By now, it is clear: We are in a new game. The iterative approach of manually compiling and submitting documents to Notified Bodies, followed by working through nonconformity reports and remediation cycles, will very soon no longer work.
What you should do about it is outlined in today’s Executive Corner.
At the suggestion of the International Medical Device Regulators Forum (IMDRF), work has begun on a new standard intended to provide guidance on applying ISO 14971 to combination products across the entire lifecycle.
The new document will take into account and build upon the “Risk Management Guidance for Combination Products” issued by AMII.
b. EU
Commission versus Notified Bodies
In January, the European Commission organized a webinar to facilitate an open discussion with Notified Bodies regarding the draft implementing regulation. The topic: requirements for conformity assessment activities.
Reading through the questions and answers, I had the distinct impression that relations were somewhat strained. So I asked AI for a “neutral” assessment using the following prompt:
“Can you give me a very brief summary of this webinar and an assessment of the general mood between the European Commission and the Notified Bodies? Perhaps you can infer the latter from the discussion and the Q&A.”
The AI’s summary:
“The Notified Bodies apparently feel unheard and consider the regulation impractical and burdensome. The Commission maintains its position and points to flexibilities that the Notified Bodies perceive as insufficient. In the end, the Notified Bodies thanked the Commission for the “constructive discussion,” but the chat reveals clear frustration and fundamental disagreements.”
Whether the Commission is doing the system any favors at this point is open to doubt.
RoHS
There is nothing new in the field of RoHS. However, a reader once again points out that key RoHS lead exemptions (including 7a for high-temperature solder and 7c-I for glass/ceramic components) will expire between December 2026 and June 2027.
This is particularly critical for medical devices, as lead-free alternatives may have technical disadvantages.
For those who wish – or need – to dive deeper into the topic (e.g., to assess which exemptions remain applicable, to analyze alternatives, or to conduct a risk-benefit evaluation), I recommend reaching out to our Product Safety Experts for support. Just contact them through the usual channels.
Executive Corner: The End of Sampling
a. When the Notified Body Reviews in Seconds What Takes You Weeks to Fix
Notified Bodies are upgrading. AI-supported review of technical documentation is no longer a vision of the future – it is happening. The rules of the game are changing fundamentally: full review instead of sampling. Seconds instead of days.
That sounds drastic. It is. But it becomes entirely plausible once you understand the new asymmetry.
b. The New Asymmetry
A false positive generated by a Notified Body’s AI costs the Notified Body seconds – an auditor takes a brief look and checks it off.
An overlooked error in your documentation costs you weeks: additional requests, internal alignment, corrections, iteration, and new approvals.
The Notified Body’s AI is infinitely scalable. Your capacity to correct errors is not.
Those who remain reactive will drown in findings. The question is therefore not whether you will change something – but what, and how quickly.
c. Three Levels of Action
1. Reactive: The Minimum (immediately implementable)
With your next submissions, you will “harvest” more findings. Not because your documentation has deteriorated – but because the review process has improved.
Damage control
AI-supported quality checks before every submission
A strong focus on consistency: Are cross-references correct? Are version statuses synchronized? Does Chapter 3 contradict Chapter 7?
This is damage control. Necessary – but not sufficient. Reactive correction does not scale. At some point, the bucket overflows.
2. Preventive: Avoid Errors Instead of Finding Them (mid-term)
If you search for errors at the end, you have already made them – in a hundred places simultaneously. The solution: integrate AI into every step of your workflow.
Three modes of operation:
Assistance: AI supports in real time. It suggests improvements and warns of inconsistencies.
Creation: AI generates deliverables itself (e.g., drafting instructions for use based on your product data).
Immediate verification: AI reviews every deliverable instantly and demands correction, not three months later.
The Cultural Shift: The role of your experts changes – from creators to curators. AI produces; humans take responsibility.
This is not a demotion – quite the opposite. But it requires new competencies and honest conversations within your teams. Not everyone will celebrate this shift.
3. Structural: The Necessity of Structured Information (strategic)
Now it becomes uncomfortable: The narrative-text-based approach is no longer viable. Why?
Documents create redundancy. The same information appears in ten different places.
Redundancy creates inconsistency: one version here, another version there.
Inconsistency creates findings.
This is not an AI problem. It is a structural issue. One that AI merely exposes without mercy
The Paradigm Shift
Today
Tomorrow
Create documents
Generate structured information
Copy/paste of content
Single source of truth
AI extracts data from documents
AI works directly on data
Versioning via file names,
e.g. “TD_final_v3_FINAL2.docx”
Versioning inherent in the system
Documents as the primary deliverable
Documents only as exports when externally required
Concretely, structuring information means using data formats such as YAML or JSON; schemas (e.g., JSON Schema) that map your regulatory requirements and describe your products; data stored in databases, not in folder hierarchies.
Yes, this partly resembles an IT project. But above all, it is a decision about how you intend to work in the future. The technology can be solved. Habits are harder.
(If you would like to familiarize yourself with these formats, we offer dedicated seminars. But that only as a side note.)
d. The Long-Term Perspecitve
Where is this all heading?
AI as state of the art. Today, your auditor may still raise an eyebrow and speak of “hallucinations” when you use AI in regulatory processes. However, it will not take long until the question reverses: Why do you believe you can still develop state-of-the-art products without AI? What is currently viewed with skepticism will soon be considered baseline professional practice.
AI talks to AI. The manufacturer’s AI communicates directly with the Notified Body’s AI. Reviews are prepared machine-to-machine. Humans intervene only in cases of escalation, but then with full focus on the truly complex questions. Difficulties arise? Your AI consults our AI – that of the Johner Institute.
The human as final authority. Experts will concentrate on what AI (still) cannot do: exercising judgment, weighing risks and benefits, taking responsibility. These are not residual tasks. They are the core.
Paradoxically, the value contribution of experts increases, not decreases. Because they can focus on what truly matters instead of maintaining cross-references.
And what if the manufacturer’s AI and the Notified Body’s AI disagree? Then you need a competent third party with a clear system architecture and well-designed data structures. There are solutions for that as well.
e. And Now?
Don’t panic. But don’t wait, either. Notified Bodies are already using AI.
First step: an honest assessment. How much redundancy exists in your documentation? How often do you copy content from one document into another? If the answer is “constantly,” you know where to begin.
Second step: a pilot project. Take one product. Convert its documentation – or at least one file within it – to structured data. Make mistakes. Learn. Then scale.
Those who create the data foundation today will be capable tomorrow.
Those who wait will be forced to react – under time pressure, at increasing cost, and with the uncomfortable feeling that the Notified Body was faster once again.
Practical Tips
a. Operating LLMs Yourself
Many manufacturers feel uneasy about allowing large language models to process “company secrets.” If you decide to operate LLMs yourself, here are some tips:
We use a self-hosted version of LibreChat for this purpose.
b. Use Set Theory
What should you do if AI-based Technical Documentation reviews produce multiple largely correct – but not identical – results for identical queries?
My tip: Do what you would do with multiple human reviewers. Their findings are usually correct as well, but rarely identical.
If you want to maximize sensitivity and be on the safe side, build the union of findings.
If you would like to minimize effort and false positives and increase specificity, build the intersection of findings.
Of course, you should not perform this set logic manually. I will soon explain how to automate this step as well.
Best regards,
Christian
Founder Johner Institute
Podcast “AI Deployment at Notified Bodies”
Andreas Purde from TÜV SÜD reports on the use of AI at his Notified Body. He describes and evaluates areas of application, provides clear recommendations for manufacturers, and anticipates the broader implications of this epochal disruption.
Practical Workshop “AI Power for the development and approval of medical devices”
March 10, 2026
Learn in just one day how to use large language models (LLMs) specifically in developing medical devices – from the initial idea to post-market surveillance.
This practical workshop with product expert Mario Klessascheck and regulatory and AI expert Prof. Dr. Christian Johner will teach you specific methods that will enable you to drastically reduce the time needed to develop and document your medical devices while remaining compliant with the law.
IEC 60601-1 Safety of Medical Electrical Equipment
March 23-24, 2026
In this seminar, you will develop a comprehensive understanding of the basic concepts of the IEC 60601 family of standards. You will learn how to successfully implement the relevant requirements in practice using various exercises and how to create standard-compliant documentation.
Class IIa medical devices are among the most common products on the European market. Our latest article explains which products fall into this class, which classification rules are relevant, and what regulatory consequences result from this classification.
