# Keydris > Keydris checks AI agent permissions before governed actions run. Keydris Gateway routes supported model requests to approved cloud or self-hosted models. An organization using AI makes two decisions. The first is where a model may process a workload. The second is what an agent may do on someone’s behalf. A team might approve a model to summarize customer records and still require a person to approve exporting those records. Keydris lets the team express both choices as policy. **How a team sets up a governed workflow** 1. **Define the work.** Identify the task, the people or team responsible, and the resources the agent needs. A finance agent preparing an internal report may need to read that team’s files. Publishing the report or accessing another team’s files are separate actions with different consequences. 2. **Choose the model destination.** Decide which models may process each kind of material. For example, a team can use Claude in an approved cloud deployment for public documents and a self-hosted model for confidential records. The organization sets these rules to match its own data classifications. 3. **Set action permissions.** Decide which actions may proceed, which require a reviewer, and which are outside the agent’s authority. A software team might permit reading a repository, require approval to merge a pull request, and reject changes to a different repository. 4. **Run the work.** For a model request routed through Keydris Gateway, the workload’s routing policy determines the permitted destination. When an agent attempts a governed action, Keydris checks the authority it has at that point. Approval to use a model does not by itself grant permission to change a file or send data elsewhere. 5. **Review the result.** A decision record shows the agent, the action it requested, the policy in effect, the decision, and its time. The execution outcome is recorded separately. An operator investigating a failed change can distinguish a policy rejection from a permitted request that failed in the receiving system. **What happens when an action needs approval?** ALLOW permits the agent to attempt the action. APPROVAL REQUIRED holds the action for a human decision. REJECT prevents the governed action from proceeding. A team can therefore let an agent do routine work without granting it standing permission for a consequential step. These decisions cover actions connected to Keydris through a supported integration. **How does model routing support AI sovereignty?** In the Keydris Gateway enterprise pilot, the organization chooses who may process a supported request and where inference may run. For public material, it can approve a cloud provider and deployment region. For a workload that must stay within a particular region, it can select a model running on its own infrastructure there. Before approving a cloud destination, the organization should confirm that the deployment’s processing location and the provider’s data-handling terms meet its requirements. **What changes when an agent’s assignment ends?** The organization can narrow its policy or revoke the agent’s authority. For example, after a repository review ends, revocation prevents the agent from using that authority for its next governed action. The agent process may still be running. Earlier actions and their decision records remain available for review, and revocation does not undo work that already completed. ## Product - [AI agent authorization](https://keydris.com/agent-authorization): permissions and human approval for governed actions. - [Keydris Gateway](https://keydris.com/ai-gateway): model routing, enterprise pilot scope, and data residency. - [Model routing example](https://keydris.com/#model-routing): choosing a model based on data sensitivity. - [Authorization decisions](https://keydris.com/#decisions): how identity, policy, and human review apply to a governed action. - [Frequently asked questions](https://keydris.com/#faq): answers about authorization, model routing, and data residency. ## Learn more - [Documentation](https://keydris.com/docs): public product guides and getting started. - [Contact Keydris](https://keydris.com/contact): discuss supported workloads and deployment requirements. ## Optional - [Research](https://keydris.com/research): Keydris investigations into AI agent authority. - [Founder thesis](https://keydris.com/thesis): why software that acts needs explicit authority.