The 8-Requirement Compliance Checklist for High-Risk AI
For systems that are classified as high-risk but not prohibited, the Act specifies eight categories of requirements that must be satisfied before deployment. Here is what each means in practice.
1. Risk Management System
A documented, ongoing risk management process covering the entire lifecycle of the AI system — from initial design through post-market monitoring. This is not a one-time assessment. The Act requires that the risk management process be continuously updated as new risks emerge or the system's behavior changes with retraining. You need a living process, not a static document.
2. Data and Data Governance
Training, validation, and testing datasets must satisfy specific quality requirements: they must be relevant, representative, and free of errors and biases to the extent possible given intended use. You must document the origin of each dataset, the preprocessing applied, any known limitations, and any data quality checks performed. Retroactively documenting data provenance on a system built before compliance was a consideration is significantly harder than building with documentation in mind from the start.
3. Technical Documentation
Before a high-risk AI system can be placed on the EU market, providers must complete detailed technical documentation covering: system description and intended purpose; design specifications; development methodology; training approach; performance metrics and known limitations; instructions for use; cybersecurity measures; and information required for downstream deployers to understand the system's behavior. This documentation must be maintained and updated whenever the system changes materially — think of it as the technical equivalent of a drug's prescribing information.
4. Record-Keeping and Automatic Logging
High-risk AI systems must be capable of automatically logging their inputs, outputs, and the circumstances under which decisions were made. The Act specifies that logging capabilities must be enabled by default and must cover events relevant to identifying risks. For systems handling EU residents, these logs must be retained for a minimum period and must be accessible to authorities upon request. If your current system does not have structured decision logging with retention policies, this is an engineering work item that belongs on your roadmap today.
Deployers — organizations that put the system to use — must be provided with clear instructions covering the intended purpose, performance levels, limitations, human oversight requirements, and any conditions under which the system should not be used. If you are building AI systems that other organizations will deploy (APIs, SaaS platforms, enterprise software), your documentation obligations extend to making compliance feasible for your downstream customers.
6. Human Oversight
Perhaps the most operationally complex requirement: high-risk AI systems must be designed to allow human oversight throughout operation. Specifically, systems must enable humans to understand the system's outputs and decision rationale, monitor for anomalies, override or overrule the system's outputs, and shut the system down. Building genuine override capability — not just a nominal “human in the loop” toggle — requires rethinking the UX and workflow design of many AI-integrated products.
7. Accuracy, Robustness, and Cybersecurity
High-risk AI systems must achieve appropriate levels of accuracy for their intended purpose, be robust against errors and adversarial inputs, and implement cybersecurity measures proportionate to the risks. The Act does not specify numeric accuracy thresholds — providers must define appropriate benchmarks for their use case, document those benchmarks, and demonstrate the system meets them. Adversarial testing — including prompt injection testing for LLM-based systems — is expected as part of the robustness requirement.
8. Quality Management System
Providers must implement a quality management system covering all aspects of the system's lifecycle from design through monitoring. This includes documented procedures, responsibilities, configuration management, and post-market surveillance plans. For software companies not accustomed to ISO-style quality management systems, this requirement often requires the most organizational change.
Conformity Assessment: The Gate Before Market Entry
Before a high-risk AI system can be deployed in the EU, the provider must complete a conformity assessment demonstrating that the system meets all eight requirements. For most Annex III systems, this is a self-assessment — the provider conducts the assessment internally, documents it, draws up an EU Declaration of Conformity, affixes the CE marking, and registers the system in the EU AI Act's central database.
There are two exceptions where third-party assessment by a notified body is required: remote biometric identification systems, and AI systems that are safety components of products already subject to CE marking under existing product safety legislation.
The conformity assessment is not a box-checking exercise. If a regulator investigates a compliance incident, the conformity assessment documentation is the primary evidence base. Organizations that conducted perfunctory assessments to meet the deadline without genuine engineering remediation will face the same enforcement exposure as organizations that skipped compliance entirely — the Act assesses compliance against technical requirements, not the quality of paperwork.
The Agentic AI Compliance Problem
Autonomous AI agents — systems that take multi-step actions, use tools, and operate with minimal human intervention — are the category that creates the most AI Act compliance complexity in 2026, precisely because they are also the fastest-growing enterprise deployment pattern.
Most enterprise AI agents deployed in HR, financial services, or healthcare contexts qualify as high-risk systems under Annex III. But their compliance challenges go beyond meeting the eight standard requirements.
Decision traceability: Agents that chain multiple model calls, tool uses, and data lookups to reach conclusions create audit trail challenges that traditional AI systems do not. Logging “the agent approved the loan application” is not sufficient — you need to trace which inputs, retrieved documents, and reasoning steps produced that output. This requires structured observability from the ground up, not a post-hoc logging layer.
Human oversight at scale: An agent completing 10,000 actions daily cannot be supervised by a human at each step. The Act does not require per-action human review — it requires that humans can intervene, override, and stop the system. Building architectures with meaningful override points while maintaining agent efficiency is a non-trivial engineering challenge. The OWASP Agentic Top 10 framework (published in early 2026) provides useful architecture guidance that maps directly to AI Act Article 14 requirements for human oversight.
Third-party model governance: Most agents use foundation models from OpenAI, Anthropic, or Google as components. The Act places primary compliance responsibility on the entity that deploys the system in the EU — the fact that the underlying model is provided by a third party does not transfer compliance liability. You need documentation about the foundation model's capabilities and limitations, and your conformity assessment must account for behaviors that emerge from your specific integration, not just the base model's published specifications.
Multi-agent systems: When agents delegate tasks to other agents — an orchestrator spawning subagents — compliance responsibility follows the system boundary, not the individual component. If the orchestrated system as a whole qualifies as high-risk, every component in the chain must meet the relevant documentation and logging requirements.
The Digital Omnibus Wildcard
There is one significant regulatory wildcard worth tracking: the European Commission proposed a “Digital Omnibus” package in late 2025 that includes provisions that could push the high-risk AI enforcement date from August 2, 2026 to December 31, 2027 for Annex III systems. As of April 2026, this proposal has not been finalized or adopted into law. The European Commission also missed its own deadline for publishing AI Act guidance on high-risk systems, creating additional uncertainty.
Prudent compliance planning assumes August 2, 2026 is the binding deadline. Organizations that bet on a postponement and lose will face the full enforcement regime with no preparation. If the Digital Omnibus extension does pass, it will be a grace period — not a permanent reprieve. The requirements will still apply in December 2027.
What to Do in the Next 100 Days
If your organization has not begun formal AI Act compliance work, here is the minimum viable sequence:
- Inventory: List every AI system your organization develops or deploys that touches EU users. Include third-party AI integrations in your products, not just systems you built from scratch.
- Classify: Determine which systems fall under Annex III high-risk categories. If any system is borderline, get legal counsel — the cost of a misclassification is the difference between self-assessment and enforcement action.
- Gap assessment: For each high-risk system, assess the gap between current state and each of the eight requirements. Most teams will find data governance documentation and automatic decision logging as the largest gaps.
- Remediation roadmap: Build a prioritized implementation plan. Human oversight architecture and technical documentation are typically the longest-lead items and should be started immediately.
- Conformity assessment: Begin the formal conformity assessment process, targeting completion by July 1 to leave buffer before the August 2 deadline.
- Register: Complete EU database registration for high-risk systems before August 2. Late registration is itself a compliance violation.
The organizations that will struggle most are those waiting for absolute regulatory certainty before beginning — in AI regulation as in engineering, waiting for perfect information is a strategy for missing deadlines. The 100-day window is tight but executable. More importantly, organizations that complete this sequence will be in a stronger commercial position: AI Act compliance is increasingly a procurement requirement in European enterprise deals, and demonstrating certified compliance before competitors will become a meaningful differentiator in regulated verticals over the next 18 months.
Comments · 0
Beta: comments are stored locally on your device and not visible to other readers.
No comments yet. Be the first to share your thoughts.