Clinical AI governance — EU AI Act enforcement begins in Q3.

The competent authority postures: BfArM, AIFA, HAS, AEMPS, Swissmedic

BfArM — the Bundesinstitut für Arzneimittel und Medizinprodukte — has been the most institutionally aggressive of the EU competent authorities in preparing for enforcement, and the German posture has consequently shaped the broader implementation pattern. BfArM's published guidance under the Leitfaden für Hochrisiko-KI-Systeme im Gesundheitswesen, finalised on 14 March 2026 after a six-month consultation period, establishes the most detailed conformity-assessment requirements of any national competent authority and is being treated by industry as the de facto reference text across the German-speaking jurisdictions. The Leitfaden requires deployers of high-risk clinical AI systems to maintain documented institutional governance covering eight specific elements: a designated AI Officer with formal responsibility within the deployer organisation; documented training records for all clinical staff interacting with the system; a quality management system aligned to either ISO 13485 (for medical-device-classified systems) or ISO/IEC 42001 (for AI management systems generally); incident reporting procedures with documented escalation paths to BfArM; post-market surveillance arrangements with documented frequency and scope; data governance documentation covering training-data provenance, deployment-data flows, and retention periods; risk management documentation aligned to ISO 14971; and a continuous monitoring obligation that requires the deployer to track performance metrics against pre-specified thresholds. The Leitfaden's eight-element architecture is the most granular set of institutional obligations any EU competent authority has published. BfArM has signalled that compliance with the Leitfaden will be the benchmark German enforcement actions are calibrated against.

AIFA — the Agenzia Italiana del Farmaco — has been less prescriptive than BfArM in its public guidance but materially more aggressive on the cross-border data flow provisions. AIFA's published implementation guidance under the Linee guida sull'applicazione del Regolamento (UE) 2024/1689 nei sistemi sanitari, dated 22 April 2026, focuses substantively on the data residency obligations the Regulation imposes for high-risk clinical AI systems whose deployment involves cross-border data flows within the EU. The Italian posture is that any deployment touching Italian patient data must maintain processing infrastructure within EU member-state jurisdictions and that transfers to third countries — including the United Kingdom following the post-Brexit adequacy decision and the United States under any successor framework to the Trans-Atlantic Data Privacy Framework — require additional safeguards beyond what GDPR alone requires. The Italian data residency posture is stricter than the position the European Data Protection Board has taken in its general AI Act guidance, and it creates a structural conflict for vendors operating multi-national deployments that include Italy. Several US vendors operating European deployments — Microsoft Nuance, Glass Health, and Aidoc among them — have indicated through procurement channels that they will maintain dedicated Italian processing infrastructure to accommodate the AIFA posture, which materially raises the cost of operating an Italian deployment for any vendor not already at a scale that supports dedicated national infrastructure.

HAS — the Haute Autorité de Santé in France — has taken a posture distinct from both BfArM and AIFA. HAS's published guidance under the Cadre de référence pour l'évaluation des systèmes d'IA de santé à haut risque, dated 8 May 2026, focuses on the clinical-evidence requirements deployers must maintain for high-risk clinical AI systems. The French posture is that the conformity-assessment evidence required by the Regulation's Annex IV must include institution-specific clinical validation evidence demonstrating the system's performance in the actual deployment context, not just the general clinical validation evidence the vendor produces for CE marking. This is a meaningfully stricter posture than either Germany or Italy and is closer to the FDA Pre-Submission posture in the US than to the standard CE-marking framework that has historically governed medical-device evaluation in Europe. The HAS guidance has produced substantive concern among European vendors who anticipated that CE marking alone would suffice for French deployment; the institutional clinical validation requirement adds a deployment-cost layer that the original AI Act drafting did not unambiguously require. HAS's position is that the institutional validation requirement is implicit in the Regulation's Article 9 risk management system obligations and that French deployments must comply. The legal interpretation will likely be tested before the AI Office and possibly the European Court of Justice before the question is definitively resolved.

AEMPS — the Agencia Española de Medicamentos y Productos Sanitarios — has taken the most permissive posture among the four principal competent authorities, with published guidance under the Guía para la implementación del Reglamento (UE) 2024/1689 en sistemas de inteligencia artificial sanitaria of high risk, dated 16 May 2026, that interprets the Regulation's institutional governance requirements at the floor rather than at any extended interpretation. The Spanish posture privileges legal compliance over operational gold-plating and is consistent with AEMPS's broader regulatory philosophy of producing predictable, minimally-modified-from-EU-baseline implementation guidance. The structural effect is that vendors and deployers operating in Spain face a lower compliance overhead than those operating in Germany, Italy, or France, which has begun to influence procurement decisions at multinational hospital networks. Two of the European university hospital networks tracked for this report have indicated through procurement channels that their first cross-border deployment expansions over 2026 will be into Spain rather than the larger German, Italian, or French markets, on the documented argument that the compliance-overhead differential favours Spain at this stage of the enforcement cycle. Whether AEMPS retains its permissive posture as enforcement begins or shifts toward the more demanding interpretations seen in Germany and France is the open question for Spain over the second half of 2026.

Swissmedic — Switzerland's medical-device competent authority, operating under the bilateral agreement between Switzerland and the EU rather than as a Regulation-binding authority — occupies an observer role with substantive practical influence. Switzerland is not a member state of the EU and is not formally bound by the AI Act. However, Swiss-resident medical-device manufacturers seeking to sell into the EU single market must demonstrate AI Act compliance, and Swissmedic has signed a memorandum of cooperation with the European Commission's Directorate-General for Health and Food Safety to align Swiss regulatory practice on AI-enabled medical devices with the AI Act framework. Swissmedic's published guidance, dated 28 April 2026, is structurally aligned to the AI Act but introduces two specifically Swiss provisions: a requirement that Swiss-resident manufacturers maintain conformity-assessment documentation in at least one of the four Swiss national languages (German, French, Italian, or Rumantsch) in addition to the English documentation the EU framework requires; and a requirement that deployments involving Swiss patient data be processed within Switzerland or under specific bilateral cross-border data flow arrangements that Switzerland has been negotiating with EU member states on a case-by-case basis. Swiss-resident clinical AI vendors — a small but commercially consequential category including Sophia Genetics, Mindmaze, and several smaller players — have built their compliance architecture around the Swissmedic posture and are positioned to meet the 2 August enforcement date.

Conformity-assessment requirements and the CE-marked vendor landscape

The Regulation's conformity-assessment requirements for high-risk clinical AI systems sit on top of the Medical Device Regulation (MDR) framework that has governed clinical software in Europe since 2021. The interaction between the AI Act and the MDR is structurally complex and is the principal source of implementation difficulty for vendors operating European deployments. A high-risk clinical AI system that qualifies as a medical device under the MDR must comply with both regulatory frameworks: the MDR's classification as Class IIa, IIb, or III medical device with corresponding clinical evidence and post-market surveillance obligations, and the AI Act's institutional governance, conformity-assessment, and continuous monitoring obligations under Title III, Chapter 2. The two frameworks overlap substantially on documentation requirements but diverge on specific obligations: the MDR focuses on the medical-device manufacturer's responsibility, while the AI Act extends additional obligations to the deployer (the hospital or healthcare provider operating the device). This deployer-side obligation set is the principal new compliance burden the AI Act introduces.

As of 28 May 2026, the CE-marked vendor landscape for high-risk clinical AI systems comprises approximately 340 distinct devices across 87 vendors, according to the European Commission's MDR/IVDR EUDAMED database. Of those 340 devices, approximately 215 fall within the high-risk AI system category as defined by the Regulation's Annex III; the remainder are either MDR-classified medical devices that do not meet the AI Act high-risk threshold (lower-risk categorisations such as Class I devices or devices that do not include the autonomous decision-making characteristics the AI Act specifies) or are AI-enabled but not classified as medical devices under the MDR (administrative and operational AI tools that fall outside the MDR scope entirely). The 215 high-risk AI medical devices comprise the operational core of the AI Act's enforcement phase. The largest vendors in this category are Microsoft Nuance with 23 distinct CE-marked devices, Aidoc with 18, Siemens Healthineers with 17, Philips Healthcare with 15, GE HealthCare with 12, Glass Health with 8, Abridge with 7 (a recent expansion from US-only operations), and Suki AI with 4. The remainder of the 215 high-risk devices are distributed across smaller vendors, with no single additional vendor holding more than 4 distinct devices.

The structural readiness of the named vendor landscape for the 2 August enforcement date is highly variable. The major established vendors — Microsoft Nuance, Siemens, Philips, and GE HealthCare — have substantive compliance infrastructure already in place from their MDR compliance programmes and have signalled through their published statements that they will be operationally ready for enforcement. Aidoc, Glass Health, Abridge, and Suki AI are smaller but well-capitalised vendors with sophisticated regulatory affairs functions and have signalled similar readiness. The 79 smaller vendors holding the remainder of the 215 high-risk devices are a more concerning category. Industry surveys conducted by the European Medical Device Manufacturers Association (EMDMA) suggest that approximately 30 to 35 per cent of these smaller vendors will not be operationally ready for the 2 August enforcement date and will face material compliance risk during the early enforcement period. The competent authorities have signalled, through various published statements, that the early enforcement period will be administered with what BfArM's published guidance calls "proportionate flexibility" — that is, that the competent authorities will focus initial enforcement actions on flagrant non-compliance rather than on technical or procedural omissions in good-faith compliance efforts. The proportionate flexibility window is not formally bounded but is generally understood to last between six and twelve months from the enforcement date.

The deployer-side obligations are creating their own substantive compliance challenges. The Regulation requires deployers of high-risk clinical AI systems to maintain documented institutional governance covering the elements that BfArM's Leitfaden specifies and that the other competent authorities' guidance documents identify in more or less equivalent form. For hospitals operating multiple high-risk AI systems across different specialties, the documentation burden is substantial. The three European university hospitals adopting Cleveland Clinic's ACIS framework — Charité, Karolinska, and AKH Vienna — have all built their ACIS implementation around AI Act compliance, with their dual-log architectures, DPIA gating, and clinical review panel structures explicitly designed to produce the conformity-assessment evidence the Regulation requires. Hospitals without comparable governance infrastructure face a more difficult readiness path. The institutional compliance overhead is the principal reason hospital procurement officers cite for choosing vendors with strong compliance infrastructure over smaller, more specialised vendors who may offer superior clinical capabilities but cannot underwrite the deployer-side documentation requirements as comprehensively.

Cross-border data residency rules and hospital-level obligations

The cross-border data residency provisions of the Regulation, operating in interaction with the GDPR's data transfer framework and the relevant national data-protection laws, produce one of the most operationally complex compliance landscapes any of the major jurisdictions have produced for clinical AI deployment. The AI Act itself does not prescribe specific data residency requirements — the data residency framework operates under the GDPR — but the AI Act's institutional governance obligations require deployers to document the data flows associated with each high-risk AI system and to maintain those flows within compliant data-transfer architectures. For deployments that span multiple EU member states or that involve data transfers to third countries, the documentation burden is non-trivial.

Italy's strict data residency posture, articulated by AIFA's April 2026 guidance, is the most operationally consequential of the national positions. The AIFA framework requires processing of Italian patient data to occur within EU member-state jurisdictions and adds additional safeguards for any third-country transfers beyond what the standard GDPR framework requires. Vendors operating Italian deployments have responded in three ways. Microsoft Nuance has committed to dedicated Italian processing infrastructure for any deployment touching Italian patient data, structured around its Azure Italy North region. Aidoc has chosen a different approach, partnering with Italian data centre operator Aruba to maintain Italian-resident processing infrastructure under a contractual structure that AIFA has endorsed. Glass Health has committed to similar Italian-resident infrastructure but has not yet finalised the operational architecture. Smaller vendors, including several US-based clinical AI startups, have indicated that they will withdraw from the Italian market for the early enforcement period rather than build the Italian-resident infrastructure required to comply with AIFA's posture. The vendor-withdrawal pattern is the most operationally consequential signal: it suggests that AIFA's strict posture is creating market-exit decisions among the smaller vendors and is consolidating Italian deployments around the larger, better-capitalised vendor cohort.

The Germany-Switzerland data flow question is a second consequential dimension. Switzerland is not a member state of the EU but maintains an adequacy decision under GDPR allowing data flows between Switzerland and EU member states without additional safeguards. The Switzerland-EU adequacy decision is bilateral and is reviewed periodically by the European Commission. Several Swiss-resident clinical AI vendors maintain operational architectures that depend on cross-border data flows between Switzerland and Germany, with Germany being the largest single EU market for clinical AI deployment. The bilateral framework currently in force supports those data flows under specific contractual safeguards that Swissmedic and the German competent authority BfArM have jointly endorsed. The framework's stability through the enforcement period is essential to Swiss vendors' commercial operations in the EU market, and any disruption to the bilateral framework would create material consequences for the Swiss clinical AI sector. The framework is presently stable but is subject to broader Swiss-EU political dynamics that are not under the competent authorities' control.

The hospital-level governance obligations under the Regulation are the deployer-side obligations that the AI Act extends beyond the MDR framework. Each deployer of a high-risk clinical AI system is required to maintain a designated AI Officer or equivalent role with formal responsibility for compliance with the Regulation; documented training records demonstrating clinical staff training on appropriate use of the system; institutional governance documentation covering the eight elements BfArM's Leitfaden specifies; incident reporting procedures with documented escalation paths to the relevant competent authority; and continuous monitoring documentation demonstrating ongoing tracking of system performance against pre-specified thresholds. These hospital-level obligations are the structural innovation of the AI Act and are the principal new compliance burden the framework introduces relative to the prior MDR-only regulatory landscape. For the three European university hospitals operating under Cleveland's ACIS framework — Charité, Karolinska, and AKH Vienna — the hospital-level obligations are substantively satisfied by their ACIS implementation. For hospitals without comparable governance infrastructure, the obligations represent a material build-out, and the institutional capacity to satisfy them is presently uneven across the European hospital landscape. The first six to twelve months of enforcement will likely produce a clearer picture of which hospitals can comply and which face structural compliance gaps.

The US-EU posture gap and its consequences

The structural gap between the US FDA's posture on clinical AI and the EU's posture under the AI Act has widened materially over the eighteen-month implementation window. The FDA Digital Health Center of Excellence under Director Troy Tazbaz has maintained a deliberately incrementalist posture: case-by-case Pre-Submission engagement, the published pre-determined change control plan guidance, and the recent favourable Pre-Sub feedback letter to Aidoc on multi-modal triage. The FDA's posture is principled, consistent with the agency's broader regulatory philosophy, and structurally cautious. The EU's posture is structurally interventionist: the AI Act applies a comprehensive regulatory framework to high-risk clinical AI systems, extends institutional governance obligations to deployers as well as manufacturers, and creates competent-authority-specific enforcement architecture that varies across the EU member states. The two postures are not necessarily inconsistent — both are committed to the principle that high-risk clinical AI requires structured regulatory oversight — but they create different operating environments for vendors and deployers active in both jurisdictions.

For vendors, the consequences are operationally meaningful. A US-headquartered clinical AI vendor seeking to operate in both jurisdictions must maintain compliance infrastructure that satisfies FDA Pre-Submission requirements and produces 510(k) or De Novo clearance documentation in the US, and that simultaneously satisfies the AI Act conformity-assessment requirements and the relevant national competent authority's institutional governance obligations in the EU. The documentation overlaps substantially but not entirely, and the compliance overhead of maintaining dual-jurisdiction infrastructure is material. Several large US-based vendors — Microsoft Nuance, Glass Health, Abridge, and Suki AI most prominently — have absorbed this overhead and operate in both jurisdictions. Smaller US vendors have increasingly chosen to operate in one jurisdiction or the other rather than both, with the choice typically influenced by where their early-stage customer base is concentrated. The US-only operating posture is more common among smaller vendors and reflects the structural advantage of the US single-jurisdiction regulatory environment for early-stage commercial development. The EU-only operating posture is rarer and is typically chosen by European-headquartered vendors whose strategic positioning favours the EU regulatory framework's perceived legitimacy advantages in the European clinical market.

For deployers, the consequences are different but equally consequential. US academic medical centres operate under FDA regulation alone, and the institutional governance choices they make — whether CARS-aligned, ACIS-aligned, or hybrid — are not regulatory obligations. They are operational choices designed to produce the evidence that FDA Pre-Submission engagement will require, but they are not compelled by FDA rule. European university hospitals operate under regulatory obligations that compel a specific architecture of institutional governance, with the AI Act's deployer-side obligations producing requirements that hospitals must satisfy regardless of which clinical AI vendor they choose. The structural effect is that European hospitals are required to build governance infrastructure that US hospitals build only voluntarily. This creates a transatlantic capacity differential: European hospitals are, on average, materially further along in formalised clinical AI governance than US peer hospitals at comparable operational scale. The transatlantic capacity differential will, over the next eighteen to twenty-four months, likely produce structural advantages for European hospitals in international research collaboration on clinical AI outcome research and in cross-border vendor partnership negotiations. The advantage will not be uniform — US institutions like Mayo, Cleveland, Geisinger, and Stanford have built sophisticated governance infrastructure voluntarily — but the average European hospital will be governance-mature relative to the average US hospital, and that average matters for the broader institutional landscape.

The US-EU gap will not narrow quickly. The FDA's posture is unlikely to shift toward a more interventionist framework absent a structural clinical AI failure that would create the political conditions for legislative action. The European posture is locked in by the Regulation's text and by the institutional capacity the competent authorities have built to administer it. The two frameworks will coexist in their current postures for the foreseeable future, and the cross-jurisdictional operational complexity will remain a structural feature of the global clinical AI vendor and deployer landscape. The implications for procurement decisions, framework adoption, and institutional governance design are substantial and continuing. Vendors and deployers operating in both jurisdictions must build dual-compliance architectures. Those operating in one or the other will optimise their architecture for the regulatory environment they are in. The bifurcation is structural and is here to stay.

What to watch

The 2 August 2026 enforcement date is six weeks away as of the analytical cut-off for this report. The first twelve months of enforcement will determine whether the AI Act's regulatory architecture produces the institutional governance outcomes its drafters intended, whether the national competent authority posture variations stabilise or produce demands for harmonisation, and whether the US-EU gap creates structural disruptions to the cross-Atlantic clinical AI market. Five signals are the leading indicators.

• Whether the early enforcement period — the six-to-twelve-month "proportionate flexibility" window that BfArM and the other competent authorities have signalled — produces actual enforcement actions or remains a regime of soft-touch warnings; the first formal enforcement action by any of the four principal competent authorities will be the leading indicator of how serious the enforcement phase will be in operational practice.
• Whether the AIFA strict data residency posture is sustained or moderated as smaller vendors withdraw from the Italian market and the consolidation effect on larger vendors becomes visible; if the consolidation effect produces meaningful competitive concerns at the Italian hospital procurement level, AIFA may face political pressure to moderate the posture, and the question of whether it does will be visible by Q1 2027.
• Whether the HAS clinical-evidence requirement for French deployments is challenged before the AI Office or the European Court of Justice by vendors operating French deployments; the legal interpretation of whether institution-specific clinical validation is required by the Regulation's Article 9 risk management system obligations or is a French-only extension is the most consequential interpretive question across the four jurisdictions, and the resolution will shape the broader interpretation of how aggressive national competent authorities can be in their guidance.
• Whether the European hospital landscape's capacity to satisfy the deployer-side institutional governance obligations proves sufficient over the early enforcement period; the three institutions operating under ACIS governance (Charité, Karolinska, AKH Vienna) are positioned to comply, but the larger European hospital landscape — with several hundred high-acuity hospitals across the twenty-seven member states — has variable institutional capacity, and the first six months of enforcement will produce a clearer picture of which hospitals can comply and which face structural gaps.
• Whether US-EU regulatory cooperation produces a substantive Trans-Atlantic Clinical AI Framework or remains a series of bilateral conversations between the FDA Digital Health Center of Excellence and the European Commission; the published statements from both sides have signalled continued cooperation but have not yet produced a structural framework agreement, and whether the cooperation deepens into a formal agreement over 2026 or stays at the conversational level will materially shape the cross-jurisdictional vendor and deployer operating landscape.

Frequently asked

When does the EU AI Act enforcement phase for high-risk clinical AI systems actually begin?

The enforcement phase begins on 2 August 2026, twenty-four months after the Regulation's entry into force on 1 August 2024. On that date, the conformity-assessment requirements, the institutional governance obligations, the data residency rules, and the cross-border deployment provisions of Title III, Chapter 2 of Regulation (EU) 2024/1689 all become enforceable across the twenty-seven EU member states. The national competent authorities — BfArM, AIFA, HAS, AEMPS, and others — administer enforcement within their respective jurisdictions. Swissmedic operates as an observer authority under Switzerland's bilateral agreement with the EU. The early enforcement period is widely understood to involve "proportionate flexibility" for technical or procedural omissions in good-faith compliance efforts, with the period generally expected to last between six and twelve months from the enforcement date.

How do the postures of BfArM, AIFA, HAS, and AEMPS actually differ?

BfArM has been the most prescriptive on institutional governance obligations, with the Leitfaden für Hochrisiko-KI-Systeme im Gesundheitswesen establishing eight specific elements deployers must document. AIFA has been the most aggressive on cross-border data flow provisions, requiring Italian patient data processing to occur within EU member-state jurisdictions with additional safeguards beyond standard GDPR. HAS has been the most demanding on clinical evidence, requiring institution-specific clinical validation evidence in addition to the general validation evidence vendors produce for CE marking. AEMPS has been the most permissive, interpreting the Regulation's requirements at the floor rather than at any extended interpretation. The four postures are all defensible under the Regulation's text but produce materially different compliance environments across the four jurisdictions, with consequences for vendor strategy and hospital procurement decisions.

What are the hospital-level (deployer-side) governance obligations that the Regulation introduces?

Each deployer of a high-risk clinical AI system must maintain: a designated AI Officer or equivalent role with formal responsibility for Regulation compliance; documented training records for all clinical staff interacting with the system; institutional governance documentation covering eight specific elements (per BfArM's Leitfaden, the most detailed national guidance); incident reporting procedures with documented escalation paths to the relevant competent authority; and continuous monitoring documentation demonstrating ongoing tracking of system performance against pre-specified thresholds. These obligations are the structural innovation of the AI Act and are the principal new compliance burden the framework introduces relative to the prior MDR-only regulatory landscape. They apply to the hospital as deployer, not just to the vendor as manufacturer, which extends regulatory responsibility into the institutional environment in ways the MDR did not.

How is the US FDA's posture different from the EU's AI Act posture?

The FDA Digital Health Center of Excellence has maintained a deliberately incrementalist posture: case-by-case Pre-Submission engagement, published pre-determined change control plan guidance, and structured 510(k)/De Novo clearance pathways. The FDA's posture extends regulatory authority to the vendor as manufacturer but does not extend specific governance obligations to the hospital as deployer; US institutional governance choices like CARS or ACIS are operational, not regulatory. The EU's AI Act extends regulatory authority to both vendor and deployer, applies a comprehensive framework to high-risk clinical AI systems, and creates competent-authority-specific enforcement architecture. The two postures are not inconsistent in principle but create materially different operating environments for vendors and deployers active in both jurisdictions, with substantive consequences for compliance infrastructure, procurement decisions, and institutional governance design.

Which clinical AI vendors are best positioned for the 2 August 2026 enforcement date?

The large established vendors — Microsoft Nuance, Siemens Healthineers, Philips Healthcare, and GE HealthCare — have substantive compliance infrastructure already in place from their MDR compliance programmes and have signalled operational readiness. Aidoc, Glass Health, Abridge, and Suki AI are smaller but well-capitalised vendors with sophisticated regulatory affairs functions and similar readiness signals. The 79 smaller vendors holding the remainder of the 215 high-risk CE-marked medical devices are a more concerning category; industry surveys suggest approximately 30-35 per cent will not be operationally ready and will face material compliance risk during the early enforcement period. The competent authorities have signalled "proportionate flexibility" for good-faith compliance efforts during the first six to twelve months, but the smaller vendor cohort faces structural disadvantages relative to the larger established players.

Will the US-EU regulatory gap narrow over time, or is it structural?

The gap is structural and unlikely to narrow quickly. The FDA's posture reflects principled US regulatory philosophy and is unlikely to shift toward a more interventionist framework absent a structural clinical AI failure that would create political conditions for legislative action. The European posture is locked in by the Regulation's text and the institutional capacity the competent authorities have built to administer it. The two frameworks will coexist for the foreseeable future, and the cross-jurisdictional operational complexity will remain a structural feature of the global clinical AI landscape. Vendors and deployers operating in both jurisdictions will continue to build dual-compliance architectures. Those operating in one or the other will optimise for that regulatory environment. The bifurcation is here to stay, and the strategic question for vendors and deployers is whether to operate dual-jurisdiction or to focus on one. There is no general right answer; the choice depends on commercial positioning and operational scale.