Doctolib and Nvidia: the freedom to switch tools

A practice using a paid version of Doctolib Médecin can request its medical database as 45 CSV files, accompanied by attachments. Patients, consultations, prescriptions and links from documents to patients occupy separate files. That documented exit route offers a concrete starting point for a question raised by Nvidia’s planned investment: what must a clinician be able to take away to keep working somewhere else? ^[7]

On 13 October 2026, Doctolib announced a planned minority investment from Nvidia and a deeper collaboration in artificial intelligence. Reuters reported the announcement; neither the amount nor the precise ownership percentage was disclosed. The deal brings a computing-technology supplier closer to a business already connecting appointments, medical records and consultation tools. ^[1][3]

The public documents reveal two different kinds of autonomy. Doctolib wants control over its models and their operation in Europe. In the practice, clinicians need to retain usable information and working arrangements when they change supplier. Those freedoms sit at different points in the chain. This investment provides an opportunity to examine them together.

The partnership and its public commitments

In its official 13 October community post, Doctolib said the minority holding would carry neither governance rights nor a board seat. It also stated that partners and investors have no access to health data. These are public company commitments. Neither the shareholder agreement nor the technical contracts was obtained for this investigation. ^[2]

The industrial announcement concerns developing and training in-house models, building on open models with Nvidia’s technical support. According to Doctolib, the collaboration has already lasted around a year. The company plans to expand administrative and medical assistants, while describing clinical decision support as an area for development. That timetable separates documented note-taking functions from capabilities announced for the future. ^[1][2]

For a practice, the significance will depend on what is delivered. Better handling of medical vocabulary could reduce corrections. A suggestion incorporated into the record could avoid re-entering information. At each stage, a service becomes more useful when it understands context and fits the clinician’s routine. That convenience deserves to be assessed alongside the work needed to preserve an alternative.

The documents needed to measure any contractual dependence on Nvidia remain unavailable: purchasing commitments, exclusivity provisions and conditions for changing infrastructure. Without them, that question stays open. Inferring investor access to consultations merely from a place on the shareholder register would exceed the available evidence.

A conversation passes through several states

The integrated assistant’s guide describes a specific process. The software combines a transcript, additional information entered by the clinician and the existing record to generate suggestions. The clinician must review the entire summary: Doctolib warns that the output can introduce or omit information. Once approved, information feeds the consultation note and selected record sections. ^[4]

This transition is consequential. A provisional sentence becomes something a future consultation may rely on. Human review is the bridge between proposed text and the practice’s memory. Its effectiveness also depends on available time, visible corrections and the ability to establish where a statement came from. These are operational criteria to investigate in actual use.

The stated retention periods reflect these different statuses. Doctolib says audio is no longer retained after the consultation. Its workflow guide specifies deletion of unapproved suggestions 48 hours after creation, and deletion of the transcript within 48 hours after approval. Information added to the medical record follows the record’s separate retention policy. These are documented commitments; no deletion test was performed for this article. ^[4][5]

The guide also warns that leaving the window during continuous recording, or an unexpected closure before note generation, irreversibly loses the transcript and information not yet summarised. This is a documented limitation; no loss test was performed for this article. ^[4]

Method and scope

Integrated workflow described by Doctolib, read on 15 October 2026. Unapproved suggestions: deletion 48 hours after creation. Transcript: deletion within 48 hours after approval. The privacy guide gives a general 48-hour post-consultation summary; the usage guide specifies starting events. No combined 96-hour period is inferred. Approved records follow a separate policy. Qualitative diagram, without a deletion audit or volume encoding.

Removing an audio recording limits the accumulation of sensitive material. It also makes the retained evidence important. Once the temporary source has gone, the record’s quality rests on what was checked and preserved. Investigating a later error would require knowing which versions, corrections and logs remain available. The material reviewed does not provide a complete reconstruction of that history for every circumstance.

Doctolib also requires advance notice to the patient and says the patient can object to the assistant’s use at any time. A practice therefore needs a way to conduct the consultation without it. That is a practical continuity test too: can the team carry on when the tool is declined, suspended or unavailable? ^[5]

Measuring the work left behind

Benefits deserve the same scrutiny as risks. Doctolib describes a medical-expertise framework tailored to intended uses and their possible effects on care. This sets out its approach; results still need to be assessed against their study design and comparator. ^[18]

A US study published on 23 February 2026 in the Journal of the American Medical Informatics Association illustrates the point. Its crossover trial randomised 160 volunteer outpatient clinicians across two sequences using two anonymised commercial assistants in autumn 2024. The adjusted difference between B and A was −3.19 minutes per day in notes, with a 95% confidence interval from −4.87 to −1.50. For time spent in the electronic health record outside scheduled hours, the interval crossed zero. The study is not an evaluation of Doctolib. ^[9]

Method and scope

Chowdhury et al., JAMIA, published 23 February 2026, Table 3. US crossover trial randomising 160 volunteer outpatient clinicians: 7 October–4 November and 14 November–12 December 2024. Two anonymised commercial products, not identified as Doctolib. Unit: minutes, adjusted B minus A difference; a negative value means less time with B. Notes per day: −3.19 [−4.87, −1.50]. EHR outside scheduled hours: −2.19 [−4.74, +0.37]. Shared −6 to +2 scale. EHR analyses included 158 clinicians for notes and 157 for outside-hours time, with some period-specific data missing. Short, open-label trial, without multiple-testing adjustment. Outside-hours time measures EHR activity, not all work. No total work reduction or clinical improvement is calculated.

The choice of outcome changes the conclusion available to an evaluator. Less time in the note editor can improve a working day without reducing total work by the same amount. Reviewing, preparing and completing other tasks belong in the assessment. A favourable average also leaves open how benefits are distributed across specialties and kinds of consultation.

For a purchaser, the useful test follows a complete activity: create a usable note, correct it, incorporate it and recover the information needed at the next appointment. Stopping the clock when the first text appears measures one particular stage. It leaves out the work that turns a suggestion into reliable information.

This also matters for competition. Comparable outcomes allow an assistant to be chosen for its performance. Different measures, populations and observation dates make that comparison much less secure. A demonstration may justify a trial; the design of that trial determines what the practice actually learns.

Patients’ habits already have economic value

Integration is advancing on a platform whose network effects have been examined in detail. In its 6 November 2025 decision, France’s competition authority describes a two-way attraction: more participating clinicians make the service more useful to patients, while that audience attracts clinicians. Paragraphs 304–325 set out Doctolib’s objections and the authority’s assessment. ^[12]

The authority sanctioned exclusivity provisions applied until 30 September 2023, and the tying of teleconsultation to the appointment-booking subscription. It also sanctioned the 2018 acquisition of MonDocteur. These findings concern defined markets and historical conduct. The official case page states, as read on 15 October 2026, that an appeal is pending. ^[11][13]

Doctolib disputes the decision and argues that integration supports patient follow-up and daily practice. That defence deserves serious consideration: linking the record, prescriptions and billing can provide genuine utility. Competition analysis needs to examine access conditions and the ability to select components while accounting for those benefits. ^[14]

Method and scope

Qualitative feedback loop based on the authority’s reasoning, notably §§304–325 of decision 25-D-06, 6 November 2025. The official case page lists an appeal as pending on 15 October 2026; Doctolib argues for integration benefits. Workflow dependence in B is an analytical hypothesis, with no effect estimate or legal finding about the current AI offering. Arrows have equal width and encode neither amounts nor probabilities.

AI may add a different reason to stay, within the practice itself. A team with familiar note templates, shortcuts and connected tasks has good reasons to retain its software. A competitor offering a better standalone result may also need to rebuild its links to the rest of the job. This is an economic hypothesis about integration; its magnitude has not been measured here.

The mechanisms remain distinct. A network effect comes from the other users on a platform. Dependence on a workflow comes from what the team has organised around the software. A file export may address part of the latter. Patients’ booking habits travel much less readily inside an archive.

Our investigation into how workflows become dependent on business applications follows the same problem beyond medical records.

The assistant can also sit alongside the record

Doctolib’s own documentation provides a material counterweight to the idea that the whole practice must use one suite. A standalone assistant allows all or part of the summary to be copied into another medical software product. Access requires activation or a subscription, but the documented workflow leaves the primary record in a different environment. ^[6]

The same guide says standalone summaries remain available for 48 hours, only on the computer where they were created. It instructs users to copy them into their medical software. Transfer is therefore part of the job to complete: moving to another workstation before doing so can remove access to that temporary copy. ^[6]

The integrated assistant follows another path: after review, its suggestions can feed sections within Doctolib Médecin. The difference lies in the connection between the result and the destination record. In one route the clinician transfers the content; in the other, integration handles part of that work. ^[4][6]

Method and scope

Two workflows described in the company’s public guides, read on 15 October 2026. Standalone use requires activation or subscription; its guide documents copying all or part of a summary. Integration also uses the existing record for context. The diagram assumes neither identical models nor equivalent performance, and depicts no universal automatic connector. No practice test was performed.

These routes enable a more honest comparison of choices. Manual copying keeps a visible boundary between assistant and record, while requiring an action and a check in the right place. Integration can remove part of that step and connect structured information. Its value should be assessed over time, together with the consequences of changing a component.

Modularity becomes especially useful when tools improve at different speeds. A clinician may want a new speech-recognition system while keeping the record, or a different record while retaining a preferred assistant. The standalone documentation establishes a genuine option in one direction. Automated connections, retained functionality and the commercial terms of each combination require separate examination.

This investigation does not establish how many practices use that option or how often they switch suppliers. Its existence nevertheless qualifies a blanket diagnosis of closure. A useful assessment of choice starts with the routes actually offered and then tests their limits.

Forty-five files, followed by relationships to rebuild

Data portability means being able to take reusable information elsewhere. For the paid versions covered by its help page, Doctolib describes a free export on request, following authority checks. The 45 CSV files and attachments are prepared for retrieval through the desktop application. The download is one-time; a fresh request is required after the retrieval window expires. The stated two-month window starts when the export is prepared and concerns retrieval. It is neither a migration duration nor a general period running from contract termination. ^[7]

An individual-record export instead provides a PDF summary and associated documents. That format supports reading a case. Migrating a whole database has a different purpose: enabling the destination software to recover the information and its relationships. ^[8][15]

Consider an entirely fictional example. Patient P7 has consultation C12 and document D4. Finding all three objects is a first check. The software must then know that the document belongs to the right patient and context. Those identifiers illustrate a relationship; they do not reproduce columns from an actual Doctolib export.

Method and scope

Doctolib’s stated format: 45 CSV files and an attachment folder. Small rectangles represent those files, without assigning size or patient counts to each one. Relationships are illustrative and do not reproduce an export column dictionary. The maximum two-month period concerns retrieval after export preparation; the axis is schematic. The examples establish neither success nor failure of any particular migration.

This is a question of semantic interoperability: two systems understanding the same information. A numerical observation needs its unit, date and context. France’s digital-health agency uses the CI-SIS framework to organise technical and semantic exchange requirements. The framework helps define what must remain understandable when information moves. ^[15]

Verification therefore needs to reach a working action: opening the correct record, finding a document or interpreting an observation. The number of files received demonstrates delivery. Comparing source and destination data, alongside functional tests, demonstrates usability. No actual export or import into a competing system was carried out for this article.

Our investigation into preparing and testing an exit from a provider examines the resources and exercises needed to maintain that continuity.

European rules may strengthen that ability over time. The European Health Data Space regulation entered into force in March 2025; exchanges of patient summaries and electronic prescriptions or dispensations are due to apply in March 2029, followed by medical images, laboratory results and discharge reports in March 2031. That phased timetable leaves current contracts, formats and tests operationally important. ^[16]

A privacy control always has a scope

The investor’s identity cannot describe every possible use of information. France’s data-protection authority, the CNIL, explains that a processing purpose determines necessary data, notice and retention. Reusing material to train a system requires examination of the conditions specific to that further use. ^[19][20]

One documentation detail makes this tangible. Doctolib’s 14 October community reply links to a privacy centre. The linked help page describes appointment, usage and connection information, with a reuse control managed by the administrator. Its stated scope cannot by itself establish a universal switch for every clinical, research or training activity. ^[2][17]

The usage guide describes a specific feedback channel: the rating and comment a clinician submits about a summary’s quality. Doctolib says this feedback contributes to product improvement and AI learning. The documents reviewed do not fully specify the information accompanying that feedback or its exact processing route. ^[4]

Doctolib also publicly distinguishes scientific research from product development. Examining a particular reuse would require the project description, the information given to people, recipients and applicable conditions. This article establishes neither a transfer of medical records to Nvidia nor comprehensive compliance by every processing operation. ^[18][19][20]

That precision also helps the reader understand a choice. Having software assist with a note during care, permitting service improvement and participating in research are different questions. Clinicians and patients need to understand the relevant choice when it matters, and how to exercise the rights attached to it.

The vendor’s autonomy and the clinician’s

Doctolib presents open models as a way to maintain independence. Model weights are the numerical parameters that determine much of a model’s behaviour. Being able to run a copy under its licence reduces dependence on another vendor’s remotely operated service. The announcements reviewed do not establish that clinicians would receive Doctolib’s own models to operate themselves. ^[1][2]

Our analysis of model licences and competition between providers follows the shift in bargaining power between model developers, hosts and applications.

Changing a computing supplier poses another test: running the model elsewhere with the required resources and service level. A licence permitting use and a functioning operational migration address different questions. No complete infrastructure bill of materials or report of such a transfer was obtained.

Method and scope

Analytical diagram based on Doctolib announcements and ANS frameworks. Open models are a stated strategy; no complete agreement, transferred in-house model or infrastructure migration was obtained. Export is a documented function; no complete practice changeover is verified here. HDS addresses a hosting scope, distinct from clinical-performance and reversibility evidence.

Hosting requires similar precision. For this assistant, Doctolib says data is kept in the European Union in an HDS hosting environment. France’s digital-health agency describes HDS as an assessment of hosting and its security. Its scope differs from a clinical-performance measurement or a reversibility test. Each kind of evidence needs to retain its proper function. ^[5][10]

For clinicians, autonomy is tested much closer to the patient. Can they use another assistant, keep intelligible access to records and continue consultations through a change? For Doctolib, it is tested through model, infrastructure and contractual choices. An improvement at one level can coexist with a remaining dependency at the other.

The planned partnership may improve practice tools and strengthen the ability to develop them in Europe. The public documents also offer tangible ways to assess openness: a standalone assistant, a structured export and an interoperability framework. The next useful evidence would be a tested changeover, from a validated consultation to work resumed in the destination. Freedom to choose is measured by the service that can still be delivered.

Sources and limits

Documentary research as of 15 October 2026. Investment announcements and usage/export guides are read alongside public security and interoperability frameworks, a competition decision and a published clinical trial. Doctolib’s described features remain documentary statements. No interviews, patient-record access, practice trial, deletion audit, investment agreement or actual migration was obtained or performed.

Doctolib corporate pages that did not return directly usable text, the JAMIA study and the two CNIL guidance pages were checked through indexed primary-source passages. The Reuters report was read in its Euronext republication. Relevant passages of the public competition decision were read, with visual checks of printed pages 2, 85 and 89. The official case page lists an appeal as pending. Cited rules and historical findings are not a legal determination concerning the new AI offering.

The US trial observations date from 2024 and publication from February 2026. Its products are anonymised and are not attributed to Doctolib. Figures distinguish published data, documented commitments and analytical mechanisms. Record identifiers are entirely fictional; no real personal data is reproduced.

1. Doctolib · Doctolib va accélérer dans l’IA en santé grâce à un investissement de NVIDIA. 2026-10-13. Planned minority investment; technical collaboration and in-house models. Company disclosure. Official indexed passages read; direct page returned no usable text. Investment amount, ownership percentage and agreement unavailable.
2. Doctolib Communauté · NVIDIA rejoint Doctolib pour accélérer l’IA en santé, en Europe. 2026-10-13. Official 13 October post and 14 October reply: stated absence of governance rights or board seat, European data hosting and optional AI modules. Company statements, without independent audit or shareholder agreement. Private users’ comments are not treated as evidence.
3. Reuters · Nvidia buys stakes in French tech startups Doctolib, Genesis AI. 2026-10-13. Reuters report republished by Euronext, updated 13 October at 20:56; size and amount undisclosed. Reporting relies on company statements; no closing instrument obtained.
4. Doctolib / aide · Comment utiliser l’Assistant de consultation avec Doctolib Médecin ?. changing webpage. Generation from conversation, additional information and the record; full review; writing after validation; event-specific 48-hour deletion rules. Changing documentation read on 15 October. Described behaviour, without clinical session or deletion test.
5. Doctolib / aide · Confidentialité et sécurité des données de l’Assistant de consultation. changing webpage. Audio not retained after the consultation; temporary transcript and suggestions; separate retention for validated records; patient notice and objection; stated EU/HDS hosting. The page gives a general 48-hour summary; the workflow guide specifies starting events. No combined retention period is inferred.
6. Doctolib / aide · Comment utiliser l’Assistant de consultation autonome ?. changing webpage. Summary can be copied in full or in part into another medical software product; access requires trial activation or subscription. A documented modular option; no universal automatic connector or tested migration established.
7. Doctolib / aide · Exporter ma base patients Doctolib Médecin. changing webpage. Paid versions: free export on request of 45 CSV files and attachments; authority checks; one-time retrieval within two months after preparation. Medical-database export description. No actual export, full column dictionary, firm preparation time or destination import obtained.
8. Doctolib / aide · Exporter un dossier patient dans Doctolib Médecin. changing webpage. Individual export: PDF summary and associated documents, distinct from the full-database export. Documented route; PDF readability and structured re-import are separate criteria.
9. Chowdhury et al. / JAMIA · Comparing ambient scribes: a randomized crossover clinical trial addressing ambient scribe technologies’ impact on physician burnout. 2026-02-23. DOI 10.1093/jamia/ocag018. Crossover trial, 160 participants; 7 October–4 November and 14 November–12 December 2024; Table 3, B minus A differences and 95% confidence intervals. Two anonymised commercial products in the United States; no identification with Doctolib. Available samples differ by metric; short, open-label trial. Out-of-hours time concerns the EHR only; no multiple-testing adjustment. Indexed primary-source passages read; no full text downloaded.
10. Agence du numérique en santé · Certification HDS : tout savoir. changing webpage. Scope of health-data hosting certification, independent conformity assessment and covered activities. Hosting certification does not measure the assistant’s clinical quality. No individual certificate was audited for this investigation.
11. Autorité de la concurrence · The Autorité fines Doctolib €4,665,000 for abusing its dominant position. 2025-11-06. Summary of exclusivity, tying and the MonDocteur acquisition; booking and teleconsultation markets. Sanctioned historical conduct, distinct from the current AI offering. Appeal status is documented by source 13.
12. Autorité de la concurrence · Décision 25-D-06, version publique. 2025-11-06. 195-page public decision. Summary pp. 2–3; network effects §§78–89 and 304–325, notably pp.85 and 89; company arguments and the authority’s assessment. Relevant passages read, screenshots of printed pages 2 and 85 checked. Public PDF version dated 26 November 2025 on its landing page. Decision under appeal.
13. Autorité de la concurrence · Décision 25-D-06 : suivi du recours. changing webpage. Official decision page states that an appeal is pending. Status read on 15 October 2026; no final appellate outcome inferred.
14. Doctolib · Doctolib will appeal the decision of the French Competition Authority. 2025-11-06. Challenge to the market definition and decision; defence of integration benefits for follow-up and practice operations. Position of the sanctioned party; official indexed passages read.
15. Agence du numérique en santé · CI-SIS : Cadre d’interopérabilité des systèmes d’information de santé. changing webpage. Technical and semantic interoperability; content, service and transport layers; conformity testing. Reference framework, without certifying Doctolib’s complete conformity or any particular import.
16. Commission européenne · European Health Data Space Regulation. changing webpage. Entry into force on 26 March 2025; 2029 and 2031 implementation milestones for priority exchange categories. Phased timetable. No complete practice migration guaranteed as of 15 October 2026.
17. Doctolib / aide · Désactiver la réutilisation des données par Doctolib. changing webpage. The linked privacy-centre guide lists appointment, usage and connection information; control by the administrator. The described scope alone does not establish a universal control over every clinical or research processing purpose.
18. Doctolib · Expertise médicale. changing webpage. Stated use-specific medical assessment; declared distinction between scientific research and product development. Company framework and statements, not independent clinical-trial results. Official indexed passages read.
19. CNIL · IA : définir une finalité. 2024-04-08. Explicit processing purpose; notice, minimisation and retention depend on that purpose. General guidance informing the analysis; no finding of compliance or sanction concerning this partnership.
20. CNIL · IA : réutilisation des données, tests et vérifications nécessaires. 2024-04-08. Reuse for another purpose, compatibility under applicable legal conditions and required checks. Research, improvement and care require distinct examination; no particular Doctolib legal basis is assumed.