Skip to content
Beyond Prompt AI Studio
The EU AI Act in Practice

Provider or Deployer? The Question That Decides Everything

Two companies deploy the exact same high-risk AI system – and carry completely different obligations. The reason isn't the system, it's the role. This module lays out the full obligation catalogues for both roles and the three concrete ways you cross the line without noticing.

Four role scenarios – worth remembering

Try it yourself: use unchanged, or substantially modify?

Stays deployerBecomes provider

You use an open language model unchanged for internal customer support.

No substantial modification, no new high-risk purpose – you stay a deployer with lighter obligations.

Two roles, two sets of obligations

A provider develops an AI system or places it on the market under its own name. Its obligations for high-risk systems are extensive: a risk management system across the entire lifecycle, technical documentation, conformity assessment, CE marking, entry in a public EU database, a quality management system, and post-market monitoring. A deployer uses an AI system for its own professional purposes. Its obligations are noticeably lighter even in the high-risk case: use the system per the instructions, ensure human oversight by trained personnel, inform affected people, keep automatically generated logs for at least six months, and – in certain cases – carry out a fundamental rights impact assessment (see “Human Oversight and the Fundamental Rights Impact Assessment”).

The provider trap: three concrete triggers

You don't become a provider only through intent, but through three clearly defined actions (Art. 25):

  • Your own brand: you place an existing high-risk system on the market under your own name or brand – even if you didn't develop it yourself.
  • Substantial modification: you change a high-risk system already on the market such that it's no longer covered by the original conformity assessment.
  • Repurposing a general-purpose system: you take a system that wasn't originally classified as high-risk and change its purpose so it now falls into a high-risk category.

What counts as a “substantial modification”?

A substantial modification isn't a trivial threshold. It exists when a change wasn't part of the original conformity assessment and affects compliance with the requirements – or when the system's intended purpose changes. Deliberately fine-tuning a model for a new, more sensitive use case is already enough, even if the underlying technical architecture stays the same.

Practice section: the role check before every new AI project

Before introducing a new AI system or adapting an existing one, three questions help: Is the system being passed on under your own name or brand? Is something being changed about the system's behaviour that goes beyond mere configuration? Is a previously harmless system being used for a new, potentially high-risk purpose? If the answer to any is yes, the role check belongs at the start of project planning – in the architecture decision, not as an afterthought once branding and code are already fixed. This classification doesn't replace case-by-case legal advice.

The key points

  • Your role decides your workload more than the risk tier: providers carry documentation, conformity assessment, CE marking, and EU registration; deployers mainly human oversight, logs, and information duties.
  • Three concrete actions trigger the provider role: your own brand, substantial modification, repurposing a general-purpose system for a high-risk use.
  • A “substantial modification” isn't a trivial threshold – deliberately fine-tuning for a new, more sensitive purpose can already be enough.
  • The role check belongs at the start of a project, in the architecture decision, not as an afterthought.
  • This classification doesn't replace case-by-case legal advice – when unsure about your own role, bring in a specialist.

The EU AI Act: what companies really need to know – and what's just panic

Quick check: did it sink in?

1 / 3

Which obligation typically applies ONLY to the provider, not the deployer?

Want to make sure your own AI application doesn't unknowingly trigger provider obligations?