← Back to Insights
November 18, 2026 · 6 min read · AI Act
AI Act High-Risk: Is Your AI System in Scope?
The EU AI Act classifies AI systems into risk levels. Most companies assume they're "low risk" or "no risk" and move on. Some are wrong. If your AI system is high-risk, you need governance architecture in place before deployment — not retrofitted after. Here's how to tell which category you're in.
The four risk levels
- Unacceptable risk — banned entirely. Social scoring, real-time biometric surveillance in public spaces. You can't deploy these.
- High-risk — allowed, but heavily regulated. Must meet requirements before deployment. This is where most compliance work is.
- Limited risk — transparency obligations. Users must know they're interacting with AI. Content labelling requirements.
- Minimal/no risk — no specific obligations. Most AI systems fall here. But you need to assess and document why.
What makes a system high-risk
The AI Act lists high-risk systems in Annex III. The categories that catch most companies by surprise:
- AI used in employment decisions — CV screening, candidate evaluation, performance assessment. If your HR tool uses AI to rank candidates, it's high-risk.
- AI used in education — admissions, exam scoring, learning assessment. If your EdTech platform uses AI to evaluate students, it's high-risk.
- AI used in essential services — credit scoring, insurance pricing, healthcare triage. If your AI affects access to finance, insurance, or health, it's high-risk.
- AI used in law enforcement — lie detection, evidence evaluation, criminal assessment. If your tool supports policing or legal decisions, it's high-risk.
- AI used as safety component — if your AI is part of a product covered by other EU safety law (machinery, toys, medical devices), it's high-risk.
The trap is "we don't think we're high-risk." The classification isn't about whether your AI is dangerous. It's about whether it's used in a regulated domain. A harmless AI that scores CVs is high-risk. A dangerous AI that generates marketing copy is minimal risk. The use case determines the level, not the technology.
What high-risk requires
If your system is high-risk, you must implement before deployment:
- Risk management system — documented risk assessment for the AI system, not just the company
- Data governance — training data quality, bias testing, data provenance
- Technical documentation — how the model works, what it was trained on, how decisions are made
- Record-keeping — logs of decisions, automated where applicable
- Transparency — users must know they're interacting with AI
- Human oversight — a human must be able to override or intervene
- Accuracy, robustness, and cybersecurity — tested and documented
Why this is an architecture problem
These requirements aren't compliance checkboxes. They're architecture decisions. Risk management means your system design must include risk controls. Data governance means your data pipeline must include quality checks. Human oversight means your system must allow intervention — not after deployment as a patch, but in the original design.
Retrofitting governance onto an AI system that wasn't designed for it is expensive, unreliable, and sometimes impossible. If the training data doesn't have provenance, you can't add provenance retroactively. If the model doesn't log decisions, you can't add logging without retraining. Governance architecture must be built in.
When to assess
Before deployment. Not after. The AI Act is clear: high-risk systems must meet requirements before they're placed on the market. "We'll fix it in version 2" is not a compliance strategy. And regulators can pull non-compliant systems off the market.
Questions about AI Act classification?
Get in touch. 15-minute call, no obligation.
Get in touch
Aethel Guard is a Netherlands-based security architecture firm serving EU companies of all sizes. Your architect, not your vendor.