AI Governance Agent and Model Evaluation Metrics

OneTrust AI Governance provides native support for ingesting, governing, and reporting on agent and/or model evaluation metrics generated by external AI platforms. This capability enables organizations to centrally monitor agent and/or model quality, reliability, and risk indicators. The platform is designed to support continuous AI governance, audit readiness, and regulatory compliance by operationalizing model evaluation data within the OneTrust AI inventory and governance workflows. 

OneTrust AI Governance ingests agent and/or model evaluation metrics directly from native LLM evaluation jobs (e.g., Amazon Bedrock), preserving the original system as the source of truth.

Key Benefits

Note:

Amazon Bedrock & AgentCore: Evaluations are only available on models. For AgentCore, evaluations are available on agents.

Microsoft Foundry: Evaluations are available on agents and models.

Databricks: Evaluations are available on agents.

Vertex (now Gemini Enterprise Agent Platform): Evaluations are available on agents.

• Centralized visibility into model evaluation outcomes 

• Automated, audit-ready evidence collection 

• Strong data integrity with non-editable metrics 

• Reduced manual effort in AI governance operations

Prerequisites

• Your AI models are evaluated using a supported AI platform that provides native evaluation jobs (for example, automated LLM-as-a-Judge evaluations). 

• You have access to OneTrust AI Governance with permissions to configure AI Automation. To deploy a lightweight worker node, follow the steps here. 

• Supported Model Evaluation jobs have been executed at least once in the source AI platform.

Known Limitations

• Evaluation Metrics are not currently supported in Reports or Insights Dashboards. 

• Custom Metrics are not currently available.

OneTrust AI Governance integrates with AI platforms that natively generate model evaluation results (for example, evaluation jobs executed within managed AI services). These evaluation jobs assess the quality of model outputs using standardized metrics and scoring approaches. 

The platform supports the ingestion of built-in evaluation metrics provided by the source AI platform. For Amazon Bedrock, please refer to AWS documentation.

Metrics are typically scored on a normalized scale (e.g., 0–1), reflecting the quality of agent and/or model outputs as determined by the evaluation job configuration. OneTrust displays the latest available evaluation results for each agent and/or model, under Agent/Model > Details & Evaluation Metrics, ensuring stakeholders always see the most current assessment of model performance. 

To preserve trust and auditability, evaluation metrics brought through this native integration are non-editable within OneTrust AI Governance. OneTrust stores and presents evaluation outcomes. This design ensures that governance teams, auditors, and regulators can rely on the integrity of reported metrics to make informed governance decisions.

Evaluation metrics per supported platform and AI entity type:

To monitor and enforce thresholds for reported metrics, it is highly encouraged to set up custom policies and track individual or groups of metrics against organizationally defined thresholds. A breach in these thresholds results in policy violations and allows deviations in a large agent inventory to surface.

For more information on creating custom policies, see AI Policies Feature Guide.

Evaluation metrics are collected through scheduled observability jobs that run at a configurable interval (commonly every 24 hours). Each execution captures any new or updated evaluation jobs from the connected AI platform and synchronizes results into OneTrust. This enables:

• Continuous monitoring of model quality 

• Automated evidence collection for compliance programs 

• Reduced operational overhead for AI governance teams

Within OneTrust AI Governance, evaluation metrics are surfaced directly on the model record, allowing users to:

• Review the most recent evaluation scores at a glance. 

• Correlate evaluation outcomes with associated risks, controls, and policies. 

• Include evaluation metrics as part of governance reporting and documentation.

While the platform displays the latest evaluation result, historical evaluation data is retained in the system to support future enhancements such as trend analysis and longitudinal reporting.

Responsible use metrics help assess fairness and potential disparities in outcomes produced by AI models. They complement the quality, safety, relevance, and other evaluation metrics imported from supported AI platforms.

There is no single, universally accepted metric for summarizing fairness as one overall score. Today, OneTrust collects Responsible use metrics and will continue to expand them over time to produce a Model Fairness report.