The general-purpose model rules in the EU AI Act bind the companies that build and release large models, not the businesses that run AI agents on them. Deployers inherit information, not obligations: documentation that helps them understand the model. Everything an agent does with tools, data and money sits outside those rules and with the business.
What do the general-purpose model obligations require?
General-purpose AI model obligations have applied since 2 August 2025. In substance, every provider of a general-purpose AI model must:
- prepare and keep up to date technical documentation of the model, including its training and testing, for the authorities;
- prepare and make available information and documentation to downstream providers who intend to integrate the model into their own AI systems, so they understand its capabilities and limits and can meet their own obligations;
- put in place a policy to comply with EU copyright law;
- publish a sufficiently detailed summary of the content used to train the model.
Providers of models with systemic risk, the most capable models, carry more: evaluating the model, including adversarial testing; assessing and mitigating systemic risks; tracking and reporting serious incidents; and ensuring adequate cybersecurity for the model and its infrastructure. Some relief exists for models released under free and open-source licences, except where they carry systemic risk.
Our glossary entry on what a general-purpose AI model is explains the term.
Who carries them, and where does a deployer sit?
The model provider carries them. A business that builds an agent by calling a model through an interface is not a general-purpose model provider. It is building an AI system on top of a model, which makes it a downstream provider of that system if it puts it into service under its own name, or a deployer if it uses a system someone else built.
The question whether general-purpose model rules apply to agent builders usually has the answer no, with one caveat: a business that modifies a model significantly, for example through extensive fine-tuning, may take on provider obligations for the modification. Configuring a brief, connecting sources and setting limits is not modifying the model.
The timing also differs. General-purpose model rules already apply; high-risk system rules are phasing in on dates that have been subject to proposed changes, as set out in EU AI Act timeline for AI agent deployers.
What does the agent builder actually inherit?
Three things pass downstream in practice:
- Documentation about the model. Capabilities, limits, intended and unsuitable uses, and how the model was tested. This is the input you need for your own risk assessment and, if your agent is high-risk, for your technical documentation.
- Evidence about systemic risk testing, for the most capable models. Useful when you assess how your agent might be misused, but it covers the model in general, not your agent's tools.
- A dependency. If the model provider changes the model, your agent's behaviour changes. Their obligations do not include telling you every change in a way that suits your release process.
What does not pass down is any assurance about how the model behaves inside your agent, with your sources, your connections and your customers.
Where is the gap?
The general-purpose model rules govern the model. The high-risk rules govern AI systems in listed uses. Between them sits most of what makes an agent risky:
- Actions. Calling a payment interface, updating a customer record, sending an email. The model rules say nothing about what a system built on the model may do.
- Permissions. Which tools and data each capability may touch, and within what limits.
- Sources. Which documents and records the agent reads at runtime, and whether any could carry injected instructions.
- Oversight. Who approves what, and when.
For a high-risk agent, the Act's system-level duties fill part of that gap. For the many agents that are not high-risk, nothing in the Act does. Sector rules, data protection law and the firm's own controls have to. The full list of deployer duties is in what the EU AI Act requires deployers of AI agents to do.
How do you meet your side of the line?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Understand the model you build on | Model provider's downstream documentation obtained and reviewed | Filed documentation with review notes |
| Assess model limits in your use | Agent tested on your own cases, not only on the provider's benchmarks | Test results per capability |
| Manage model change | Model version pinned per agent; re-test before adopting a new version | Version record; re-test results |
| Govern actions | Each capability bounded by scope, amount and counterparty | Bound settings and refusal records |
| Govern sources | Source register with trust class; outside content cannot authorise actions | Register and provenance in records |
| Meet system duties if high-risk | Deployer or provider duties for the agent, as applicable | Duties checklist with evidence |
Our deployer obligations checklist under the EU AI Act covers the system-level duties in detail.
What will a supervisor or auditor ask?
- Which model does each agent run on, and which version?
- What did the model provider tell you about its limits, and how did you use it?
- What happens when the model changes?
- Which controls sit outside the model and would hold if the model misbehaved?
- Have you modified the model itself?
What do the model rules not cover?
They do not cover your agent, its tools, its data or its actions. They do not make the model provider answerable for how you deploy it. They do not certify that a model is suitable for a regulated use. And they do not replace your own testing: a model that passed its provider's evaluation can still fail on your cases. For the full set of duties, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
Pryme treats the model as one component and governs the agent around it. We never update model weights for a customer; qualification changes the agent's brief, sources, bounds and certification instead. Every action meets five bounds enforced outside the model, so a model change cannot widen what an agent may do without fresh evidence from Attest.
Questions people ask
Does using a model with systemic risk make my agent high-risk?
No. Systemic risk is a property of the model; high-risk is a property of the system's use. A drafting agent on a very capable model can be minimal risk, and a credit agent on a modest model can be high-risk.
Can I ask the model provider for its downstream documentation?
Yes. The obligation exists so that downstream providers can understand the model and meet their own duties. Ask for it before you build, and again when the model changes.
Does fine-tuning make me a general-purpose model provider?
It can, if the modification is significant. Light configuration does not. Keep a record of what you changed and why you judged it below or above that line.