department works. Its connectors, rules and access rights become the work tool. A model can be replaced in an afternoon. An agent that carries a hospital's procedures stays for years. That is what Epic and the labs are fighting over in the United States. And no French rule says today who may operate an agent wired into a patient record, or from which country.
France starts with some advantages. Epic is not installed there, and hospital health records mostly come from European vendors such as Dedalus, Softway Medical or Maincare. Mon espace santé, the national patient health space, was awarded to a consortium led by Atos until 2028. SecNumCloud-qualified offerings exist, at OVHcloud or at S3NS, the Thales and Google joint venture, which incidentally shows that US technology can be operated under a European legal regime. But the balance of power is well known: Amazon, Microsoft and Google hold 70% of the European cloud market, and European providers' share has halved since 2017. Even Nabla, a French medical AI start-up, hosts its European customers on Google Cloud.
There are counterarguments. Microsoft says no such request has ever happened, and AWS says it has never disclosed enterprise content stored outside the United States. Europe's supply of healthcare agents is still thin, and a hospital will hesitate to pay 25 to 40% more for qualified hosting. Finally, refusing US tools without a serious alternative would deprive clinicians of real time savings. We do not think the answer is a ban. It lies in the architecture, and in setting the rules before the agents arrive, not after. Our page on sovereign AI details the criteria we apply to our own providers.
What a hospital, a vendor or a healthcare CIO can take away
For a healthcare provider or a medical software vendor, a few principles emerge from what is happening in the United States and from the French cases.
- A map of the three layers. For every tool that touches patient data, the point is to know who hosts it, who provides the model, and which parent company could be compelled by foreign law. Since the March 2026 decree, the host must list those laws in the contract: the information is already there.
- Read-only as the starting rule. An agent that prepares and a clinician who signs off, with a separate identity for each agent and the permissions of the user who runs it: Google did it, and it is a good model whatever the provider. Our principles for enterprise AI agents point the same way.
- An agent kept separate from the model. Procedures, connectors and business rules are best kept with the provider or with a European vendor, using standard connections such as the MCP protocol. The model then becomes a component that can be swapped, including for an open-weight model hosted in Europe.
- AI kept for what is not deterministic. Checking a patient's eligibility or applying a coding rule is often a job for a classic algorithm, more reliable and sending nothing to anyone. We explain why in our article on lean automation.
- A requirement that can come before the obligation. Hospitals are not subject to SecNumCloud, but nothing stops them from requiring it for their most sensitive data. The extra cost is easier to negotiate before a deployment than during a forced migration, as the Health Data Hub is finding out.
Frequently asked questions about Gemini for Healthcare and health data dependence
What is Gemini for Healthcare?
Gemini for Healthcare is an AI agent launched by Google Cloud on November 10, 2026 for US hospitals and health insurers. It connects read-only to Epic and Oracle Health patient records, to billing systems and to Medicare reference data, to summarize a patient's history, prepare a visit or assemble a prior authorization request. It works with models from Google, Anthropic or OpenAI, at the customer's choice.
Is Gemini for Healthcare available in Europe?
No. Google launched it in preview in the United States only, with no timeline for Europe, and its connectors target US software and reference data. The equivalent offerings from OpenAI and Anthropic do not target Europe either. The only US clinical assistant deployed in French hospitals is Microsoft's Dragon Copilot, available in France since October 2025.
Does Doctolib use US AI models?
Yes. According to its July 2026 privacy policy and its data processing agreement with doctors, Doctolib uses models provided by Microsoft Azure, Google (Gemini) and Anthropic for its AI features, including the consultation assistant. Doctolib says these providers act on its instructions and that their contracts forbid them from training their own models on this data.
Is Doctolib's health data stored in the United States?
No, it is stored in Europe, with Amazon Web Services in Frankfurt and Paris, and encrypted with a key held in France by Evidian (Atos group). But Doctolib states itself that because AWS's parent company is American, AWS may be required to comply with an order under the CLOUD Act or FISA, in which case data may be accessed from the United States.
Does HDS certification protect against the CLOUD Act?
No. HDS certification requires health data to be stored in the European Economic Area and customers to be informed of possible access from abroad. It does not shield a US host from the CLOUD Act. Only ANSSI's SecNumCloud qualification excludes providers subject to non-European law, and it is only mandatory for the French state, its agencies and a few public bodies.
Where does the Health Data Hub migration stand?
On April 23, 2026, the French government chose Scaleway to replace Microsoft Azure as the host of the national health data platform, with a standalone copy of the main database targeted between late 2026 and early 2027. The contract requires the SecNumCloud qualification, which Scaleway has not obtained yet. No public announcement says the migration is complete.
Gemini for Healthcare will not land in French hospitals tomorrow morning. It does show what is coming: agents that read the patient record, sold by three US companies. France took seven years to decide to move the Health Data Hub off Microsoft, and has not finished doing so. The agents that will read patient records will arrive faster than that.
For hospitals, software vendors and the state, what remains is to decide, before they arrive, which procedures stay with us, and on which infrastructure.


