EU AI Act vs Council of Europe AI Convention: Which Obligations Reach Private Companies?

Two different legal mechanisms

The Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law sets common principles and rules so that activities within the life cycle of AI systems respect human rights, democracy and the rule of law. The European Commission describes it as the first legally binding international instrument on AI. Its obligations run to the Parties, which must give it effect in domestic law.

The AI Act, Regulation (EU) 2024/1689, works differently. It is a Regulation that applies to operators by role. A company does not wait for a national measure to find out whether it is a provider or a deployer. The role and the risk tier determine the duties.

The two instruments are linked. In the EU, the Convention is implemented exclusively through the AI Act and other relevant EU acquis, where applicable. The EU signature was authorised by Decision (EU) 2024/2218, and Decision (EU) 2026/1080 approves conclusion of the Convention and the attached EU declarations. The Commission also describes the Convention as risk-based and differentiated, and as including key concepts of the AI Act.

Side-by-side comparison

The table summarises the points that matter most when scoping compliance work for a private company.

• Who is bound: Convention: the Parties. AI Act: providers, deployers, general-purpose AI model providers and other operators, inside and outside the EU where the Act's triggers are met.
• How duties arise: Convention: through domestic implementation and declarations. AI Act: directly, through the Regulation.
• Public sector: Convention: applies to AI life-cycle activities of public authorities and of private actors acting on their behalf. AI Act: applies to public and private actors alike.
• Other private actors: Convention: Parties must address risks and impacts consistently with it and declare how, either by applying its obligations directly or by equivalent measures. AI Act: direct duties by role and risk.
• Exclusions: Convention: national defence is outside scope, national security activities need not be covered, and the Commission notes exemptions for research and development. AI Act: not covered here beyond the scope triggers.
• Geographic reach: Convention: the territories where the EU treaties apply, for the EU. AI Act: systems placed on the EU market, put into service, or used in the EU.

Private-sector reach: direct, derived, or declared

It helps to sort obligations into three groups. The first is duties that apply to your company directly under the AI Act. The second is Convention commitments that reach private actors only through EU and national implementation. The third is other Union law that applies in parallel, including data protection, consumer protection, fundamental rights, employment and product safety. The AI Act is stated to be without prejudice to, and complementary to, that law, so AI Act scoping does not replace other regimes.

For a private company operating in the EU, the second group should not be read as a source of separate duties. The EU applies the Convention's obligations to private actors through the AI Act and other relevant EU acquis. The practical question is therefore whether the AI Act and related law cover your use, not whether the Convention does.

Two cases need closer attention. If you act on behalf of a public authority, the Convention's principles and obligations apply to that activity more directly, so your contract and role should be examined. If you operate in non-EU Convention Parties, private-sector reach depends on each Party's declaration and domestic implementation, and you should assess each jurisdiction separately rather than extrapolate from the EU position.

Role, risk tier and timing under the AI Act

The AI Act applies to actors that place an AI system or general-purpose AI model on the EU market, put an AI system into service, or use it in the EU. This includes actors outside the EU. The Commission's examples show how roles split: a developer of a CV-screening tool is a provider, and a bank using that tool is a deployer. The same enterprise can hold both roles for different systems, so role is best recorded per system, not per company.

Providers of general-purpose AI models carry a distinct set of obligations. These have applied since 2 August 2025, together with the governance structure and penalty rules. Providers of models above a capability threshold must also meet additional risk management and cybersecurity obligations.

High-risk timing is the point most likely to change. As of November 2026, the Commission's Digital Omnibus on AI, proposed on 19 November 2025, would set 2 December 2027 for stand-alone high-risk AI systems and 2 August 2028 for high-risk systems embedded in products. Plan high-risk work against both the current dates and the proposed dates, keep a change log of amendments, and re-baseline when the status changes.

Which obligations need operational controls and retained evidence

Some duties can be met largely with documents: a policy, a classification memo, a role assignment. Others describe what a system must do or what must be shown about how it operated. Human oversight of high-risk systems, record-keeping, and logging of high-risk system operation are the clearest examples. A written oversight policy does not show that a person reviewed a consequential output. A retention policy does not show that a specific interaction was captured.

The exact form, content and retention period for each of these duties must be taken from the Regulation's text and official guidance, and counsel should confirm them. The architectural point holds regardless of the final reading. For each system, you need to know where it executes, where oversight can be exercised, and where the evidence of what happened is produced and kept.

A workable approach is to hold role and risk tier as data attributes on each entry in your AI system inventory, along with the EU-market, put-into-service or use-in-EU trigger and the legal basis for the classification. Controls and evidence can then be mapped per system, and an auditor can trace a classification decision to the controls attached to it. Assign named owners for role classification, risk-tier determination and evidence retention.

Where runtime enforcement supports AI Act obligations

This page is educational and not legal advice. Where counsel concludes that an obligation depends on controls operating at runtime and on retained evidence, Trussed is one way to implement that layer. Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants: allow, redact, block, rate limit, and escalate for human approval, at the level of individual interactions and tool calls. Enforcement is at runtime, not reviewed after the fact.

That enforcement generates continuous audit-ready evidence, mapped to the EU AI Act among other frameworks. Trussed deploys self-managed in the customer cloud, with configurable retention and data residency. It does not decide your role or risk classification. Those remain legal determinations for your counsel and control owners.