リスク管理¶

Risk management helps evaluate and document how a Use Case, model, or deployment complies with an organization's risk framework, including the built-in EU AI Act framework. Apply a framework, select a risk level—optionally with help from the guided EU AI Act questionnaire—and DataRobot automatically evaluates the in-platform mitigation methods that the framework requires against the asset the framework is applied to.

A risk management framework can be applied at several points across the AI lifecycle. Each surface applies frameworks and evaluates mitigation methods independently:

Risk management is always available for a Use Case, regardless of the target types of the models it contains. It’s also available for registered model versions, deployments, and custom model versions with these target types:

• Predictive: Binary, regression, multiclass, multilabel, and geospatial
• Generative: Text generation and agentic workflow

Experiments, projects, and the models on their Leaderboard don’t have their own risk management. A Leaderboard model gets its own risk management once you register it. Until then, it’s covered by its Use Case.

Because a framework can be applied at more than one stage of the model lifecycle, several other pages show a preview of previously applied risk management frameworks. These previews are read-only; to apply or change a framework, go to the asset's own Risk management location listed in the preceding table.

Apply a risk management framework¶

To apply a risk management framework to a Use Case, registered model version, deployment, or custom model version:

1. In the asset's Risk management section, click + Apply risk management framework. See above for location information.

2. From the Risk management framework dropdown, select a framework. If the framework includes instructions, they're displayed under Risk management framework instructions.

3. Select a Risk level. The available options are defined in the selected framework's configuration (as set by the organization administrator). The selection applies to this risk assessment only; other frameworks applied to the same asset keep their own risk level. If you selected the EU AI Act framework, proceed to the next step to set the risk level using the EU AI Act questionnaire.

4. If you selected the built-in EU AI Act framework, review the disclaimer and click Acknowledge and start questionnaire to launch the EU AI Act questionnaire. When complete, review the outcome and either restart the questionnaire or click Apply results.

   Once applied, the assessment outcome automatically sets the risk level on the Apply risk management framework page. You can review, restart, or delete the questionnaire from that page at any time.

   

5. Optionally, enter a Risk description describing the risk associated with this asset.

6. Review Mitigation and compliance requirements. This lists the mitigation methods the framework requires, grouped by category.

   For the EU AI Act framework, this section is split across the In-platform mitigation methods, Out-of-platform compliance requirements, and GPAI Downstream Information tabs.

7. Review all settings, then click Apply.

Review an applied framework¶

Once a framework is applied, the asset's Risk management section shows the selected risk level and any provided description as well as a section on mitigation and compliance. Mitigation methods are organized into categories: Security, Audibility, Documentation, Risk management, Human feedback, Evaluation, External validation, and Monitoring.

DataRobot automatically evaluates a method's status wherever it can be determined from the asset configuration. Otherwise, evidence may need to be provided manually (for example, by filling in the risk description) or confirmed by an administrator. Click a method to expand and review the compliance status:

Use the Edit or Remove framework actions to modify the framework assignment. The remove action is labeled by asset type, for example Remove framework from Use Case. If more than one framework is applied to the same asset, use the Risk management framework dropdown to switch between them.

Check and raise a compliance score¶

After you apply a framework to a registered model version, custom model version, or deployment, its Risk management section shows a Compliance score. The score represents the percentage of the selected risk level's required mitigation methods that are Passing, out of every method that's Passing, Failing, or Not set. Methods that don't apply to the asset aren't part of the calculation. A Use Case's Risk management page doesn't show a score, since a Use Case doesn't evaluate mitigation methods the way a model or deployment does.

You can check and raise a compliance score by following these steps for a registered model version, custom model version, or deployment:

1. Open the asset's Risk management section (see above for where to find it for each asset type).
2. If a framework isn't already configured, apply a framework and select a risk level (see Apply a risk management framework). This works the same way for any framework. Once a framework is applied, the Compliance score appears above the mitigation methods.
3. Next to the score, use the Current asset's methods / All methods tabs to display either the methods that count toward this asset's score or the framework's complete list, respectively.

4. Expand a method and if not set, the entry lists what it requires. The next steps depend on the method and the asset:

   • Filling in a Risk description on the same page satisfies Risk description filled for any asset.
   • Enabling a guard under Evaluation and Moderation satisfies a guard-based method like Prompt injection metric on a custom model version
   • Enabling drift detection in Settings satisfies Drift monitoring on a deployment.

5. Make the change described by the method, then return to the asset's Risk management section. The method evaluates to Passing and the Compliance score increases.

Changing the Risk level can also affect the score: a higher risk level typically pulls in more required mitigation methods, so the same asset can show a lower score at a stricter risk level even when nothing else changed.

EU AI Act framework¶

This section applies if you applied the built-in EU AI Act framework template to a Use Case, model, or deployment. When using the built-in framework template, instead of choosing a risk level yourself, DataRobot walks you through a guided questionnaire that classifies the AI system under Regulation (EU) 2024/1689. It then evaluates the asset against the mitigation methods and obligations that apply at that level. In-platform methods (monitoring, guards, documentation, and similar controls) are checked automatically where DataRobot can determine status from the asset configuration; out-of-platform requirements and GPAI downstream information remain your organization's responsibility outside the product. Use the results to see what still needs to be configured in DataRobot and what must be handled through legal, operational, or third-party action.

EU AI Act questionnaire¶

The EU AI Act questionnaire is a guided, roughly 5-minute flow that helps explore how the AI system-related provisions of the EU AI Act apply to an AI system. Answer a series of short questions about how, where, and for what purpose the AI system operates to get help with:

• Classifying the risk level of an AI system under the EU AI Act.
• Identifying GPAI downstream information.
• 出力内容の開示要件を特定する。
• そのリスクレベルに必要な軽減措置を特定する。
• Identifying the obligations that apply to the provider and deployer of the AI system.

Launch the questionnaire when you apply the EU AI Act risk management framework. The questionnaire walks through the following sections:

The questionnaire reaches one of the following outcomes, based on the answers given:

• Potentially out of the EU AI Act's scope of application (Article 2): The EU AI Act may not apply to the AI system. Voluntary compliance with its requirements is nonetheless recommended.
• EU AI Act prohibited AI practice detected (Article 5): The AI system may involve one or more practices prohibited under the EU AI Act. Systems in this category cannot be placed on the market, put into service, or used in the EU. AI規制への準拠について、計画されているユースケースを評価するにあたり、法務部門または法務顧問にご相談ください。
• High-risk AI system, with or without a high-risk exception, and with or without an underlying GPAI model (with or without systemic risk). To be compliant with the EU AI Act: establish and maintain a risk-management system across the AI system's lifecycle; apply data-governance and quality criteria to training, validation, and test data; create technical documentation and maintain automatic event logs; ensure appropriate human oversight and provide clear information to deployers; meet accuracy, robustness, and cybersecurity requirements; and complete a conformity assessment, affix CE marking, and register the AI system in the EU database.
• Limited risk AI system, with or without an underlying GPAI model (with or without systemic risk). To be compliant with the EU AI Act: provide technical documentation sufficient to demonstrate compliance and maintain a basic level of transparency appropriate to the system's purpose; adherence to voluntary codes of conduct is encouraged.

EU AI Act obligations¶

In DataRobot, EU AI Act obligations are the compliance requirements that apply as a result of the framework and questionnaire's assigned risk level. Access them from the Mitigation and compliance requirements section, which is organized into three tabs:

• In-platform mitigation methods: The mitigation methods DataRobot evaluates automatically.
• Out-of-platform compliance requirements: EU AI Act obligations that must be fulfilled outside DataRobot, through the organization's own legal, operational, or third-party action.
• GPAI Downstream Information: Transparency information a GPAI provider must supply to downstream AI system providers under Article 53(1)(b).

プラットフォーム内での軽減手法¶

In-platform mitigation methods are controls DataRobot can evaluate on the asset itself—for example AI literacy, RBAC, compliance documentation, drift/accuracy/fairness monitoring, logs and tracing, and guards such as prompt injection or PII detection. See the full list of mitigation methods. Mitigation methods are evaluated for compliance status as either passing, failing, or not set. These methods influence the compliance score that is generated for an asset.

プラットフォーム外のコンプライアンス要件¶

Out-of-platform compliance requirements are obligations an organization must meet outside of DataRobot. The following table lists the EU AI Act obligations, which must be addressed through your organization's own legal, operational, or third-party action.

GPAI downstream information¶

The GPAI downstream information tab lists the obligations for transparency information a GPAI provider must supply to downstream AI system providers under Article 53(1)(b). The following table lists the specific articles; what is shown in the tab—the requirements—depends on whether the GPAI model carries systemic risk. See also the LLM GPAI reference for descriptions of each available LLM.

次のステップ¶

After assessing risk for your asset, continue with Use Case management, register your model, or review your deployments.

• Track Use Case value: Define the value you expect to get from your Use Case and monitor the actual value received over time.
• Register custom models: Add the custom model to Registry so you can deploy it and make predictions.
• Deployment overview: View the model- and environment-specific summary for a selected deployment, including replacement activity.