Healthcare AI Under the EU AI Act

Validaitor
7 days ago
6 min read

Updated: 5 days ago

Clinical AI usually qualifies as a medical device under EU rules and is therefore high-risk under the AI Act, while administrative AI usually is not.

The product's stated intended purpose decides which category applies.

Requirements

Classification Examples

Classifying Clinical AI under the EU AI Act

The EU AI Act's "high-risk list" names only a few healthcare uses, but that list is not the main route for clinical AI. AI that diagnoses, guides treatment, or monitors a patient's condition is usually a medical device under EU rules (MDR/IVDR), and being considered a medical device makes it high-risk under the AI Act as well.

Most healthcare administration AI, such as scheduling, transcription, and billing, is in neither category and carries only light duties.

The boundary between the two follows the product's claims. An app described as helping users track sleep is a wellness product; the same app described as detecting sleep apnea is a medical device.

Hospital procurement teams, investors, and notified bodies request evidence of this classification, both whether you classify as High-Risk or classify outside High-Risk for your Healthcare AI under EU AI Act.

When do the rules apply?

Dates checked against the European Commission's pages on July 2, 2026.

Date	What applies
Feb 2025	In force. Bans, plus AI literacy: clinical and admin staff using AI must be adequately trained.
Aug 2026	Transparency: patient-facing chatbots must say they are AI; AI-generated content needs a machine-readable marker.
Dec 2, 2027	High-risk rules for the Annex III spots: emergency triage, life and health insurance pricing, public healthcare benefits.
Aug 2, 2028	High-risk rules for AI in medical devices (the Route 2 cases).

The later dates apply only to the AI Act's high-risk requirements. Medical device rules (MDR/IVDR) are already in force.

Intended purpose determines the category

Under EU device rules, classification follows the intended purpose the manufacturer states for the product. The website, app-store description, sales deck, and manual all count as statements of intended purpose. A general wellness app is not a medical device; the same app claiming to detect a condition is.

Products often change category through incremental feature changes:

A scribe tool that starts suggesting diagnoses in the note
A wellness tracker that adds diagnostic feedback "this pattern looks like atrial fibrillation — see a doctor"
An appointment chatbot that begins assessing symptoms to decide urgency

A single feature or marketing claim can change the category, so claims review belongs in product review.

What are the requirements?

If your AI is high-risk (either route), before selling or using it in the EU you need:

Risk management across its life
Good training data, checked for bias
Technical documentation and automatic logging
Clear instructions for users
Human oversight
Accuracy, robustness, and cybersecurity
A quality system, a conformity assessment (via your notified body for medical devices), CE marking, and EU registration
Monitoring and incident reporting after launch

If you deploy someone else's high-risk system — hospitals, clinics, insurers:

follow the provider's instructions, assign trained staff to oversee it, monitor its operation, keep logs (at least six months), and tell patients when AI makes or shapes decisions about them.

If your product is a wellness or admin tool:

maintain the classification file; it is the basis for stating that the product is neither a medical device nor high-risk.

Whatever your category:

AI literacy applies now; transparency disclosures from Aug 2026.

Common Healthcare AI classification examples under EU AI Act

The nearby function	Usually outside high-risk	Usually high-risk
Clinical documentation	Transcribing consultations, drafting notes the clinician signs	Generating diagnosis or treatment suggestions in the note
Diagnosis	Searching literature and guidelines for clinicians	Analyzing patient data to flag or rule out conditions
Triage	Booking appointments, answering practice FAQs	Symptom checking that directs urgency or care
Monitoring	Step counts and general wellness trends	Detecting deterioration or arrhythmia and alerting care
Insurance	Explaining coverage, helping with forms	Pricing or risk-scoring individuals for life and health cover
Public benefits	Explaining rules, helping with applications	Recommending who qualifies

No authority pre-approves a classification.

Providers document their own assessment before launch, and for medical devices a notified body checks it. Classification is therefore within the provider's control: the product's design and claims determine the category, and the classification file is the evidence.

The classification documentation

Most AI in healthcare is not high-risk: scheduling, documentation, billing, and general wellness features carry only light duties. The classification file is the evidence for that conclusion, and it is what hospital procurement teams, enterprise customers, and investors request. It contains:

The system's intended purpose and the claims made on the website, app store, sales material, and manuals
What the system does not do: no diagnosis, no triage, no individual risk scoring
The assessment against the high-risk list and medical-device rules, with the reasoning
The changes that trigger reassessment, such as a new claim, feature, model, or market

For systems close to the high-risk list, the Act requires the assessment to be documented before launch. For all others there is no legal duty, and the file exists because buyers and due-diligence reviewers request it. The classification depends on the claims and features the file describes, so it is reviewed whenever those change.

Two routes into high-risk for healthcare AI

Route 2, the more common one: AI as a medical device

AI intended to diagnose, treat, monitor, predict, or alleviate disease is usually a medical device. Most diagnostic and treatment software needs review by an independent body (a "notified body") under medical-device rules, and where that is the case, the AI Act's high-risk requirements apply as well.

In practice, expect this route if your AI:

Analyzes images, signals, or lab data to flag or rule out conditions
Predicts deterioration, sepsis, or readmission risk for individual patients
Recommends or calculates treatment or dosing
Interprets ECGs, monitors arrhythmias, or alerts care teams
Runs a symptom checker that directs people to or away from care

Route 1: three named healthcare spots on the high-risk list

The Act's high-risk list ("Annex III") adds three healthcare-adjacent uses that are high-risk even without being a medical device:

Emergency call triage and dispatch, including triage of patients in emergency care
Life and health insurance risk assessment and pricing for individuals
Public healthcare benefits eligibility decisions

Neither: admin and back office

Scheduling, ambient scribing and transcription, billing and coding, literature search, and drafting letters a clinician reviews and signs are usually in neither route. Light transparency duties can still apply (a patient-facing chatbot must say it is AI), but not the heavy rules.

The overlap with existing MDR work

Much of what the AI Act asks for resembles an MDR file: risk management, technical documentation, a quality system, post-market surveillance. The Act is designed to integrate with that regime, so one conformity assessment through your notified body covers both, rather than a second parallel process.

What the AI Act adds is AI-specific:

Training data governance — data quality and bias checked and documented
Automatic logging across the system's life
Human oversight designed in — clinicians must be able to understand, question, and overrule outputs
Accuracy and robustness evidence for the AI itself, including behavior on data it wasn't trained on
Transparency to users — capabilities, limitations, and conditions of use spelled out

Treating this as an extension of existing MDR work avoids running two separate compliance processes.

Penalties and earlier costs

Fines are the higher of a fixed amount or a share of worldwide annual revenue:

Violation	Maximum fine
Using a banned practice	€35 million or 7%
Breaking high-risk and other rules	€15 million or 3%
Misleading authorities	€7.5 million or 1%

Startups and small companies pay the lower of the two. Medical-device enforcement applies separately: an unclassified device can be pulled from the market regardless of the AI Act.

Other costs usually arrive earlier than fines: hospital deals stalled in procurement, notified-body reviews failed late, investors treating unclear exposure as compliance debt, and expensive retrofits of data governance and logging after the model is built.

How to limit the impact

List your AI systems and features — shipped and in the pipeline, including AI features inside a larger product.
Classify each one. Medical device or not? Which class — is a notified body needed? On an Annex III spot? Or genuinely admin? Record the conclusion and the reasoning in a classification file.
Audit your claims. Website, app store, sales decks, manuals. Marketing copy counts as a statement of intended purpose, so it must match the classification you rely on.
Design to stay in the intended category. If a feature would cross into diagnosis or triage, make that a deliberate, planned step rather than scope creep.
Plan the conformity work once. If you're heading into high-risk, align AI Act evidence (data governance, logging, oversight) with your MDR timeline from the start.
Re-check on every change. A new claim, feature, model, or market can change the product's classification. Assign this re-check to a named owner.

How Validaitor can help

Our goal is to enable innovation and the use of AI, not to add compliance burden. Most AI use cases do not need heavy compliance. The work is to identify which ones do, document the assessment, and keep it current as the product evolves. Validaitor is an AI governance platform built for that work:

Boundary scan A short review of one AI system: which role you hold (provider, deployer — often both), which rules apply, red flags, missing evidence.	"Not high-risk" classification The classification file that documents why a tool is admin or wellness, where the facts support it.
High-risk readiness Risk management, documentation, logging, oversight evidence, aligned with your MDR work.	Change monitoring So new features and claims don't silently break your classification.