EU Released Its Draft High-Risk AI Guidelines. Here's What Life Sciences Companies Actually Need to Read.

On May 19, 2026, the European Commission published the long-anticipated draft Commission guidelines on the classification of high-risk AI systems under Article 6 of Regulation (EU) 2024/1689. The package is distributed across three documents: a general principles section, the Article 6(1) and Annex I route covering AI as a product or safety component of a regulated product, and the Article 6(2) and Annex III route covering the eight stand-alone use case categories. The Commission opened a targeted stakeholder consultation through the EU Survey platform that closes June 23, 2026.

The guidelines were originally due February 2, 2026 under Article 6(5) of the AI Act. They are now out, more than three months late, alongside a separate political agreement between Parliament and Council, reached May 7, 2026, that pushes high-risk application dates further into the future under the Digital Omnibus on AI. Both moves changed the operating reality for any life sciences company building an AI/ML-enabled product intended for the EU market.

This piece is a clause-level walkthrough of what the draft guidelines actually say, where they confirm the existing framework, where they clarify ambiguity, and what medical device, SaMD, IVDR, and life sciences AI manufacturers should do before the consultation closes.

Why the Guidelines Matter Now

Article 6 is the gateway to the high-risk regime. Once a system is classified as high-risk, Chapter III applies: a risk management system under Article 9, data governance under Article 10, technical documentation under Article 11 and Annex IV, record-keeping under Article 12, transparency under Article 13, human oversight under Article 14, accuracy/robustness/cybersecurity under Article 15, a quality management system under Article 17, conformity assessment under Article 43, post-market monitoring under Article 72, and CE marking and EU database registration. The classification decision determines whether all of that attaches.

Until May 19, the only binding interpretive text on Article 6 was the AI Act itself, recitals 47 to 51, and the medical-device-specific guidance in MDCG 2025-6 from the Joint Artificial Intelligence Board. Operators across machinery, toys, radio equipment, lifts, automotive, aviation, and medical devices were classifying systems by reading the recitals against the sectoral harmonisation acts and making defensible calls. The new draft is the first detailed interpretation the Commission has put on the table.

It is a draft. Final adoption is expected later in 2026 or early 2027. The Commission has indicated responses submitted outside the questionnaire will not be reflected in the final summary report. The CJEU remains the only authoritative interpreter. But for practical compliance work between now and final adoption, this is the document that matters.

The Two Statutory Routes to High-Risk Classification

The guidelines organize around the two routes set in Article 6. Both routes still depend on the threshold question that any system must first qualify as an AI system within the meaning of Article 3(1), as interpreted by the Commission Guidelines on the definition of an AI system adopted July 29, 2025. Software systems that do not meet the seven-element definition (machine-based, varying autonomy, optional adaptiveness, explicit or implicit objectives, inference, output generation, environmental influence) are not in scope at all.

For systems that clear the definitional threshold, the two routes are:

• Article 6(1) and Annex I. The AI system is intended to be used as a safety component of a product, or the AI system itself is a product, covered by Union harmonisation legislation listed in Annex I, and the product or system is required to undergo third-party conformity assessment under that legislation. Two cumulative conditions. Both must be met.
• Article 6(2) and Annex III. The AI system falls within one of the eight Annex III use case areas, subject to the Article 6(3) carve-out filter for systems that do not pose significant risk to health, safety, or fundamental rights.

For life sciences, Article 6(1) is the live question for AI/ML medical devices and IVDs. Article 6(2) lands less often, but matters when a life sciences company uses AI for clinical investigator selection (Annex III employment), patient access to essential services, or biometric processing.

The Two-Prong Safety Component Test

The most consequential interpretive contribution in the Annex I document is the Commission's reading of "safety component" under Article 3(14). The definition contains two alternative scenarios, and the draft confirms that either is sufficient.

The first prong is intent-based. An AI system qualifies as a safety component where it is intended to fulfil a "safety function," meaning its intended purpose, as declared by the provider in instructions for use, technical documentation, and promotional materials, is to prevent or mitigate risks to the health and safety of persons or property. The draft gives illustrative examples: an AI computer vision system that detects human presence in a robot cell and triggers a safe stop, an AI system monitoring gas concentrations in ATEX equipment and commanding shutdown, an AI system in a train monitoring speed limits to prevent collisions.

The second prong is consequence-based. An AI system also qualifies as a safety component where its failure or malfunctioning would endanger the health and safety of persons or property, regardless of whether the intended purpose is a safety function. The Commission gives three examples that should stop manufacturers in the lift, automotive, and agricultural sectors from assuming they are outside scope. An AI system managing lift door closing timing is intended for efficient operation, but a malfunction could cause injury. An AI lane assistance system is intended to enhance user experience, but unexpected steering could cause a collision. An AI system targeting areas for chemical spraying is intended to optimize chemical use, but failing to detect nearby persons could endanger them. All three are safety components on the failure-mode prong, even though none of them was designed as a safety function.

The implication for life sciences AI: a medical device manufacturer cannot avoid the safety component classification by declaring the intended purpose as "diagnostic decision support" or "workflow assistance" rather than "safety function." If the system's failure or malfunctioning could endanger health and safety, and most clinically-relevant AI/ML medical devices meet that threshold, the safety component test is satisfied on the second prong.

Failure or malfunctioning, the draft clarifies, includes incorrect outputs (false positives or false negatives), loss of function or availability, performance drift, latency errors, and misclassification leading to hazardous control decisions. Endangerment includes illness, injury, temporary or permanent impairment of body structure or function, chronic disease, and damage to property. The Commission expressly excludes reputational harm, purely financial loss, and minor service degradation. The likelihood of failure must not be only theoretical; it must lead to actual endangerment in the product context.

What This Means for AI/ML Medical Devices

For medical devices and IVDs, the existing MDCG 2025-6 framework from the Joint Artificial Intelligence Board still controls. The new Commission draft does not alter the medical-device-specific mapping; it provides the horizontal interpretive layer above it. The two read together as follows.

High-risk Not high-risk under Article 6(1) · MDCG 2025-6 Table 1, read with Commission draft guidelines May 19, 2026.

The Commission draft adds two clarifications that have practical bite. First, the choice of conformity assessment module under the sectoral legislation does not change the high-risk classification. A manufacturer that could rely on Module A internal control under the Toys Safety Regulation by applying harmonised standards is still operating a high-risk AI system under Article 6(1) and still owes all the Chapter III obligations. This was confirmed for toys by Recital 15 of the Toys Safety Regulation (Regulation (EU) 2025/2509) and Articles 26(2) and 26(3), and the Commission draft extends the same logic horizontally. Manufacturers who assumed they could avoid Chapter III by selecting a lighter conformity assessment module should re-read the analysis.

Second, the AI Act does not in itself extend the scope of any sectoral harmonisation legislation. The classification under Article 6(1) is dependent on, and operates within, the sectoral framework that already applies. The AI Act does not change the risk class of a medical device under MDR or IVDR. The MDR or IVDR classification determines whether AI Act Article 6(1) applies; not the other way around. This was already stated in MDCG 2025-6 and is reinforced in the draft.

The Annex III Route and the Article 6(3) Filter

For systems that fall outside Annex I but within one of the eight Annex III areas, Article 6(2) still imposes high-risk classification by default. The eight areas are biometrics, critical infrastructure, education and vocational training, employment and workers management, access to essential private and public services and benefits, law enforcement, migration and asylum and border control, and administration of justice and democratic processes.

Article 6(3) provides a derogation. An Annex III system is not high-risk if it does not pose a significant risk of harm to health, safety, or fundamental rights, including by not materially influencing the outcome of decision-making. The article lists four conditions that, if any one is met, bring the derogation:

1. The AI system is intended to perform a narrow procedural task.
2. The AI system is intended to improve the result of a previously completed human activity.
3. The AI system is intended to detect decision-making patterns or deviations from prior decision-making patterns and is not meant to replace or influence the previously completed human assessment without proper human review.
4. The AI system is intended to perform a preparatory task to an assessment relevant for the purposes of the use cases listed in Annex III.

The fifth rule, set in the third subparagraph of Article 6(3), overrides all four conditions. Any Annex III system that performs profiling of natural persons is always high-risk. Profiling is defined by reference to GDPR Article 4(4) and covers automated processing to evaluate aspects of a person's life including work performance, economic situation, health, preferences, interests, reliability, behaviour, location, or movement.

The draft guidelines treat the four conditions narrowly. The Commission's posture, consistent across the Annex III document, is that high-risk classification is the rule, the filter is the exception, and the exception should be read tightly. Providers who consider an Annex III system is not high-risk are required under Article 6(4) to document the assessment before placing the system on the market, register under Article 49(2), and produce the documentation on request from national competent authorities. This is not a quiet self-exemption; it generates a paper trail that the surveillance authority can audit.

Intended Purpose and General-Purpose AI

The general principles document reinforces a point that matters for any provider whose AI system has multiple applications or whose system is built on a general-purpose AI model. Where the instructions for use, contractual terms, usage policy, promotional materials, or technical documentation present the AI system as broadly applicable across a generality of contexts, and do not consistently limit application or exclude high-risk uses, the system's intended purpose will be deemed to encompass high-risk use cases. The system qualifies as high-risk.

Merely asserting in the terms of service that high-risk uses are excluded is not enough. The Commission states that if the provider's overall presentation, examples, or product positioning effectively provides for or promotes high-risk uses, the disclaimer is overridden. Limitations of use must be "clearly, concretely, and coherently described across all materials."

For LLM-based products, vendor-provided clinical decision support agents, AI-powered SOP generators sold to regulated companies, and general-purpose models that providers position as fit for regulated workflows, this is the operative test. The Article 25(1) provisions also pull distributors, importers, deployers, and third parties into provider obligations when they brand-label a high-risk system, substantially modify it, or modify the intended purpose of a non-high-risk system in a way that makes it high-risk. Anyone integrating a general-purpose AI model into a regulated medical device or IVD workflow needs to assess whether the integration triggers Article 25.

The Timeline Reality After the Digital Omnibus

The Commission published the draft guidelines twelve days after a political agreement between Parliament and Council on the Digital Omnibus on AI, announced May 7, 2026. The political agreement, if formally adopted, pushes the application dates for high-risk obligations:

The Digital Omnibus political agreement still requires formal adoption by Parliament and Council and publication in the Official Journal, after which it enters into force three days later. The Commission press release on May 7 confirms the political agreement; until formal adoption, the legal default in Article 113 remains in force. Manufacturers planning placement on the EU market under the Article 6(1) route still face the original August 2, 2027 date until the Omnibus is published.

Two things follow. First, the breathing room is real but contingent. The Commission used the Omnibus framing in the draft guidelines themselves, noting that the dates "are now postponed with the AI Omnibus to 2 December 2027 and 2 August 2028 respectively." That is the Commission's working assumption. It is not yet the legal text. Second, the additional time does not reduce the work; it redistributes it. Article 111(2) provides a narrow grandfathering clause: it applies to high-risk AI systems placed on the market or put into service before August 2, 2026, and only where those systems are not subject to significant design changes after that date. A separate provision in the same article requires providers and deployers of high-risk AI systems intended for use by public authorities to comply by August 2, 2030 in any event. Most life sciences AI products do not fit the grandfathering scenario.

What Manufacturers Should Do Before June 23

The consultation closes on June 23, 2026. The next interim states are a Commission summary of feedback, a revised draft (potentially with further consultation), and final adoption. Between now and final adoption, three things should be on a regulated AI/ML product team's desk.

Re-run the classification assessment against the two-prong safety component test. Any AI/ML system embedded in or accompanying a medical device, IVD, machinery, lift, radio equipment, toy, or other Annex I-regulated product should be re-evaluated against both prongs of Article 3(14). The intent-based prong is straightforward; the failure-mode prong is the one that catches manufacturers off guard. Document the assessment, including the failure modes considered and the basis for concluding endangerment is or is not present in the product context.

Lock down the intended purpose declaration across all materials. The Commission has stated that the provider's overall presentation governs over isolated disclaimers. Instructions for use, technical documentation, marketing materials, sales presentations, contractual terms, and product positioning need to be coherent and consistent. For multi-purpose systems and general-purpose AI integrations, this is the harder discipline. Audit what the marketing pages, sales decks, and product documentation actually say. If they imply high-risk uses, the system will be deemed high-risk regardless of the terms of service.

Begin Annex IV technical documentation in parallel with the regulatory submission. Annex IV is the technical documentation requirement under Article 11 for high-risk AI systems. It is substantive, system-design-specific, and overlaps with MDR Annex II technical documentation for medical devices but is not identical. The post-market monitoring plan under Article 72, the quality management system under Article 17, and the conformity assessment pathway under Article 43 all interact. Starting Annex IV after the rest of the technical file is complete creates rework. Starting in parallel does not.

Submit feedback if the draft missed something material. The Commission has been explicit that feedback received outside the EU Survey questionnaire is not reflected in the final summary. Sector-specific gaps (life sciences operators have several) should be submitted through the questionnaire by close of consultation. The AI Act Service Desk is the Commission's working channel for ongoing questions after the consultation closes.

Where This Leaves the High-Risk Question

The draft guidelines do useful interpretive work. The two-prong safety component test resolves a question that has been answered inconsistently across early operator self-assessments. The Annex III filter is read narrowly, which means more systems land in the high-risk category than some providers had assumed. The conformity assessment module clarification closes a workaround. The intended-purpose discipline closes another. The medical-device-specific mapping in MDCG 2025-6 is reinforced and operationally usable today.

What the guidelines do not do is reduce the surface area of work. The Annex IV technical documentation, the Article 17 QMS, the Article 9 risk management system, the Article 14 human oversight controls, and the Article 72 post-market monitoring plan all still need to be built and maintained. The Digital Omnibus extension does not change the deliverables; it changes the date by which they must be ready.

For a medical device manufacturer with an AI/ML-enabled product on the trajectory toward EU placement, the answer to "should we wait for the final guidelines" is no. The interpretive direction is now clear enough to build against. The companies that compress the work into the final months before the application date will do so under time pressure, with less room for diligence and with a regulator who has already published the position they will enforce.