A tool that answers dosing questions, one that reads your ECG, and one that writes down what was said. All three are called KI. Only one of them is the one you use when documenting.
You have probably noticed it yourself: «AI in healthcare» has become a catchphrase. In the same week, you can read about a tool that suggests treatment, one that detects disease in an image, and one that prints out draft journal notes – and all of this is presented as the same technology, with the same capabilities and the same risks.
That’s a problem. Not because the language is inaccurate, but because it makes it harder for you to assess what you’re actually dealing with when a new tool appears in the clinic. The requirements are different. The responsibilities are different. And the risks are quite unevenly distributed.
The warning that made the difference clear
The Managing Director of Felleskatalogen, Bente Cecilie By Jansen, wrote an article in Dagens Medisin earlier this year that is worth reading for everyone working in clinical settings. The topic is medical «knowledge chats» – KI solutions similar to ChatGPT, but limited to medical questions and based on selected professional sources.
Her point is counterintuitive: Even when the model has access to reliable professional sources, the answers can become poorer. The reason they point to is that the structure of the sources disappears along the way. The common catalog texts are built up with tables, headings, and sections that give the content precise meaning – what dose applies to what indication, which patient group, and what age. When the text is broken down into smaller pieces before it is sent to the language model, the model loses its overview.
The results they describe from their own testing are serious: dosages that are mixed together across indications, and dose recommendations given for the wrong patient group. They have also seen that such tools perform dose calculations – and thus practically function as medical calculators, without being regulated or approved for that purpose.
Her concern is not primarily that the model is wrong. It is that the reference to credible sources creates False security, Thus, one assumes that the answer is correct because a reputable source is listed at the bottom.
Three categories that are mixed together
The article does something useful: it shows how different things we talk about when we say AI. Roughly speaking, there are three categories in clinical everyday life.
Knowledge chat. You ask a factual question and receive an answer. The tool draws on subject-specific sources and formulates a reasoning. This is the category Common Catalogue warns against in its current form – not because the technology is worthless, but because the errors are difficult to detect and the consequences can be severe.
Clinical decision support. The tool analyzes clinical data and points out something: a finding in an image, a pattern in an ECG, a risk in a set of test results. At Akershus University Hospital, a doctoral student has developed and tested a model based on data from over 200,000 patients, which reads ordinary ECGs to detect heart failure earlier. This is real clinical AI, and it is – rightly – strictly regulated.
Documentation support. The tool converts what was actually said in the consultation into a draft journal note. It does not assess anything clinically. It does not suggest any treatment, calculate any doses, or make any diagnoses. It sorts through what already exists, so you don't have to write it down twice.
The difference between the first and third categories is not a difference in degree; it is a difference in species. A knowledge chat claims something about medicine. A documentation tool reproduces something that was said. When the latter is wrong, you see it immediately in the draft you are reading, because you were present in the conversation. When the first is wrong, you have no such reference.
What the regulations actually require
It is worth being precise here, because the three categories meet different requirements.
An AI tool can be medical equipment. The Directorate of Health’s cross-departmental guidance service makes it clear that this must be assessed concretely by the manufacturer for each individual tool, and that it is crucial whether the tool is used on humans for a medical purpose. If the answer is yes, CE marking is required before the tool is put into use, and The Directorate for Medical Products has its own guidelines for software that is medical equipment. A tool that calculates dosages or suggests diagnoses is obviously in this category. A tool that writes down what was said is basically not one – but the assessment should still be made, not assumed.
For documentation support specifically, Normen, the health sector's own industry standard for information security and privacy, has published a fact sheet about safe use of AI in documentation of healthcare. The advice is specific: use tools that are designed for documentation of healthcare, not open general AI services. We have More detailed information about the fact sheet can be found here.
The guidance service has also clarified two questions many people are wondering about. The legal basis for processing the data is the documentation requirement under the Health Personnel Act §§ 39 and 40 – not consent. And the audio recording and the transcribed notes must be deleted when the journal note is quality-assured and completed; they cannot be used to train the model further without a separate processing basis. The latter is a reason to ask each provider specifically what happens to the audio afterwards. We have explained How pseudonymization works in practice in a separate article.
Where Medivox belongs
Let's be absolutely clear, since the topic invites it: Medivox is documentation support.
Medivox transcribes what is said in the consultation and creates a draft journal note in the template you have created yourself. The tool does not provide clinical advice, does not calculate doses, does not suggest treatment, and has no decision support. It is not a disclaimer – it is a description of what the product is intended to do and where the limits lie.
The liability line follows from this. Health Director Cathrine M. Lofthus expressed it like this when the fact sheet was launched: healthcare professionals are responsible for the content of the journal and must ensure that KI-generated text is correct before it is stored. This is not a disclaimer on our part. This is the entire way the tool is meant to work: you read the draft, correct what needs to be corrected, and approve. The draft is a draft until you have looked at it.
And that is precisely why the risk picture is different than for the knowledge chat. You were in the room. You know what was said. Your control is real, not theoretical.
The question you should ask
Next time someone presents an AI tool for your clinic, one question will separate the categories faster than anything else:
Does this tool say anything about medicine – or does it reproduce something that already exists?
If the supplier answers the first question, you should request CE marking, documented validation, and a description of what the tool is not intended to be used for. If they answer the second question, you should still ask about how the data is processed, what is in the data processing agreement, and what happens to the audio file – but you are faced with a different kind of risk, and you can assess it for yourself every time you read through a draft.
It’s a simpler conversation than the debate about «AI in the healthcare sector» often becomes. And it’s far more useful.
Frequently Asked Questions
Are AI tools for record-keeping allowed to be used in Norway?
Yes. The Norwegian Directorate of Health has clarified that the legal basis is the documentation requirement under the Health Personnel Act §§ 39 and 40. The requirements for privacy, information security, and quality assurance apply as usual, and the Normens factsheet on speech-to-text describes what is expected.
Do I need to obtain consent from the patient to use voice-to-text?
The Directorate of Health’s assessment is that consent for healthcare also includes the use of AI tools, in the same way as other technologies, so that no separate consent is required. They nevertheless recommend informing patients, for example with notices in the waiting room – especially during an initial phase.
Is Medivox clinical decision support?
No. Medivox documents what is said during the consultation and creates a draft of the journal note. The tool does not provide any clinical assessments, suggest any treatment, or calculate any doses.
How do I know if an AI tool is medical equipment?
This must be assessed concretely by the manufacturer, based on whether the tool is used on humans for a medical purpose. If it is medical equipment, it must be CE-marked before it is put into use. The Directorate for Medical Products has guidelines for software as medical equipment.
Does this work for other professional groups other than doctors?
Yes. The distinction between documentation support and decision support is the same regardless of profession, and the documentation obligation applies to all healthcare professionals – from dentists and physiotherapists to psychologists and nurses.
Use Medivox for free – Get started completely free
Are you wondering which category a tool you are considering belongs in? Contact us – We are happy to answer what Medivox does, and just as happily what it does not do.
Sources:
- Jansen, B. C. B., Felleskatalogen (2026): Artificial intelligence in clinical tools – when access to professional sources leads to poorer answers, Today's Medicine
- Norwegian Directorate of Health / The Norm (2026): Safe Use of AI in Health Care Documentation (Speech-to-Draft)
- The Directorate of Health (2026): Safe use of AI in documentation of healthcare – news item at the launch of the fact sheet
- The Norwegian Directorate of Health (2025): Questions and answers from the KI guidance service
- Directorate for Medical Products: Software for medical equipment
- Norwegian Directorate of Health Duty to Document – Health Personnel Act with Comments
Related posts
Share