#42 In healthcare, the best screen is the one you switch off

If you want to know what the next relationship between humans and machines looks like, do not watch startup demos. Watch a doctor's surgery. It is the one place where the cost of the keyboard is measured in visible damage to a human relationship, and therefore the first ground where computing is learning to disappear.

The problem nobody dared name

For fifteen years, the electronic health record has put a screen between the doctor and the patient. The clinician types while the patient talks. They document in the evening, at home, in what is known as « pyjama time ». Burnout among health professionals has several causes, but the documentation burden is one of them, and it is extensively documented.

Note the nature of the problem: this is not a medical problem, it is an interface problem. The tool meant to help introduced a data-entry task that did not exist and that eats into consultation time. It is exactly the anomaly described in the first two articles of this series, in an aggravated form.

The ambient scribe, or computing that listens instead of being driven

The technical answer is called ambient clinical documentation. A microphone captures the conversation between clinician and patient. The system separates the speakers, transcribes, structures, produces a draft note, and the clinician reviews and signs off, usually in under a minute. Nobody types. The computer goes back to being what it should always have been: a scribe, not an interlocutor.

Results are starting to be measured seriously, which changes everything. A study published in JAMA across five American academic medical centres reports a fall of about 13.4 minutes in time spent in the health record and 16 minutes in documentation time, associated with roughly half an additional consultation per week. A matched before-and-after analysis of 6,026 outpatient notes at Mass General Brigham was published in February 2026, covering the Nuance DAX and Abridge tools.

Let us stay measured. These gains are real but modest against the enthusiasm of the market, part of the literature points to a persistent gap between expectations and results, and the effect on note quality remains debated. This is not a revolution, it is a significant and verifiable improvement. That is already far more than most digital projects deliver.

In France, the movement is funded and framed

This is not an American subject. The 2026 e-health barometer suggests 55 % of hospitals already use AI copilots, around 70 % of them for diagnostic support or imaging. An « AI-augmented hospital » programme backed by France 2030 was announced in spring 2026 with funding of around 15 million euros. An envelope of 119 million euros has been announced to train 500,000 health professionals. And the national AI and health data strategy for 2025-2028 is built around three priorities: imaging, clinical decision support and the automation of medical notes.

Look at the third priority. A state that writes « automation of medical notes » into its national strategy is officially acknowledging that data entry was the problem.

The guardrail is solid

It is worth stating what is authorised and what is not, because that is where the ridge line runs.

In the United States, more than a thousand medical devices incorporating AI have been cleared by the FDA, about 95 % of them through the simplified 510(k) pathway, and three in four fall under imaging. The first autonomous diagnostic system, IDx-DR, was authorised in 2018 for diabetic retinopathy. Since then, no autonomous prescribing service has been cleared. In Europe, the Medical Device Regulation and the AI Act overlap, and medical AI falls into the high-risk category.

Translation for a decision-maker: the market authorises assistance on a massive scale and autonomy almost never. This is not passing caution, it is doctrine, and it is an excellent mental model for every other sector. The machine prepares, the human decides and signs.

What this says about your front desk, your office or your reception

Transpose it. In a tourist office, an adviser spends a significant share of their time re-entering data, formatting it, hunting for information across three different systems while the visitor waits in front of them. At a town hall reception desk, an officer types while the citizen talks. The problem is structurally the same as the one in the doctor's surgery, less severe and therefore less addressed.

The right question to ask your teams this week is not « how do we use AI ». It is: at what moments does a screen come between us and the person we are serving, and how many minutes does that cost per interaction. You can measure it in half a day of observation, with no software and no budget.

So what now?

Healthcare shows the way because it is the sector where the cost of a bad interface was the most painful and the most measurable. It also shows the limit: assistance yes, autonomy no, and human validation as a compulsory checkpoint.

What remains is to decide what to do concretely in your own organisation, in what order, and with what budget.