The EU Artificial Intelligence Act came into full application in stages beginning August 2024. The prohibition on unacceptable-risk AI practices applied from February 2025. The obligations for high-risk AI systems under Annex III, including the conformity assessment requirements, quality management system obligations, and post-market monitoring requirements, became binding for most affected deployers in August 2026. This briefing is designed to give board members and senior executives a practical understanding of what high-risk classification means, what it requires, and what governance structure it demands.
Key findings
- The high-risk classification covers a broader range of AI applications than most organizations have assessed. Annex III of the Act covers eight domains: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and justice. Within each domain, the scope is specific and requires application-level analysis, not sector-level assessment.
- The conformity assessment for Annex III systems requires documented board authorization for most deployers. The designated responsible person for a high-risk AI system must have organizational authority to withdraw the system from use. For systems that are material to core operations, this authority appropriately sits at board or board-committee level.
- The quality management system requirements under Article 17 are operationally demanding. They require documented risk management, data governance, technical documentation, transparency measures, human oversight mechanisms, accuracy and robustness standards, and cybersecurity measures. Each has specific documentation requirements.
- Most organizations have not conducted a systematic Annex III assessment of their current AI deployments. The assessment is not optional. Deploying an Annex III system without completing the conformity assessment is an infringement subject to administrative fines.
Understanding the high-risk classification
The EU AI Act establishes a risk-based classification framework. At the top are prohibited practices: AI applications that the Act prohibits entirely, including social scoring by public authorities, real-time remote biometric identification in public spaces in most circumstances, and AI systems that exploit vulnerabilities. Below prohibited practices are high-risk AI systems, which are permitted but subject to significant obligations before and during deployment. Below high-risk are limited-risk systems with transparency obligations, and at the base are minimal-risk systems with no specific regulatory requirements.
Annex III defines the high-risk category through eight domains and specific application types within each. The employment domain, which is most commonly encountered in non-specialist organizations, covers AI systems used for recruitment, selection, targeted job advertising, assessment during interviews, and evaluation of performance and behavior of persons in work-related contractual relationships. An AI system used to screen CVs, rank candidates, or assess employee performance that affects decisions about employment, promotion, or termination is high-risk under the Act.
The essential services domain
The essential services domain is the most broadly applicable to financial institutions, insurance companies, and public service organizations. It covers AI systems used in the evaluation of creditworthiness and credit scoring of natural persons, with the exception of systems used to detect fraud. It also covers risk assessment and pricing for life and health insurance, and AI used in the dispatch of emergency services including police, ambulance, and fire services.
For banks and insurance companies, this means that AI-assisted credit scoring, underwriting, and pricing systems for natural persons are likely to be high-risk under the Act and require conformity assessment before deployment or continued use. Organizations that have deployed these systems without conducting an Annex III assessment should treat this as a priority compliance gap.
The question is not whether an organization uses AI. It is whether the AI systems it uses are making or influencing decisions about people's access to credit, employment, essential services, or legal rights. If they are, the Act is not optional.
KIG Field Intelligence, Briefing, November 2025What conformity assessment requires in practice
For Annex III AI systems where the deployer is not also the provider, the deployer's obligations under Article 26 include: verifying that the system has an EU declaration of conformity from the provider; ensuring the system is used in accordance with the instructions for use; monitoring the system's performance against the intended purpose; informing the provider of any serious incidents; and maintaining a post-market monitoring system.
The most demanding deployer obligation is the human oversight requirement. Article 14 requires that high-risk AI systems be designed and deployed so that natural persons can effectively oversee the system during use, including the ability to override or stop the system. This is not a passive requirement. It requires documented procedures for how human oversight is exercised, training for the persons responsible for oversight, and a record of oversight activities.
For organizations that have deployed AI systems that generate recommendations for decisions affecting employment, credit, or essential services, and where the workflow effectively treats the AI output as the decision without meaningful human review, the human oversight requirement is a substantive compliance gap. The fix requires redesigning the workflow, not only documenting it.
The board governance structure the Act requires
The EU AI Act does not prescribe a specific board governance structure. It prescribes outcomes: high-risk AI systems must be governed in a way that allows meaningful human oversight, that ensures the system operates within its approved intended purpose, and that produces timely responses to serious incidents. For most organizations, achieving these outcomes requires governance structures at board or board-committee level.
The minimum governance architecture has four elements. First, an AI system inventory classified by the Act's risk tiers, maintained at a level of specificity sufficient to determine whether each system's current deployment is compliant with its applicable obligations. Second, a designated responsible person for each high-risk AI system, with documented authority and a clear escalation path to board level. Third, a board-level review at minimum annually of the organization's AI risk exposure, including any material changes to the inventory, any serious incidents, and any changes to the regulatory environment. Fourth, a documented decision authority specifying which governance body must authorize the deployment, material modification, or withdrawal of high-risk AI systems.
Organizations that build this governance structure before they need to demonstrate it to their national supervisory authority will find the required documentation straightforward to produce. Those that attempt to reconstruct the governance record after an inquiry has begun will find it significantly more difficult. The Act's administrative fine structure, which mirrors GDPR in its severity, makes the cost of deferred compliance substantially higher than the cost of timely preparation.