夜雨聆风
Key Takeaways
The EU AI Act turns “high-risk” classification into an operational, contractual, and evidentiary issue. Companies cannot treat it as a label to be added at the end of a product launch.
For companies selling AI-enabled software, SaaS, robotics, industrial systems, HR tools, fintech products, education technology, or smart devices into Europe, the practical question is not whether the company is incorporated in the EU. The question is whether the AI system is placed on the EU market, put into service in the EU, or produces outputs used in the EU.
The European Commission’s public consultation on draft guidelines for high-risk AI system classification closed on June 23, 2026. The final guidance remains to be seen, and the broader AI Omnibus discussion may affect some implementation timelines. The core legal framework, however, is already in the AI Act: Article 6, Annex III, and the obligations imposed on providers, deployers, importers, distributors, and parties that materially modify or repurpose AI systems.
Background
Regulation (EU) 2024/1689, known as the EU Artificial Intelligence Act, entered into force on August 1, 2024 and applies in phases. Its central regulatory design is risk-based. Some AI practices are prohibited. High-risk systems are permitted but subject to substantial compliance obligations. Lower-risk systems are subject to lighter transparency or governance requirements.
For many non-EU businesses, the high-risk analysis is the point at which the AI Act becomes a product launch and commercial contracting issue. The classification determines whether the business needs a risk management system, data governance controls, technical documentation, logging capability, transparency information, human oversight, robustness and cybersecurity controls, conformity assessment, registration, and a credible evidence file.
That classification also affects who must do what under the contract. A customer may ask the supplier to provide technical documentation, test records, data descriptions, version-change notices, logs, incident support, and regulatory cooperation. A distributor or systems integrator may want indemnities or clear responsibility allocation. A deployer may need operational instructions and monitoring tools. If the agreement is silent, each party may discover the classification problem only after a procurement questionnaire, tender, incident, or regulator request.
Legal Framework
Article 6 and Annex III are the primary entry points for high-risk classification.
The first category covers AI systems that are products, or safety components of products, subject to listed EU product safety legislation. This may matter for medical devices, machinery, toys, lifts, radio equipment, pressure equipment, personal protective equipment, aviation, vehicles, marine equipment, and other regulated product categories. If an AI module affects a safety function, it may not be enough to describe the feature as an “intelligent assistant.”
The second category covers specific use cases listed in Annex III. These include biometric systems, critical infrastructure, education and vocational training, employment and worker management, access to essential private and public services, law enforcement, migration and border control, and administration of justice and democratic processes. The focus is not the size of the model. The focus is the function performed by the system and the consequences of its output.
Article 6(3) creates a limited carve-out for certain Annex III systems that do not pose a significant risk of harm to health, safety, or fundamental rights and do not materially influence the outcome of decision-making. The provision refers to systems that perform a narrow procedural task, improve the result of a previously completed human activity, detect decision-making patterns or deviations without replacing human review, or perform preparatory tasks.
That carve-out should not be read as a self-certification shortcut. Article 6(4) requires the provider relying on the carve-out to document the assessment before the system is placed on the market or put into service. The provider must provide that documentation to competent national authorities on request. In addition, systems that perform profiling of natural persons remain high-risk.
Why It Matters
The hardest mistakes usually happen before the legal team is asked to review the AI Act position.
A product team may classify the system as ordinary enterprise software because the output is called a recommendation. A business team may assume that the EU customer is responsible for all deployment risks. A sales team may accept a white-label arrangement without asking whether rebranding changes the role allocation. A systems integrator may modify the intended use of the system and accidentally move into the provider role. A customer may use a general analytics tool for employment, credit, education, or access-to-service decisions.
Under Article 25, certain parties may be treated as providers if they place a high-risk AI system on the market or put it into service under their own name or trademark, make a substantial modification, or modify the intended purpose of the system. This is particularly important for reseller, white-label, OEM, systems integration, and customized deployment structures.
Article 26 also matters. Deployers of high-risk systems must use the system in accordance with instructions, assign human oversight to competent persons, monitor operation, keep automatically generated logs where relevant, and take action when risks arise. Some deployers must conduct a fundamental rights impact assessment before use.
In other words, the AI Act is not only a provider statute. It is a value-chain statute. The commercial documents must match that reality.
Contract, Evidence, and Governance Implications
A company preparing to sell or deploy AI systems in Europe should not wait for a final procurement questionnaire to decide whether the system is high-risk. The analysis should begin at product scoping and continue through contracting, launch, and post-deployment monitoring.
In supplier agreements, the buyer should request enough information to classify the system: intended purpose, covered use cases, model architecture at an appropriate level, data governance approach, validation and testing methodology, logging capability, human oversight design, cybersecurity controls, and known restrictions. The supplier should resist open-ended obligations that exceed its actual control, but it should be ready to provide a structured AI Act documentation package.
In customer agreements, the provider should define the intended purpose, prohibited uses, configuration boundaries, customer responsibilities, human review requirements, logging responsibilities, incident escalation, regulatory cooperation, and consequences of unauthorized modification or repurposing. A standard disclaimer that “the customer is responsible for use” is usually not enough if the commercial facts show that the provider designed the system for a high-risk use case.
In reseller and white-label agreements, the parties should address branding, modification rights, training or fine-tuning, customer-facing claims, documentation delivery, regulatory inquiries, post-market monitoring, and who can approve a change in intended use. If the local partner changes the product substantially or sells it under its own brand, the provider analysis must be revisited.
In deployment governance, the business should preserve the documents that will matter later: classification worksheets, risk assessments, test reports, technical documentation, user instructions, data governance records, human oversight arrangements, logs, change approvals, and meeting records showing why a particular classification decision was made.
Practical Checklist
1. Map the use case before launch. Identify whether the AI system falls within Annex III or a regulated product safety framework, whether it affects natural persons, and whether it performs profiling.
2. Identify the legal role. Determine who is the provider, deployer, importer, distributor, reseller, systems integrator, or white-label operator. Reassess the role if the system is rebranded, substantially modified, or used for a new purpose.
3. Document the classification decision. If the business relies on the Article 6(3) carve-out, create a written assessment before market placement or deployment and keep it ready for competent authority review.
4. Build contract clauses around evidence. Include technical documentation, logging, version updates, incident notification, regulatory cooperation, prohibited uses, human oversight, data governance, and modification controls.
5. Preserve deployment evidence. Keep records of testing, instructions, human oversight, user training, logs, complaints, incidents, and change approvals.
6. Reassess after changes. A system that starts as a low-risk assistant may become high-risk if a customer uses it to rank candidates, price credit, evaluate students, assess workers, control safety functions, or profile natural persons.
Open Issues
The Commission’s final high-risk classification guidance is still pending after the June 23, 2026 consultation deadline. Companies should monitor how the final guidance interprets Article 6(3), Annex III examples, and profiling.
The AI Omnibus process may also affect certain implementation dates or transitional arrangements. Until any change is enacted, companies should treat the AI Act text as the controlling legal baseline and describe timing issues as an open legislative development rather than a settled exemption.
Member State enforcement practice, harmonized standards, conformity assessment capacity, and market surveillance expectations will also shape how the regime operates in practice.
Official Sources
European Union: Regulation (EU) 2024/1689, July 12, 2024
https://eur-lex.europa.eu/eli/reg/2024/1689/oj
European Commission: AI regulatory framework / AI Act implementation materials, accessed June 24, 2026
https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
ITPro: European Commission opens public consultation on long-awaited draft for high-risk AI guidelines, May 20, 2026
https://www.itpro.com/business/policy-and-legislation/european-commission-opens-public-consultation-on-long-awaited-draft-for-high-risk-ai-guidelines
────────────────────────────────────────────
作者简介/ About the Author
Frank Qin (秦国梁) has extensive experience at leading international law firms, top-tier PRC law firms and Goldman Sachs, specializing in dispute resolution, intellectual property, data compliance, commercial negotiations, and corporate overseas expansion. This article is for general legal and business information sharing purposes only and does not constitute legal advice on specific matters.
