Why XpertDPO submitted feedback on the EU AI Act high-risk classification guidelines

On 27 May 2026, XpertDPO Limited submitted feedback to the European Commission’s targeted consultation on the draft guidelines for the classification of high-risk AI systems under Article 6 of the EU AI Act.

The consultation is important because classification is the gateway into the AI Act’s high-risk regime. If an AI system is wrongly classified, the consequences are practical and serious: providers may under-document systems, deployers may misunderstand their responsibilities, and people affected by AI-assisted decisions may lose the benefit of the safeguards the law is meant to provide.

We submitted because the draft guidance is useful, but it needs to be more operational for real organisations.

What we commented on

We focused our response on three areas:

• Section II, especially intended purpose and how it should be evidenced in real deployments.
• Section IV, especially Annex III, human involvement, complex workflows, Article 6(3), profiling, and the practical meaning of material influence.
• Article 112, where the Commission is gathering input on whether the high-risk use cases in Annex III should be clarified, amended, or expanded.

We did not comment on every part of the questionnaire. We kept the response focused on the places where XpertDPO sees the greatest practical classification risk for providers, deployers, DPOs, governance teams, procurement teams, and organisations implementing AI-enabled tools.

Our core point

The final guidance should make it easier to recognise when an AI system materially shapes an outcome for a natural person, even where a human remains formally involved.

Many systems will not look dramatic. They may be sold as ordinary workflow tools, case-management features, copilots, dashboards, triage tools, summarisation tools, ranking tools, or recommendation features. But if they affect who is prioritised, flagged, routed, scrutinised, shortlisted, escalated, delayed, or given access to a service, they may have real-world consequences.

Our submission asks the Commission to make those practical indicators clearer.

Intended purpose needs evidence

We asked for more guidance on how intended purpose should be evidenced in real deployments, especially for SaaS products, embedded AI features, workflow tools, and general-purpose AI systems.

A provider may market a tool broadly, while a deployer configures it for a specific decision process. Equally, a tool may be described as “assistive” but then be used inside HR, education, public services, credit, insurance, healthcare, or casework decisions.

Our view is that organisations need clearer examples of the evidence they should keep, including:

• intended-purpose statements;
• unsupported-use limits;
• Annex III mapping;
• affected persons and user groups;
• output type;
• decision context;
• human review design;
• change-control triggers.

This matters because classification should not be a one-off label. It should be revisited when the system’s purpose, users, affected persons, workflow integration, profiling activity, or decision context changes.

Human involvement is not enough on its own

The draft guidelines rightly say that human involvement does not automatically remove a system from high-risk classification.

We supported that position and asked for it to be made more practical. The final guidance should distinguish genuine independent human assessment from rubber-stamping, automation bias, workload pressure, or review that only happens after an AI system has already ranked, filtered, flagged, routed, or prioritised a person or case.

The question should not be only whether a human makes the final formal decision. The question should also be whether the AI output materially shapes the path to that decision.

Article 6(3) should remain a narrow filter

Article 6(3) allows some Annex III systems to be filtered out of high-risk classification where they do not pose a significant risk of harm and meet specific conditions.

We agree that this filter is important for proportionality. But it must be applied carefully.

Our submission asks the Commission to define practical indicators of material influence, such as:

• ranking;
• scoring;
• prioritisation;
• eligibility flags;
• risk labels;
• recommended outcomes;
• default options;
• confidence-weighted suggestions;
• selective presentation of evidence;
• case routing;
• queue position;
• escalation;
• suppression of information.

These features may influence a human outcome even if the system is formally described as advisory.

We asked for more borderline examples

The most useful guidance for organisations will be worked examples.

We suggested examples such as:

• CV summarisation versus candidate ranking;
• benefits case-file indexing versus eligibility recommendation;
• healthcare referral indexing versus care-pathway prioritisation;
• customer vulnerability identification versus service prioritisation or restriction;
• fraud detection on transactions versus profiling or risk-scoring natural persons;
• learning analytics dashboards used only by students versus institutional tools used for progression, grading, support allocation, or monitoring;
• public-sector outsourced triage performed on behalf of an authority;
• AI meeting notes or case summaries used in disciplinary, complaints, HR, or public-service casework;
• general-purpose copilots used to draft, recommend, quality-check, or structure decisions in Annex III contexts;
• quality assurance of completed decisions versus live influence on pending decisions;
• agentic workflows where several low-risk steps combine into high-risk decision support.

These are the sorts of examples organisations will actually need when they are classifying AI systems in procurement, deployment, DPIAs, AI governance reviews, and vendor due diligence.

Education, employment, and essential services

We focused in particular on three Annex III areas: education, employment, and access to essential private and public services.

In education, we asked the Commission to distinguish genuinely student-controlled support tools from systems used by institutions to classify, stream, grade, monitor, allocate support, detect prohibited behaviour, influence progression, or assess special educational needs.

In employment, we supported a broad functional approach. Systems used for targeted job advertising, CV ranking, candidate sourcing, interview scoring, shift allocation, productivity scoring, behavioural monitoring, platform access, promotion, or termination should not be filtered out where they affect opportunity, pay, progression, or continued engagement.

In essential services, we asked the Commission to connect classification more explicitly to practical contestability. In benefits, healthcare, housing, care services, credit, insurance, emergency response, and similar contexts, affected persons may be vulnerable or dependent. The risk is not only outright denial. Delay, lower priority, additional scrutiny, escalation, or reduced effective access can also matter.

Our Article 112 position

We took a stronger position under Article 112.

XpertDPO recommended that the Commission consider adding or expressly clarifying use cases for AI systems that materially influence healthcare access, care-pathway prioritisation, allocation of care resources, waiting-list order, referral urgency, or outsourced public-service triage performed by or on behalf of public authorities.

These systems can determine whether, when, and how a natural person receives essential care or public support. The consequences of error, bias, delay, or de-prioritisation can be severe, particularly for children, older persons, persons with disabilities, patients, benefit claimants, and people with limited capacity to challenge decisions.

If the Commission considers these use cases already captured by existing Annex III areas, our view is that the final guidance should say so expressly and provide worked examples.

Why this matters

The AI Act will only work if organisations can classify systems consistently before harm occurs.

High-risk classification is not just an abstract compliance label. It is the point at which organisations decide what evidence they need, what safeguards apply, what questions to ask providers, and what people affected by AI-assisted decisions can reasonably expect.

That means the final guidance needs to be usable by the people who will apply it in practice: product teams, deployers, DPOs, compliance teams, procurement teams, risk teams, senior leaders, and public bodies. It must help them identify when an ordinary-looking tool is no longer just administrative support, but is materially shaping outcomes for people.

XpertDPO submitted because this is where practical accountability starts: in clear classification, visible reasoning, and early recognition of the people whose rights, access, opportunities, care, or livelihood may be affected.

We will add a link to the published submission here when the Commission makes consultation responses available.

Preparation note

The response was submitted by XpertDPO Limited, directed and reviewed by Philipa Jane Farley, and prepared with AI-assisted drafting support from Eliot/Codex.