The European Commission’s May 2026 draft guidelines on high-risk AI classification (AI Act Article 6 and Annex III) arrive too late for most deployers. While the document clarifies ambiguities in the original text, it also exposes a critical implementation gap: organizations have spent two years debating classification thresholds but have not operationalized the underlying governance, risk assessment, and documentation requirements. With the final guidelines due in June 2026—and enforcement of h
The European Commission’s May 2026 draft guidelines on high-risk AI classification (AI Act Article 6 and Annex III) arrive too late for most deployers. While the document clarifies ambiguities in the original text, it also exposes a critical implementation gap: organizations have spent two years debating classification thresholds but have not operationalized the underlying governance, risk assessment, and documentation requirements. With the final guidelines due in June 2026—and enforcement of high-risk obligations beginning in August 2026—deployers now face a compressed timeline to align their AI systems with regulatory expectations. This article maps the specific compliance obligations that must be addressed before the guidelines are finalized.
---
What Changed in the May 2026 Draft Guidelines on High-Risk Classification
The May 2026 draft guidelines (published as a Commission Notice) provide interpretative clarity on three key areas of the AI Act’s high-risk classification framework:
1. **Annex III Scope Expansion**
The guidelines confirm that Annex III’s eight high-risk domains (e.g., biometrics, critical infrastructure, employment, law enforcement) are not exhaustive. Deployers must assess whether their AI system poses a "significant risk of harm to health, safety, or fundamental rights" (AI Act Article 6(3)), even if it falls outside Annex III. The draft introduces a three-step test to determine this:
- Step 1: Sectoral relevance – Does the system operate in an Annex III domain?
- Step 2: Use-case specificity – Does the system influence decisions with legal or similarly significant effects (e.g., access to essential services, employment outcomes)?
- Step 3: Risk magnitude – Does the system create a "high probability of severe harm" (e.g., discriminatory hiring tools, biased credit scoring)?
The guidelines emphasize that deployers cannot rely solely on Annex III—they must conduct case-by-case assessments.
2. **Clarification on "Intended Purpose"**
The draft guidelines resolve long-standing ambiguity around the "intended purpose" criterion (AI Act Article 3(12)). Deployers must now document:
- The primary objective of the AI system (e.g., fraud detection, resume screening).
- Secondary or unintended uses (e.g., a chatbot repurposed for mental health advice).
- User instructions and training materials that define the system’s scope.
The Commission warns that deployers cannot evade classification by narrowing the stated purpose—regulators will assess the *actual* use of the system, not just its marketing claims.
3. **Proportionality and Exemptions**
The guidelines introduce a proportionality principle for systems that *could* be high-risk but pose minimal harm in practice. For example:
- A resume-screening tool used for low-risk roles (e.g., internships) may qualify for an exemption if it does not significantly influence hiring outcomes.
- A biometric identification system used for non-sensitive access control (e.g., office entry) may avoid classification if it does not process sensitive data.
However, the draft makes clear that exemptions are narrow and require documented risk assessments to justify them.
---
The Deployer Compliance Trap: Why Classification Guidance Alone Isn’t Enough
The May 2026 guidelines are a double-edged sword. While they provide much-needed clarity on *what* constitutes a high-risk system, they also reveal that most deployers have focused on classification at the expense of operational compliance. Three misconceptions are driving this gap:
1. **"We’ll Wait for the Final Guidelines"**
Many organizations assume they can defer action until the June 2026 finalization. This is a critical error. The AI Act’s high-risk obligations (Articles 8-15) apply from August 2026, regardless of whether the guidelines are final. Deployers must begin compliance work *now*—or risk enforcement actions for non-conformity.
2. **"Our Vendor Handles Compliance"**
Deployers often assume that AI providers (e.g., SaaS vendors, model developers) will shoulder compliance burdens. The AI Act explicitly holds deployers accountable for:
- Risk management (Article 9) – Deployers must implement ongoing risk assessments, not just rely on provider certifications.
- Data governance (Article 10) – Deployers must ensure training data meets quality, bias, and representativeness standards.
- Technical documentation (Article 11) – Deployers must maintain records of system performance, incidents, and human oversight.
3. **"We’ll Classify Systems Later"**
Some deployers plan to conduct high-risk assessments *after* the guidelines are final. This approach is legally and operationally flawed:
- Retroactive classification is risky – If a system is later deemed high-risk, deployers may face fines for non-compliance during the interim period.
- Documentation gaps will emerge – Many deployers lack the records needed to prove compliance (e.g., risk assessments, data provenance logs).
- Audit trails are missing – Regulators will expect evidence of *ongoing* compliance, not just post-hoc justifications.
---
Three Operational Gaps Deployers Must Close Before June 2026
The May 2026 draft guidelines expose three critical gaps in deployer preparedness. Addressing these requires immediate operational changes, not just legal or policy adjustments.
1. **Lack of a Centralized AI Inventory**
Most organizations cannot answer:
- How many AI systems are in use across the enterprise?
- Which systems fall under Annex III or meet the Article 6(3) "significant risk" threshold?
- Who owns each system, and what data does it process?
Solution:
- Implement an AI registry – Track all AI systems, their purposes, data inputs, and risk classifications.
- Assign ownership – Designate a responsible party (e.g., CISO, DPO, or AI governance lead) for each system.
- Automate discovery – Use tools to scan for AI models in production (e.g., shadow AI, third-party integrations).
2. **Inadequate Risk Assessment Frameworks**
The AI Act requires continuous risk management (Article 9), but most deployers lack:
- Standardized risk assessment methodologies – Many rely on ad-hoc evaluations or vendor-provided certifications.
- Bias and fairness testing – Few organizations systematically test for discriminatory outcomes.
- Fundamental rights impact assessments (FRIAs) – Required for high-risk systems, but rarely conducted in practice.
Solution:
- Adopt a risk management framework – Align with ISO/IEC 42001, NIST AI RMF, or ENISA’s AI guidelines.
- Conduct FRIAs – Assess impacts on privacy, non-discrimination, and other fundamental rights.
- Implement bias testing – Use tools to detect and mitigate discriminatory outcomes (e.g., disparate impact analysis).
3. **Poor Documentation and Audit Readiness**
The AI Act mandates extensive documentation (Article 11), including:
- Technical specifications – System design, data sources, and performance metrics.
- Incident logs – Records of failures, biases, or safety issues.
- Human oversight procedures – How human reviewers intervene in high-risk decisions.
Most deployers lack these records, making compliance impossible to demonstrate in an audit.
Solution:
- Centralize documentation – Store all compliance records in a secure, auditable system.
- Automate logging – Capture system inputs, outputs, and incidents in real time.
- Conduct mock audits – Test readiness with internal or third-party reviews.
---
Immediate Actions: Risk Assessment, Documentation, and Audit Readiness
Deployers must act *now* to close these gaps before the June 2026 guidelines finalize. Three priority actions:
1. **Conduct a High-Risk AI Classification Audit**
- Inventory all AI systems – Identify those in Annex III domains or with significant risk potential.
- Apply the three-step test – Assess sectoral relevance, use-case specificity, and risk magnitude.
- Document exemptions – Justify why low-risk systems do not require classification.
2. **Implement a Risk Management Framework**
- Adopt ISO/IEC 42001 or NIST AI RMF – These frameworks align with AI Act requirements.
- Conduct FRIAs – Assess impacts on fundamental rights, particularly for systems in employment, credit scoring, or law enforcement.
- Test for bias – Use tools to detect discriminatory outcomes and mitigate them.
3. **Build an AI Compliance Documentation System**
- Centralize records – Store technical specifications, incident logs, and human oversight procedures in a single system.
- Automate logging – Capture system inputs, outputs, and failures in real time.
- Prepare for audits – Conduct mock audits to identify gaps before regulators do.
---
How to Align Your AI Governance Framework with Draft Guidelines Today
The May 2026 draft guidelines provide a roadmap for compliance—deployers should align their governance frameworks with these expectations *now*. Key steps:
1. **Update AI Policies and Procedures**
- Revise AI use policies – Define acceptable use cases and prohibited applications.
- Establish approval workflows – Require sign-off for high-risk AI deployments.
- Define escalation paths – Set rules for reporting incidents or biases.
2. **Enhance Data Governance**
- Audit training data – Ensure datasets are representative, unbiased, and legally sourced.
- Implement data provenance tracking – Document the origin and lineage of all training data.
- Apply data minimization – Limit data collection to what is strictly necessary.
3. **Strengthen Human Oversight**
- Define human-in-the-loop (HITL) procedures – Specify when and how humans must intervene in AI decisions