Skip to content
FibonacciAestheticsLog inTry it free

Technology and AI

Dictation during the consultation: how it works, what it stores and what the doctor must re-read

Dictating the clinical record with artificial intelligence: how it works, where the audio goes, what the software suggests and what the doctor always decides.

The Fibonacci editorial teamUpdated on 12 min read

Microphone against blurred lights
In this article
  1. What dictation actually is
  2. The three steps: dictate, review, decide
  3. What it fills in and what it does not
  4. Why you choose the scale grade yourself
  5. Responsibility does not pass to the software
  6. Where the audio goes and what remains
  7. Informing the patient
  8. What the literature says
  9. The typical errors of a transcription
  10. Before using it in the practice

Dictation turns what you say into text and, where the form provides for it, splits it into the fields of the chart. The result is a draft: nothing enters the chart until you re-read it, correct it and press the save button yourself. The software does not assign a scale grade and does not make diagnoses, and responsibility for what is written remains with whoever signs the chart.

Before using it, a second point is worth making. The audio passes through an external artificial intelligence provider, so the patient must be informed. In Fibonacci, dictation is also still in testing: it works, but it is not used with real patients until the provider confirms in writing that it does not retain the content.

What dictation actually is

It is three steps in a row. The audio of your voice becomes text. That text is read by a language model that tries to place its parts into the fields of the form (allergies, medications, product used, notes). You see the result, correct it and decide what to keep.

The second step is what distinguishes dictation from a recorder. A recorder leaves you a file to listen back to. Dictation leaves you text already laid out in a clinical chart, and that is why the error is more insidious: a filled-in field looks like a verified datum, whereas it is your sentence interpreted by a model.

In Fibonacci the button appears in three places, with a different label in each because what you are dictating changes:

  • Dictate the history, in the chart summary, history tab;
  • Dictate the session, in the treatment form, next to the notes;
  • the clinical assessment, when the practice has that form active.

To use it you need an account with a medical role and clinical access to that patient, a microphone with permission granted to the browser and a connection, because the transcription does not happen in the browser but on an external service. Quality depends more on the noise in the room than on the microphone.

The three steps: dictate, review, decide

Dictate. Press the button: a red dot appears with the words «I'm listening» and the text appears as you speak. Two buttons: End closes and moves to review, Cancel throws everything away.

Review. The transcribed text sits in an editable area, below a notice worth reading once: automatic transcription gets things wrong above all on medications, dosages and technical terms. If the form provides for field extraction, next to the notice there is a reliability percentage. It is a single number for the whole extraction, not one per field, and it measures how clear the model found the text, not how correct what you said is.

Decide. Three buttons, which do different things:

ButtonWhat it does
DiscardThrows away the transcription.
Use textPuts the text as it is into the destination field, for example at the end of the session notes. Appears only where the text has a destination.
Propose for the chart (history) or Fill in the fields (treatment)Brings the recognised fields into the form, where they remain editable. Appears only if the extraction found something.

Even after applying the fields, saving is a separate gesture. Applying fills in the form; what goes into the chart is what you save.

What it fills in and what it does not

This is where expectations break down most often, so measurement is better than promise.

In treatment, dictation proposes product, quantity, batch, off-label use and the related justification. It does not fill in the device parameters (wavelength, fluence, spot, frequency, pulse duration, passes, cooling, endpoint), nor the dilution, nor the UDI or the batch expiry: those must be written by hand. They are numbers that matter a great deal in a treatment record and that a transcription can get wrong, so you write them.

Dictated areas do not become points on the map. They end up at the end of the notes in the form [dictated areas: ...], because marking an area requires its exact code; to bring them onto the map there is a dedicated button that extracts the areas from the text.

Dictation works in Italian. Even with the interface in another language, recognition and extraction are done on Italian.

Why you choose the scale grade yourself

In the forms there are two types of field, and dictation treats them differently on purpose.

Facts reported by the patient are allergies, medications, conditions, pregnancy, smoking, products used on the skin, treatments already performed. If you say «not pregnant», the field becomes no: that is transcription, and dictation fills it in.

Judgements graded on validated scales are the Fitzpatrick phototype, the Glogau classification, the «absent, mild, moderate, marked» steps of the physical examination. Here assigning the grade is a medical act: it is you who, by looking and touching, decide which step that patient is on. If the software derived the grade from your description, from «the skin gives a lot when I lift it» to «firmness: marked», it would no longer be transcribing. It would be interpreting clinical information to suggest a decision.

This is a distinction that European regulation takes seriously. The Commission's guidance on the qualification of software as a medical device (MDCG 2019-11) (opens in a new tab) states that software that processes, analyses, interprets or creates medical information may qualify as a medical device when it serves as a support to a decision or suggests an action to the user. It adds two clarifications that remove any shortcuts:

  • the risk of harm to the patient is not a qualification criterion, so «the doctor confirms it anyway» is not enough to exclude qualification;
  • a search function that uses natural language processing is not a «simple search» if it contributes to a medical purpose.

For software that provides information used for diagnostic or therapeutic decisions, the same guidance refers to rule 11 of the medical devices regulation: class IIa as a minimum, with higher classes when the decision can have serious consequences. What this means for anyone choosing a practice management system is in is clinical chart software a medical device?.

This is why Fibonacci holds a firm line: between «gives a lot» and «marked», you decide, and dictation does everything else. Every scale field ends up in one of three states.

If you have...Dictation...
stated the grade («moderate firmness»)fills it in and shows below your words: «You said: ...»
only described it («gives a lot when I lift it»)does not grade. It puts your sentence next to the field, under «For you to grade», and opens the selector empty: grading is one tap
not mentioned itleaves the field as it was

In the second case you lose nothing: the description stays in the free-text fields of the section (inspection, palpation, observed signs) and in the transcription. And you gain two things. The chart says what you said: next to every filled-in grade the sentence it comes from remains, and whoever reviews it later sees that the judgement is yours. And a plausible grade that is not yours never appears, which at the end of the day would be confirmed out of inertia, whereas an empty box would not.

Responsibility does not pass to the software

The code of medical ethics (opens in a new tab) (art. 78) states that the doctor, when using IT tools, guarantees the obtaining of consent, the protection of confidentiality, the relevance of the data collected and, as far as within their competence, the safety of the techniques. Art. 26 requires that the chart be drawn up with completeness, clarity and diligence, that corrections be justified and signed and that the traceability of its drafting be guaranteed.

The principle also applies to dictation, and Fibonacci applies it literally: the system writes nothing into the chart without an explicit action by the doctor. Every transcribed text and every proposed field requires review and an affirmative gesture. Italian Law 132/2025 says so too in art. 13: in the intellectual professions the use of artificial intelligence is limited to instrumental and support activities, with the intellectual work of the professional prevailing.

Where the audio goes and what remains

Fibonacci sends the audio to Mistral AI SAS, in Paris, through its European access point. It is the provider of the models used for transcription, for extracting fields from dictated texts and for the assistant. For extraction, the provider receives the text without the patient's name and tax code. Training of the models on the data sent is disabled.

There is one point still open, and it is the reason dictation is in testing: zero retention of content at the provider is not yet active and has not been confirmed in writing. Until it is, dictation must not be used with real patients. It can be switched off: the chart can also be filled in by hand, field by field, and works the same, with no audio trace anywhere.

If you upload a recording made elsewhere (with a recorder or with your phone), before choosing the file you must tick «the patient has consented to the recording of the visit». MP3, WAV or M4A audio are accepted, and texts already transcribed in TXT, Markdown or SRT; a DOCX or PDF cannot be read and must be exported as text. Then everything proceeds like a dictation. The transcription stays in the visit, with the file name and the date on which you declared consent; the audio is not retained.

From a legal point of view, the provider is a processor, and the doctor remains the controller. Art. 28 of the GDPR (opens in a new tab) requires a written contract binding the processor to process the data only on documented instruction and to engage another processor only with authorisation, and art. 32 requires security measures appropriate to the risk. For the details of where the patient's other data is, see where patient data is stored; the artificial intelligence page explains where there is a language model inside Fibonacci and what it never does.

Informing the patient

Treating the patient does not require privacy consent: the Garante has made this clear, and processing for care purposes rests on art. 9(2)(h) and (3) of the GDPR (if you want the full picture, see you don't need privacy consent to treat). The duty to inform remains. Art. 13 of the GDPR requires you to tell the patient, when you collect the data, also the possible recipients or categories of recipient and the intention to transfer data outside the Union. An artificial intelligence provider that receives the voice or the text is a recipient.

The special rule is added to the general one. Italian Law 132/2025 grants the patient the right to be informed about the use of AI technologies (art. 7) and requires the professional to communicate to the recipient of the service, in clear, simple and exhaustive language, information about the AI systems they use (art. 13).

In practice, two lines in the privacy notice you already hand out are enough: that you give your voice to a transcription tool, which provider processes it and where, and that the text is always re-read by you. If you record the whole conversation and not just your notes, tell the patient before you start: it is the reason Fibonacci asks you to declare consent to recording before uploading a file, and it is a good habit outside the software too.

What the literature says

A 2026 review on AI in non-surgical aesthetic procedures (PMC13314821 (opens in a new tab)) describes applications ranging from initial assessment to post-treatment monitoring, and concludes that the tools should be conceived as structured aids to assessment and communication, not as substitutes for clinical judgement. It notes that the evidence is fragmented: small, heterogeneous studies, with little external validation, and that integration into the workflows of a full day requires time, training and infrastructure.

Among the limitations it lists the inaccuracies of language models, the question of privacy and consent when the system processes identifiable data, unrepresentative training data and the need for human supervision. It is a prudent framework and coincides with what Italian law requires: the tool proposes, the doctor decides.

The typical errors of a transcription

The product notice says it: automatic transcription gets things wrong above all on medications, dosages and technical terms, that is, where an error costs the most.

  • Medications and dosages. Similar names, units of measurement, «five» and «fifteen» that sound alike.
  • Batches and codes. A series of digits and letters dictated aloud is fragile in any system.
  • Technical terms for a device or a technique, especially if uncommon.
  • Several voices at once. If the patient speaks while you dictate, the transcription gets worse.
  • Voice commands. «Full stop» and «new line» are not needed: the model adds the punctuation.

A few habits help. Speak at a natural pace, without enunciating: the model is calibrated on spontaneous spoken Italian and slowing down makes the result worse. Say medications in full, active ingredient and dose («pantoprazole forty milligrams, one tablet in the morning»). Dictate one item at a time. And always re-read the numbers, because they are the ones most often got wrong and the ones that matter most.

Before using it in the practice

A list to keep to hand:

  • I know where the audio goes and who the provider is, and I write it in the privacy notice
  • The privacy notice also mentions the use of artificial intelligence tools, in simple words
  • The provider is appointed as a processor with a written contract
  • I re-read every transcription before using it and check medications, dosages and batches by hand
  • I choose the scale grades myself: if the software proposes them in my place, I report it and do not accept them
  • If I record the whole conversation, I tell the patient before starting
  • In a practice with real patients, I use dictation only when the provider has confirmed in writing how it handles content

What Fibonacci does for this

This is how it looks in Fibonacci, with sample data:

The patient record in Fibonacci, with sample data

Sources

  1. Italian Law 132 of 23 September 2025, arts. 7 and 13: artificial intelligence in healthcare and in the intellectual professions (Normattiva) (opens in a new tab)
  2. Regulation (EU) 2016/679 (GDPR), arts. 9, 13, 28 and 32 (opens in a new tab)
  3. Garante per la protezione dei dati personali, measure of 7 March 2019, no. 55, on the processing of health data in the healthcare sector [doc. web 9091942] (opens in a new tab)
  4. MDCG 2019-11 rev. 1 (June 2025): qualification and classification of software under the MDR and IVDR regulations (opens in a new tab)
  5. FNOMCeO Code of Medical Ethics (2014, with the updates to arts. 54, 56, 76 and 77), arts. 26 and 78 with the application guidelines (opens in a new tab)
  6. Artificial Intelligence in Non-Surgical Cosmetic Procedures: a multi-stakeholder review, 2026, PMC13314821 (opens in a new tab)

These articles explain what regulations and guidelines require and what is documented in the medical record. They are not legal or tax advice on your situation and do not say how to treat a patient.

Try Fibonacci in your practice for 15 days, no credit card needed

Within one working day we call you: together we set up your practice and import your data.

Try it free

Read next

All articles