A patient-triage system built for hospitals outside Europe can come within the EU AI Act’s scope when a European organization starts using its output. Internal AI tools can also be in scope. The Act is a cross-sector law that classifies AI systems by risk and assigns obligations according to an organization’s role. section 7.1
Two decisions anchor an assessment: what the system does and what the organization does with it. The first determines whether it is minimal risk, limited risk, high risk, or prohibited. The second determines whether the organization is a provider, a deployer, or both. Building an application on a third-party model still makes the application builder its provider; operating it can also bring deployer duties. section 7.7
Turn those decisions into an operational workflow:
- Inventory each system’s use, users, outputs, and decision-making context. Start with the system that would cause the most damage if it failed. section 7.1 section 7.7
- Assess risk tier and provider or deployer roles with legal counsel. Vendor assurances do not remove deployer duties. section 7.0 section 7.7
- Retain traceable, system-level evidence of risk assessment, data governance, human oversight, logging, and monitoring. Organizational frameworks alone do not produce the required product-level documentation. section 7.7
- Revisit the assessment when features, decision contexts, or vendor models change; each can change the system’s obligations. section 7.7
Disclose an AI interaction
An insurance FAQ chatbot can use a persistent chat label to tell users they are interacting with AI. Assign someone to own the wording, test it across channels, and check that it survives interface updates. Disclosure can be omitted when the AI nature is obvious from context; when in doubt, disclose. section 7.2.5
Transparency work also includes marking AI-generated images, audio, and video with machine-readable metadata. Responsibility depends on how the media is generated: a vendor may handle marking through its interface, while a team building its own generation tool is responsible for marking its output. C2PA provides signed metadata inside media files; it is not a visible watermark. AI literacy for staff also applies across risk tiers. section 7.2.5
Check the supply chain before relying on a model provider’s evidence. A system provider integrating a general-purpose AI model is owed technical documentation on request; a deployer of a finished high-risk product is owed instructions for use, not the full technical file. Open-source general-purpose models have partial documentation exemptions, although systemic-risk models lose that exemption. section 7.3.5
Table 7.1 What else applies when you’re subject to the AI Act
|
Legal layer
|
When it
applies
|
Book
reference
|
|
Privacy law
(GDPR)
|
Your AI
system processes personal data
|
Chapters 4-5
|
|
Product
safety
|
AI is
embedded in physical products (medical devices, vehicles, machinery)[6]
|
Section
7.2.3
|
|
Copyright
|
You train or
fine-tune on third-party data. If a rights holder has reserved their rights,
you need a license before using their content[7]
|
Section
7.2.4 and Chapter 6
|
|
Sector
regulation
|
Finance
(MiFID II, CCD2), healthcare (Medical Devices Regulation, IVDR), critical
infrastructure (NIS2) each add obligations on top of the AI Act
|
Sector-specific
|
Check overlapping obligations, including privacy, product safety, copyright, and sector regulation; AI Act compliance is not the only consideration. section 7.1
The Act entered into force in August 2024, with obligations phasing in through 2028. Its first prohibitions took effect in February 2025, followed by general-purpose AI provider obligations in August 2025. Deadlines and technical standards require continued checking. Use this workflow to prepare evidence and questions for legal counsel, not as legal advice. section 7.0 section 7.1

