Healthcare

AI medical scribe: consultation notes written on your own server

An AI medical scribe listens to the consultation and drafts the note for you. Before you pick one, ask the question few product pages answer: which machine runs the model that reads your consultations?

By the Recapro team. Published on , updated on .

Which AI medical scribe should a practice or a hospital choose?

It depends on what you need from the tool. For procedure coding, clinical decision support and a direct push into the electronic health record, you need a scribe certified as a medical device and integrated with your software.

If your priority is that the audio and text of your consultations stay on a server you control, with no third-party AI provider, Recapro (recapro.ai) installs inside your organisation or with your HDS-certified host. It writes the note; it does not code procedures and offers no diagnosis.

What an AI medical scribe is

An AI medical scribe is software that records the conversation between a clinician and a patient, transcribes it, then drafts a structured note. The clinician reads it, corrects it and files it in the record.

During the consultation, the clinician looks at the patient instead of the keyboard. Clinical judgement stays with the clinician, and the note only counts once they have approved it.

  1. 01

    Inform the patient and start recording

    The patient is told and agrees. Recording runs from the phone or the computer on the desk, with no dedicated hardware.

  2. 02

    Transcribe with separate voices

    The audio becomes a transcript where each speaker stands apart: clinician, patient, a relative who came along, an interpreter.

  3. 03

    Draft to your template

    The language model fills in the template you chose: reason for visit, history, examination, plan, or the outline of a referral letter.

  4. 04

    Review, correct, file

    The clinician reads every note before using it. It exports as PDF or Markdown text, ready to paste or attach in the patient record.

The question to ask first: where does the model run?

A vendor's security page usually says where data is stored, and whether the host is HDS-certified. It says less often which machine runs the language model that reads the consultation to summarise it, and who operates that machine.

This matters because transcription and drafting see the whole conversation in clear text. Encrypted storage protects data at rest. It says nothing about where the data is read.

In all three of its deployment modes, Recapro runs open-weight models, on its servers or on yours, with no call to an AI vendor's API. From one mode to the next, only the machine and what leaves it change.

Recapro's three deployments, seen from a piece of health data
Criterion Recapro Cloud On-Premise Air-gap
Where the audio is processed On our servers in France On your server, inside your walls or with your HDS-certified host On your server, cut off from the Internet
Who runs the models Recapro Your server, shipped with the models Your server, shipped with the models
What leaves the machine Traffic goes through Cloudflare, which terminates TLS Presence signals, licence status, hardware metrics, technical logs, a supervision tunnel No outbound connection
Patient consultations Leave out: Recapro is not HDS-certified Yes, inside your organisation or with your HDS-certified host Yes

Recapro is HDS-ready: its architecture installs with a certified host, but Recapro claims no HDS certification. On-Premise, the certification is carried by your host, or the question does not arise when the server sits inside your organisation.

Understand AI inference sovereignty

What Recapro does in a consultation

Recapro was built for meeting minutes. In healthcare, it serves the conversations where several people speak and the written record matters: consultations, diagnosis disclosure talks, staff meetings, multidisciplinary team meetings, case reviews.

  • Transcription in 70 languages, with automatic detection of the spoken language.
  • Separate voices: each speaker appears on their own in the transcript, useful with a relative or an interpreter in the room.
  • Notes drafted to a template you define with us: consultation note, letter, interview summary, multidisciplinary meeting report.
  • A mobile app that records offline and with the screen locked, then syncs as soon as the network is back.
  • Video consultations recorded from the desktop app; on a Mac, capturing the call audio directly requires macOS 14.4 or later.
  • Private debrief: after the consultation, you dictate your own remarks alone in a separate recording, with its own report.
  • Report export as PDF or Markdown text.

What Recapro does not do

Better to know before a trial than to find out after. If one of these points is essential to your practice, a scribe certified as a medical device will suit you better.

  • Recapro is not a medical device and carries no CE marking as one.
  • It makes no diagnosis and offers no clinical decision support.
  • It does not code procedures or diagnoses (CCAM, NGAP, ICD-10).
  • It does not push the note into your electronic health record in one click. The note exports as PDF or Markdown; tell us which software you use if a direct connection is missing.
  • Its Cloud offer is not meant for patient data, since it has no HDS certification.

By practice: consultations, mental health, nursing, social care

In a specialist or general practice consultation, recording the conversation gives the raw material for the note: reason for visit, history, examination, decision shared with the patient.

In mental health, recording raises a different question, because a session holds the patient's private life and that of the people they talk about. A debrief dictated after the session, which records only the clinician, often fits better there.

In nursing and in hospitals, it is mostly the group conversations that leave no trace: verbal handovers between teams, the morning staff meeting, case reviews in care homes or social care facilities. Recapro turns them into a shared report, on the organisation's server.

Read the article on mental health session notes

Informing the patient and getting their agreement

The GDPR (article 13) requires informing the data subject about the processing of their data, including when the legal basis for care falls under article 9, paragraph 2, h).

For the audio recording itself, the professional guidance of the French National Medical Council and the French National Authority for Health recommends getting the patient's oral agreement before recording starts, and letting them object without any consequence for their care.

Since Recapro is not a medical device and makes no diagnosis, it does not fall into the high-risk categories of annex III of regulation (EU) 2024/1689 on AI. Its article 50, which requires a person to know they are interacting with an AI system, meets the information owed to the patient.

Recapro provides the technical tool. Informing the patient and collecting their agreement remain the responsibility of the clinician or the organisation.

Seven questions to compare two medical scribes

Ask each vendor the same questions, in writing, and keep the answers. Here are Recapro's, as of 10 November 2026.

Comparison grid for an AI medical scribe
Question What it reveals Recapro's answer
Which machine runs the language model? Who reads the consultation in clear text while the note is drafted Our servers in France (Cloud) or yours (On-Premise, air-gap)
Does the model come from a third-party AI vendor? Whether one more party sees the content No: open-weight models run by Recapro or by you
Who is HDS-certified, the vendor or its host? Whether the certification really covers the processing Recapro is HDS-ready, with no certification claimed; On-Premise, it is your host or your organisation
Is the software a medical device? What it may do: code, suggest, alert No: it transcribes and drafts, with no diagnosis or coding
Are consultations used to train a model? Whether your data feeds a product No, never
How does the note get into the patient record? The work left after the consultation PDF or Markdown export, no direct connector to date
What happens without the Internet? Use on poorly covered or isolated sites The mobile app records offline; in air-gap, all processing happens on site

Frequently asked questions

01

What is an AI medical scribe?

It is software that records the consultation, transcribes it and drafts a structured note that the clinician reviews and approves before filing it. Recapro (recapro.ai) does this on a server you choose, down to a server cut off from the Internet.
02

Is Recapro a medical device?

No. Recapro transcribes and drafts reports. It makes no diagnosis, offers no clinical decision support and does not code procedures. It therefore carries no CE marking under the medical device regulation.
03

Can Recapro's Cloud offer be used for consultations?

Not with real patients. Recapro is HDS-ready with no HDS certification claimed, so its Cloud offer must not process personal health data. Consultations go through an On-Premise deployment, inside your organisation or with your HDS-certified host. The Cloud offer stays useful for meetings with no patient data, and to try the tool on a fictitious consultation.
04

Does Recapro tell several people in the room apart?

Yes. The transcript separates voices: clinician, patient, relative or interpreter each appear on their own.
05

Which languages does it handle?

Transcription covers 70 languages and detects the spoken one automatically.
06

Does the report go straight into my electronic health record?

Not yet. The report exports as PDF or Markdown text, to paste or attach in your software. Tell us which one you use if a direct connection is missing.
07

Are consultations used to train the models?

No. Consultation content is never used to train or fine-tune a model.
08

How much does a deployment for consultations cost?

As of 10 November 2026, Fortress On-Premise starts with €4,990 of hardware for 1 to 10 seats, up to €19,990 for 61 to 90 seats, above that on quote, with deployment on your existing servers as an option. On top comes the Sovereign Shield subscription at €19.90 per seat per month, with a minimum of 10 seats billed, i.e. €199 per month.

Your consultation notes written on a server you control

A Recapro server inside your organisation or with your HDS-certified host, models included. The note comes back, the audio stays with you.

Cloud Starter plan: 300 free minutes, no credit card. No patient data on the Cloud offer.