Healthcare and medtech: getting AI adopted in regulated professions

It is often said that physicians, nurses, quality engineers and regulatory affairs managers resist AI. We think that diagnosis is lazy. These professionals ask a legitimate question that many projects avoid: if the tool is wrong, who answers, to whom, and on what basis? As long as the answer is vague, not using the tool is the rational behavior. Adoption in a regulated profession is therefore first a work of clarification, and only then a work of persuasion.

This is central to how we approach organization and adoption: you do not change practices by asking people to forget their obligations.

The brake is not culture, it is accountability

A Swiss clinician is bound by professional secrecy. A company handling health data falls under the revised Federal Act on Data Protection (FADP), in force since 1 September 2023, which treats such data as sensitive and in some cases requires an impact assessment. A medical device manufacturer works within a quality system in which any change of tool or process must be controlled. And the EU AI Act, in force since 1 August 2024, applies in phases: prohibitions from February 2025, general-purpose AI obligations from August 2025, most of the rest from August 2026, with their own timelines for products already covered by sector legislation such as medical devices.

Faced with this stack of rules, a professional who does not use an assistant is not being static. They are protecting themselves and their patients. We have written about why AI tools end up in the drawer: in regulated professions, the drawer is simply deeper, because the cost of an error is higher and less evenly shared.

Three zones of use, three adoption regimes

Our position is that "AI in healthcare" should not be treated as a single block. Uses fall into three zones, and they should not share the same governance.

Zone 1: administration and support

Scheduling, standard letters, summaries of internal documents, answers to non-clinical questions from patients or customers. No clinical decision is involved and health data can often be kept out. Adoption follows classic change management, and the lessons from deploying an AI assistant in customer service apply almost as they are.

Zone 2: documentation support

Clinical reports, discharge letters, complaint files, post-market surveillance records, draft regulatory documents. The tool proposes, the professional reviews, corrects and signs. This is where the gain is most visible and where adoption is actually decided. The conditions are strict: data processing compliant with the FADP, documented hosting and subprocessors, and, on the medtech side, validation of the tool when it operates within the quality system.

Zone 3: clinical decision and product

Anything that steers a diagnosis or a treatment, or is embedded in a device. This brings in qualification as a medical device, evidence and traceability requirements, and AI regulation. It is not an adoption project but a development and compliance project, with its own teams and its own calendar.

Our advice, stated plainly: do not start with zone 3 to "show ambition", and do not stay in zone 1 out of excess caution. Starting with zone 2 and systematic human review gives the best balance between gain and risk. It proves value, and it builds the governance reflexes zone 3 will later require.

An order of magnitude to size the stakes

Take a deliberately simple example, with explicit assumptions. A hospital has 40 clinicians who each write reports. Suppose a drafting aid, always reviewed by the author, saves 25 minutes per person per day over 220 working days. That is 40 x 25 x 220 / 60, or roughly 3,700 hours a year. If only half the clinicians use it regularly, we are still in the order of 1,800 hours.

These figures are not a forecast. They are a way to ask the right question: how many clinical hours returned to patients, or removed from overtime, justify a serious effort in framing and training? In most cases the answer supports the investment, provided real usage follows. The risk is not that the tool is poor. It is that it is purchased, legally cleared, and then barely used.

What leaders need to decide

Four decisions, in our view, unlock most of the work.

Name one owner per use. One person, not a committee, answers for the use: scope, data, quality of outputs. In an SME or a mid-sized clinic this is often the quality manager or the medical director, supported by data protection.

Write one page of rules, with the professionals. What is allowed, what is forbidden, what must be reviewed. One page, not a manual. If nurses or engineers do not recognize their own work in the text, it will not be applied.

Check by sampling, not by principle. Human review of every output is essential at user level, but management should also regularly inspect a sample: errors caught, frequent corrections, drift. This reassures professionals, because it shows accountability is not shifted onto them alone.

Train by role, briefly. A physician, a nurse, a regulatory affairs officer and a financial controller do not share use cases. A short session built around their real documents beats a generic program. We set out this logic in training teams in AI without drowning them.

Slow in appearance, fast in practice

It is tempting to read these precautions as friction. We see the opposite. Organizations that clarify accountability early move faster afterwards, because each new use fits into a framework already accepted. Those that improvise end up with parallel, undeclared uses, which is the worst outcome for both confidentiality and quality.

AI does not need to be everywhere in a healthcare provider or a medtech company. It needs to be where it frees professional time, under rules that professionals recognize. To place your organization across these three zones and identify the first use worth framing, you can start with our diagnostic, and read about our work in the healthcare and medtech sector.