South-Eastern Norway Regional Health Authority is introducing speech-to-draft into clinical use these weeks – and is stating outright that the results so far vary. That is not a bad sign. It is exactly what research says we should expect.
If you have been following AI in the Norwegian health service over the past six months, you have seen two narratives that do not quite add up. One is that documentation support frees up time and provides more patient contact. The other is that clinicians who have tried it often are left with a feeling that it was faster before.
Both can be true at the same time. And we now have research that explains why.
Norway's largest health region is now implementing – and is honest about the numbers
Sykehuspartner is in the implementation phase of the regional project «AI for medical records – speech to draft» in the South-Eastern Norway Regional Health Authority. The contracts were signed in May, the establishment is taking place in two phases, and the first clinical resources will implement the solutions in late summer 2026 – well now.
The interesting thing is not that they are introducing it. It is how they describe the status. On the project page, South-Eastern Norway Regional Health Authority lists the expected benefits—less administrative burden, better quality in the notes, more time for patient care—and right below are the preliminary experiences, completely unadorned:
Varying results related to actual time saved. Varying impact on working hours in the evening. Mixed experiences regarding financial effects. Unclear effects on the number of patient consultations. It is currently too premature to conclude on the effect on coding accuracy.
It is an unusually sober list to publish publicly, and it deserves praise. The project uses the early phase for baseline measurements and experience gathering together with the health trusts – precisely because benefits that are not measured from the start cannot be documented later.
This is happening at the same time as the Norwegian Directorate of Health in its first national mapping of AI in specialist healthcare services points to generative AI for documentation as the clearest development trend. The technology is no longer on the outside of the Norwegian health service. The question has moved from about until How is it going.
The research explains why early numbers look like this
Here comes the most useful insight from this year. A longitudinal study has followed 220 primary care clinicians at 36 clinics across 314,845 patient consultations – 24 weeks before and 24 weeks after an AI documentation tool was integrated into the electronic health record system.
Finding: the effect developed over time. The writing time per note fell by 7 % immediately, on day zero. By day 150, the reduction was 15 % – more than a doubling. The same pattern was repeated for work outside of planned working hours, note length and the proportion of notes closed within 24 hours.
The authors' point is methodical, and it is worth taking to heart: previous studies have largely measured at a single point in time, before and after. This obscures the actual effect, because clinicians adapt to the tool gradually. You learn how to speak, what you can leave to the draft, what you must write yourself.
In other words: a measurement made in week two does not measure the tool. It measures the learning curve.
The study is a preprint and has not been peer-reviewed yet, and the design is quasi-experimental without randomization. Take the numbers as a direction, not as an absolute truth. But the direction matches strikingly well with what South-Eastern Norway Regional Health Authority itself reports from an early phase.
And the gain is not just minutes
The second study is methodologically stronger: a randomized study published in NEJM AI in November 2025, involving 238 outpatient physicians from 14 specialties, comparing two AI documentation tools with standard practice.
The result was a modest, but broadly positive effect. The one tool gave 9.5 % shorter writing time in the note compared to the control group; the second showed no significant change. But what stood out for the users of Both the tools were validated measures of burnout, work exhaustion, and cognitive task load.
It is an important shift in what we measure. We have previously written about What the Research Says About AI Documentation, and the picture that is emerging now is clearer: the most robust gain is not necessarily that you save minutes, but that the burden of documenting goes down. Those are two different things. Eight notes that have to be finished after working hours don't just cost time – they cost energy.
And then the number that is most notable: The tools were used in under a third of the consultations. 33.5 % and 29.5 % in a study where the doctors had received the tool, training, and all the attention that comes with a research project.
That says something essential. A tool that does not fit into the workflow won't be used – no matter how technically good it is. We have touched upon the exact same point before, about how the same tools can provide completely different experiences from clinic to clinic.
What this means for your clinic
Three things follow from this, regardless of which tool you are considering.
Measure a zero point before you start. How much time do you spend on documentation today? How many notes are left until the evening? Without a baseline, you have no way of knowing if things got better—only a feeling, and the feeling in week two is unreliable.
Give it more than a few weeks. The curve from 7 to 15 percent took 150 days. That is not an argument for enduring a bad tool for half a year, but it is an argument for not concluding before you have actually learned it.
Assess the gain on multiple axes than the clock. Fewer notes left at the end of the workday. Less cognitive load. More eye contact in the consultation. Those are real outcomes, and research finds them more consistently than it finds major time savings.
Medivox is built to fit into this daily routine without requiring you to change the way you work. You dictate or let the conversation be transcribed, you get a draft, and you read, correct, and approve it before anything is saved. The medical record is and remains your responsibility – Medivox writes down what was said and done, and makes no clinical judgments. All processing takes place in the Norwegian language in Norwegian data centers, with pseudonymization before the text is processed. We have written more about the division of responsibilities in Normen's advice for the safe use of AI in documenting healthcare.
It takes time to get good at something – this too
There is something liberating about Norway's largest health region publishing a list of things they do not yet know. It allows for a more honest conversation than the one we often get: that new technology in healthcare rarely delivers full impact from day one, and that this does not mean it does not work.
You wouldn't have considered a new patient record system after two weeks. Why would you consider this any differently?
Frequently Asked Questions
How long does it take before AI documentation actually saves time?
A longitudinal study of 220 clinicians found a reduction in writing time of 7 % per note immediately, and 15 % after 150 days. The benefits increase as you become more familiar with the tool.
Is it documented that AI documentation reduces burnout?
A randomized study of 238 physicians, published in NEJM AI in 2025, found improvements in validated measures of burnout, work fatigue, and cognitive load for the users of both tools tested. The effect was modest, but consistent.
Does speech-to-text work for other professions besides doctors?
Yes. The documentation burden is not reserved for doctors, and Medivox is used across professional groups – from physiotherapists and psychologists to nurses and dental health personnel. With the template builder, you can create clinical record templates that suit your own practice and specialty.
Does Medivox make medical decisions for me?
No. Medivox documents what was said and done in the consultation. The tool does not provide clinical decision support, and you always read, correct, and approve the draft before it is saved in the medical record.
What should we measure before introducing a tool in the clinic?
Time spent on documentation per consultation, how many notes are left over after working hours, and how the employees themselves experience the workload. Measure it before you start – otherwise you have nothing to compare it to.
Use Medivox for free – Get started completely free
Are you considering AI documentation in the clinic and wondering how you should measure the impact? Contact us – we are happy to help set up a zero point before you start.
Sources:
- Sykehuspartner (2026): AI for journal – speech to draft
- The Directorate of Health (2026): AI entering Norwegian hospitals at full speed
- Kannampallil et al. (2026): Longitudinal effects of ambient AI scribe use on documentation burden and financial productivity: A quasi-experimental study (preprint)
- Lukac et al. (2025): Ambient AI Scribes in Clinical Practice: A Randomized Trial, NEJM AI 2(12)
- Norwegian Directorate of Health / The Norm (2026): Safe use of AI in healthcare documentation – speech-to-draft
Related posts
Share