AI Policies Feature Guide
Overview
The AI Policy feature allows enterprises to codify, monitor, and enforce AI policies across the organization's AI systems, AI agents, and models. It converts static, written policies and controls into enforceable configurations, enabling teams to build and deploy AI faster while consistently meeting responsible-use and compliance standards.
For an in-depth overview of guardrail enforcements, see AI Policy Enforcement Guide.
What is an AI policy?
An AI Policy is a compliance requirement that can be automatically detected and measured across your AI inventory in OneTrust. Policies can map to regulatory controls and are applied consistently across your AI inventory. Use cases could include the following:
|
Policy |
What it measures |
|---|---|
|
Bias Testing Validation |
Has bias/fairness testing been completed with results within thresholds? |
|
Content Filtering Guardrail (Input) |
Is input content filtering configured and active? |
|
Human Oversight Configuration |
Is the autonomy level defined with appropriate human-in-the-loop controls? |
|
Jailbreak Prevention (Input) |
Are safeguards implemented to prevent jailbreak attempts that bypass system or safety controls? |
|
Limitation Documentation |
Are system limitations and risk disclosures documented and published? |
|
Prompt Injection (Input) |
Are controls in place to detect and mitigate prompt injection attempts in user input? |
|
Sensitive Information Prevention Guardrail (Output) |
Are output filters for PII and sensitive data detection enabled? |
|
Toxic/Unsafe Content Mitigation (Input) |
Is toxic, harmful, or unsafe content in user input detected and appropriately mitigated? |
To use seeded AI policies
Seeded AI policies cannot be edited or deleted.
- Use the filter to sort the list screen to Owner = OneTrust User. All of the seeded AI policies will be surfaced.
- Click on the name to enter the details screen to review the policy's details, conditions for compliance, and mapped standards or frameworks. As of 202608.1.0, there are 16 seeded policies mapped to the following standards/frameworks:
-
NIST AI RMF v1.0
-
EU AI Act (2024
-
ISO 42001:2023 (AIMS)
Note:OneTrust will continue to add seeded policies overtime. Policies will be versioned as updates are applied to existing seeded policies.
Seeded Policy
AI entity
Description
Severity Level
Conditions for compliance
AI Agent Guardrail Coverage
Agents
Confirms that all applicable guardrail protections are active across AI agents discovered from connected AI platforms.
Critical
Agent source = Azure Foundry AND Has content filtering guardrail - input = True AND Has content filtering guardrail - output = True AND Has sensitive information filters guardrail - input = True AND Has sensitive information filters - output = True
AI System Intended Use and Purpose Declared
AI system
Checks integrity of baseline registry across the AI system inventory. Intended use and purpose drive risk classification, obligation scoping, and impact assessment.
High
Intended purpose - Is not empty AND Description - Is not empty
AI System Risk Profile Defined
AI system
Checks that AI systems have computed risk profiles. This means risks have been identified against the system or its components, and that inherent and current aggregated scores are present.
High
Aggregated inherent risk score - Is not empty AND Aggregated risk score - Is not empty AND Aggregated target risk score - Is not empty
Active Agent Operating Instructions Documented
Agents
Checks that agents in active operation have their type and operating instructions recorded.
High
Instructions - Is not empty AND Agent Type - Is not empty
Active Agent Tool and Model Binding Inventory
Agents
Checks that active agents have their bound models and available tools or resources recorded.
High
Model Binding - Is not empty AND Tools - Is not empty AND Resources - Is not empty
Agent Deployment Context Declared
Agents
Checks that active agents have their deployment location and infrastructure recorded.
Medium
Agent region - Is not empty AND Hosting Environment - Is not empty
Externally-Facing AI System Governance Baseline
AI system
Applies a heightened documentation baseline to AI systems exposed outside the organization.
High
Intended purpose - Is not empty AND Description - Is not empty AND Status - Is not empty
High-Risk Model Fairness Metrics Recorded
Models
Verifies that models with high risk classification also have fairness measurements recorded.
High
Demographic parity ratio - Is not empty AND Demographic parity difference - Is not empty AND Equalized odds difference - Is not empty AND Equalized odds ratio - Is not empty
Identify Unacceptable Risk
Models
Tracks models across your inventory for an unacceptable risk classification.
Critical
Aggregated risk score - Not equals to any of 7.0 + Very high
Model Accountability and Version Traceability
Models
Verifies that every model in the inventory has a recorded version and framework.
Medium
Model version - Is not empty AND Framework - Is not empty
Model Guardrail Coverage
Models
Checks that all applicable guardrail protections are active across models from connected AI platforms.
Critical
Hosting provider = Azure Foundry AND Has content filtering guardrail - input = True AND Has content filtering guardrail - output = True AND Has sensitive information filters guardrail - input = True AND Has sensitive information filters - output = True
Model Limitations and Known Biases Documented
Models
Checks that models above minimal risk have documented intended use, known limitations, and known biases via free-text fields in assessments.
High
Limitations - Is not empty AND Biases - Is not empty
Model Output Integrity Benchmark
Models
Monitors faithfulness and correctness evaluation scores to ensure model outputs are factually accurate and grounded in source material.
High
Faithfulness >= 1 AND Correctness >= 0.8
Model Response Quality Benchmark
Models
Evaluates whether model outputs meet minimum thresholds across relevance, helpfulness, completeness, coherence, and readability.
Medium
Relevance >= 0.75 AND Helpfulness >= 0.7 AND Completeness >= 0.6 AND Coherence >= 0.9 AND Readability >= 0.9
Model Safety & Bias Benchmark
Models
Enforces zero-tolerance thresholds on harmfulness and stereotyping evaluation metrics for any model in production.
Critical
Harmfulness = 0 OR Stereotyping = 0
Third-Party Model Provenance and Due Diligence
Models
Applies value-chain due-diligence expectations to models the organization did not build itself.
High
Hosting provider - Is not empty AND Limitations - Is not empty AND Model version - Is not empty
-
- To disable a seeded policy, use the Deactivate option in the Context Menu of a given policy.
- To copy a seeded policy's configuration for a custom policy, use the Duplicate option in the Context Menu of a given policy.
To create custom AI policies
- In the Add details section, complete the fields.
Field
Option
Name
Provide a unique name to identify this policy.
Description
Provide a description of the policy's intended use and purpose.
Owner
Select a OneTrust user to serve as the policy owner.
Severity
Choose a severity level for this policy:
-
Low - 1
-
Medium - 2
-
High - 3
-
Critical - 4
Standard or Framework
Click the Add framework button to select legislation that can help your team understand the implications of this AI policy.
-
- In the Evaluated entity section, choose which entity that this policy will evaluate and monitor:
-
AI agent
-
Model
-
- Under Entity attributes, use the ANY/ALL operators and the value fields to define the filter that will target the desired entities. Use the Add attribute button to add additional attributes and operators to refine your criteria.
- In the Conditions for compliance section, use the ANY/ALL operators and the value fields to define the attributes that will indicate entities as compliant. Use the Add condition button to add additional attributes and operators to refine your criteria.
- Review the policy details, entity attributes, and compliance attributes, using the Back button as needed to make changes.
To manage AI policies list screen
- Use the filter and column sorting options to narrow and arrange your desired view.
Option
Description
Filters
-
Severity
-
Status
-
Owner
Column Sorting
-
Name (alphanumerical)
-
Severity (ascending or descending)
-
Status (active or inactive)
-
- Select one or more policies to populate the bulk delete, reactivate, or deactivate management options.
To configure enforcement
To manage AI policy violations
- The list screen will populate the entities with unresolved or suppressed violations. Click on the entity name to navigate to its given inventory details screen.
- To take action on a specific violation, click the Context Menu icon for that violation. For details about available enforcement actions, see AI Policy Enforcement Guide.
To review compliant entities
- Entities that have met the compliant conditions configured for the policy will appear in this list screen.
To review the activity history of a policy
- Select View activity. A side pane appears with an audit log of actions.
To duplicate a policy
To edit a policy
You can only edit active policies.
To manage policy activation status
To delete a policy
Deleting an AI policy is a permanent action and cannot be recovered. You can deactivate instead to end monitoring but preserve the policy.