When You Host the Model, More of the Risk Is Yours: What Open Weight AI Means for Governance

For the past few years, most enterprise AI governance has assumed the model provider carries much of the model-level risk.

When a company uses a hosted model, the provider maintains and monitors the model and accepts contractual obligations about how it behaves. The enterprise configures it and governs how it is used. The division is reasonably clear.

Open weight models change that division for any team that hosts one itself.

Open weight models are reaching enterprise teams

Open weight models are models whose weights are publicly released, so anyone can download, host, and modify them. For many enterprises they have been a point of research interest. Improving model quality, cost pressure, and the wish to fine-tune on proprietary data inside your own environment are pushing more teams to evaluate them. Some teams already run them.

Why hosting a model changes the governance picture

When you host an open weight model, you take on much of what the provider used to handle.

Model behavior. A hosted provider monitors outputs and patches known failure modes. When you host the model, that work moves in-house.

Data handling. A hosted model comes with data processing terms and retention commitments written into a contract. When you host the model, your team decides how data is handled, and your governance record has to show those decisions.

Your role under the law. Laws such as the EU AI Act assign duties by role in the AI value chain. Fine-tuning or substantially modifying a model can change the role your organization plays, so legal should review that role at intake.

Framework coverage. Frameworks such as ISO 42001 and NIST AI RMF cover the model lifecycle. The more of that lifecycle you control, the more of those controls apply to you.

Governance built for hosted models leaves gaps

Many intake processes ask what a use case does and what data it touches. They ask less about where a model came from or how it was changed. When the model is yours, those questions matter.

Provenance. Where did the base model come from, what was it trained on, and what are its known limits? A hosted provider supplies much of this. With an open weight model, your team finds it, assesses it, and records it.

Fine-tuning. If you fine-tune a model on internal data, that change belongs in the record. What data was used, who approved it, and what controls applied?

Incident ownership. When a hosted model produces a problem output, the vendor is part of the response. When your own hosted model does, your team owns the response from start to finish.

Risk against your own tolerance. Your organization owns more of what can go wrong with a model it hosts, so a use case that scores medium on a hosted model could score higher on a model you run.

The use case is still the unit of governance

Airia governs by use case, not by model or tool. The use case carries the purpose, the risk, and the obligation. An open weight model is one more asset behind a use case, and one asset can serve several use cases. Risk tier, frameworks, controls, and evidence all attach to the use case, so the same process covers a hosted model and a model your team runs.

In Airia, each use case carries inherent and residual risk, scored against the company’s own risk tolerance. Risk can be mitigated, accepted, avoided, or transferred, with a recorded justification for each choice. Approval routes by risk, and high and critical use cases go to a committee. Airia covers 16 frameworks, including the EU AI Act, ISO 42001, and NIST AI RMF, and scopes each framework to the use case, so one intake can trigger different duties in different places. Approved mitigations become live controls through the AI gateway and MCP gateway.

Start before open weight becomes the norm

Most teams are still early with open weight models. That gives governance leads room to prepare. Extend intake to ask where a model came from, how it was changed, and who owns incidents. Keep those answers in the use case record. Score risk with that ownership in mind.

The question is not only whether your teams will host their own models. It is whether your governance program will see them when they do.

Gartner named Airia in the 2026 Magic Quadrant for AI Governance Platforms, positioned furthest on Completeness of Vision, and ranked it #1 for the AI Security Use Case in the Critical Capabilities report.

Request a demo to see how Airia governs every AI use case from intake to evidence.

Airia closes the gap between policy as written and policy as enforced.