Ressourcen
Zurück

Nehmen Sie an der AI + Data Tour teil und profitieren Sie von praktischen Schulungen, echten Kundengeschichten und dem Austausch mit Domo-Produktexperten in Ihrer Nähe.

Jetzt registrieren
Über uns
Zurück
Auszeichnungen
Seit 34 Quartalen in Folge als Leader ausgezeichnet
Sommer 2026 Leader in den Bereichen Embedded BI, Analyseplattformen, BI, ETL-Tools, Datenaufbereitung und Data Governance
Preise

What Is AI TRiSM? Trust, Risk, and Security Management

3
min read
Monday, October 26, 2026
Table of contents
Carrot arrow icon

Traditional IT governance handles who can access what data. Traditional security handles perimeter defense. Neither was designed for systems that generate novel outputs, retrieve context dynamically, or take autonomous actions. AI TRiSM fills that gap by adding evaluation, inspection, and enforcement at the exact points where AI interacts with data and workflows. This article covers the framework's core components, implementation steps, and how it differs from AI governance and AI security.

Key takeaways

If there's only time to remember a few things, these are the ones that matter.

  • Framework definition: AI TRiSM (Trust, Risk, and Security Management) is a framework Gartner introduced for governing AI systems from design through deployment and ongoing operations.
  • Three operational layers: The framework works through information governance at the data foundation, AI governance for policies and risk assessment, and runtime inspection for real-time enforcement.
  • Scope of application: AI TRiSM applies to models, applications, and agents (with agents requiring stricter controls because they take actions, not just generate outputs).
  • Compliance alignment: AI TRiSM is not a regulation itself but helps organizations meet requirements from the EU AI Act, NIST AI RMF, and ISO/IEC 42001.
  • Governance vs TRiSM: AI governance sets expectations through policies; AI TRiSM operationalizes those expectations with technical controls that actually enforce them.

Why is AI TRiSM important?

Gartner created the term AI TRiSM to address a specific gap: the distance between AI ambition and AI accountability.

Picture this: an AI assistant trained on company data surfaces confidential salary information to someone who shouldn't see it. A customer-facing agent invents a refund policy that doesn't exist. An automated workflow triggers a transaction nobody approved.

Not hypothetical. McKinsey's 2025 State of AI survey found 51 percent of organizations have experienced negative AI incidents. That's a clear signal that AI systems operating without runtime governance create measurable business risk.

AI TRiSM bridges this gap by adding evaluation, inspection, and enforcement at the exact points where AI interacts with data and workflows. The framework targets vulnerabilities unique to intelligent systems:

  • Prompt-based data leakage: Sensitive information in the context window (the text the model can see in a single request) can escape through model outputs.
  • Prompt injection: Malicious inputs can override instructions and hijack agent behavior.
  • Model drift: Outputs degrade over time without continuous evaluation.
  • Tool abuse: Agents with API access can take unintended actions when boundaries aren't enforced.
  • Compliance gaps: Regulations increasingly demand explainability, audit trails, and human oversight for automated decisions.

Here's what catches teams off guard: traditional role-based access controls often fail with AI. A person might lack permission to view a specific financial dashboard, but if the AI agent has broad database access, that same person can simply ask the agent for the restricted numbers. AI TRiSM forces access controls to apply at the data retrieval layer, not just the presentation layer.

{{custom-cta-1}}

What are the core components of AI TRiSM?

Different sources describe AI TRiSM as having four pillars, five pillars, or three layers. The variation reflects different ways of organizing the same underlying controls. Rather than getting stuck on which count is correct, focus on what the framework actually requires.

AI governance and risk assessment

This layer handles planning and policy. AI governance establishes which AI systems exist, what risk tier each belongs to, who owns them, and what policies apply. This is not the same as information governance. It focuses specifically on AI assets.

The key artifacts you need:

  • AI asset inventory: A catalog of all models, agents, and AI-enabled applications with metadata on purpose, data sources, and risk classification.
  • Risk tiering: Classification based on impact. An internal summarization tool carries different risk than an agent executing financial transactions.
  • Policy definition: Rules for acceptable use, required approvals, and prohibited actions.
  • Ownership assignment: Clear accountability spanning data, ML, security, and legal roles.

Organizations with fragmented AI adoption often discover they have far more AI systems than they realized once they start inventorying. Shadow AI is common, and it often only becomes visible once teams start inventorying what's already in use.

Information governance and data protection

AI systems are only as trustworthy as the data they access.

Information governance ensures data is classified, access-controlled, and traceable before it reaches any model or agent. You need data classification and labeling to identify sensitive information like personally identifiable information (PII) and protected health information (PHI) so access rules can be enforced. Role-based or attribute-based permissions must carry through from data storage to AI retrieval. Data lineage provides traceability from source to AI output, enabling audit and debugging. And data loss prevention controls prevent sensitive data from being included in prompts or surfaced in outputs.

If your data governance is incomplete, for example, sensitive fields aren't labeled or access rules aren't enforced at the data layer, AI TRiSM runtime controls can't fully compensate. The foundation has to be in place first. Teams sometimes assume they can skip this step and rely on output filtering alone. That approach leaves gaps that compound over time.

Runtime inspection and enforcement

This is where AI TRiSM moves from policy to action.

Runtime controls inspect prompts, evaluate outputs, and enforce boundaries in real time. Prompt inspection scans inputs for sensitive data, injection attempts, or policy violations before they reach the model. Output evaluation checks responses for accuracy, policy compliance, and data leakage before they reach the user or trigger an action. Tool sandboxing limits which APIs, databases, and systems an agent can access. Human-in-the-loop gates require approval for actions above a defined risk threshold. Audit logging records every interaction for compliance and incident response.

Stricter runtime controls add latency and friction. You will need to calibrate enforcement to risk tier. A low-risk summarization tool does not need the same approval workflow as an agent that modifies production data.

How does AI TRiSM work across data, models, and agents?

A sales manager asks an AI assistant for a forecast by region. The assistant retrieves data from the company's data platform, generates a response, and delivers it through a chat interface. AI TRiSM controls operate at every step.

The sequence looks like this:

  1. The person's prompt is inspected for sensitive data and policy violations before processing.
  2. The assistant queries governed data sources; access controls ensure the person only sees authorized data.
  3. The model generates a response; output evaluation checks for accuracy, hallucination, and data leakage.
  4. If the AI can take actions (updating a record, sending an email, triggering a workflow), approval gates determine whether it proceeds automatically or requires human review.
  5. The response is logged and delivered through the appropriate channel.
  6. The entire interaction is recorded for compliance and debugging.

Models generate outputs but do not take actions, so controls focus on input/output inspection. Applications embed models in workflows, so controls include access management and output evaluation. Agents take actions on behalf of users, which means controls must include tool sandboxing, approval workflows, and bounded autonomy.

Agents have a larger blast radius. The biggest risk is not hackers but poorly defined boundaries. Agents can modify data, trigger transactions, and interact with external systems.

A common failure mode happens when teams grant an agent broad API access but rely entirely on system prompts to restrict its behavior. Prompts can be bypassed. True tool sandboxing requires hard technical limits on the API gateway side, ensuring the agent physically cannot execute unauthorized commands regardless of its instructions.

{{custom-cta-2}}

How can organizations implement AI TRiSM?

Most organizations do not start from zero. They have existing AI systems, incomplete data governance, and distributed ownership. Implementation is less about building from scratch and more about retrofitting controls onto systems already in production while establishing governance for new deployments.

1. Inventory AI systems and data sources. Catalog every model, agent, and AI-enabled application. Document what data each system accesses, who owns it, and what actions it can take. Organizations frequently discover shadow AI during this step.

2. Classify risk and assign ownership. Tier each system by impact. Assign clear owners spanning data, ML, security, legal, and business functions. A simple RACI (Responsible, Accountable, Consulted, Informed) prevents gaps.

3. Establish information governance foundations. Ensure sensitive data is classified, access controls are enforced at the data layer, and lineage is traceable. If this foundation is incomplete, prioritize it before adding runtime controls.

4. Define policies and approval workflows. Document acceptable use, prohibited actions, and approval requirements by risk tier. Keep policies specific enough to enforce and simple enough to follow.

5. Implement runtime inspection and enforcement. Add prompt inspection, output evaluation, and tool sandboxing for high-risk systems. Calibrate strictness to risk tier. Over-enforcement creates friction that drives workarounds.

6. Monitor, audit, and iterate. Log interactions, review incidents, and refine controls based on what you learn. AI TRiSM is not a one-time project.

Buying an AI security tool before ownership, classification, and policies are clear often leads to shelfware, because the team still doesn't have enforceable rules to configure. Applying high-friction controls to low-risk tools slows adoption without reducing meaningful risk. Ignoring shadow AI? That leaves the highest-risk systems operating outside IT oversight entirely.

What are common AI TRiSM use cases?

AI TRiSM applies wherever AI systems interact with sensitive data or take consequential actions. Three workflow patterns illustrate how the framework operates in practice.

AI-assisted financial forecasting: A finance team uses an AI application to generate revenue forecasts by combining CRM data, historical financials, and market signals. The model could surface confidential deal information to unauthorized users or produce forecasts that decision-makers treat as authoritative without understanding the underlying assumptions. Role-based access ensures users only see data for their region. Output evaluation flags forecasts that deviate significantly from historical patterns. Audit logging captures every query and response.

Customer support agent with action authority: A support agent can answer questions, look up order status, and process refunds up to a defined threshold. Prompt injection could lead to unauthorized refunds. Customer PII could surface to unauthorized users. Prompt inspection blocks injection attempts. Tool sandboxing limits refund authority. Approval gates escalate high-value refunds to human review. Data masking prevents PII from appearing in responses.

Compliance reporting automation: A compliance team uses an AI application to draft regulatory reports by pulling data from multiple source systems. Risks include inaccurate data, missed disclosures, and draft reports being exposed to unauthorized reviewers. Data lineage ensures every figure is traceable to its source. Output evaluation checks for completeness against regulatory templates. Access controls limit draft visibility to the compliance team.

How is AI TRiSM different from AI governance and AI security?

These three terms often get used interchangeably, but they address different problems and involve different teams.

AspectAI governanceAI securityAI TRiSM
FocusPolicies, standards, accountabilityThreat protection, access controlOperationalizing trust, risk, and security across the AI lifecycle
Primary ownerGovernance, risk, and compliance (GRC), legal, data leadershipSecurity, ITCross-functional (data, security, machine learning (ML), legal, business)
Key artifactsPolicies, risk frameworks, compliance mappingsThreat models, access controls, monitoringAI catalog, runtime inspection, audit trails, approval workflows
Runtime scopeLimited (sets expectations)Active (blocks threats)Active (inspects, evaluates, enforces)

AI governance defines what should happen. AI security protects against external and internal threats. AI TRiSM brings both together into an operational system with technical controls, not just policies.

An organization can have strong AI governance policies and still lack AI TRiSM if those policies are not enforced at runtime. And this distinction matters more than most teams realize. AI TRiSM aligns with frameworks like NIST AI RMF, ISO/IEC 42001, and regulations like the EU AI Act, with enforcement powers effective August 2, 2026.

How Domo supports an AI TRiSM strategy

AI TRiSM requires capabilities across data governance, AI governance, and runtime enforcement. No single platform covers everything. Domo contributes to an AI TRiSM strategy by supporting three layers, Foundation, Activation, and Distribution.

Foundation: Domo helps make data AI-ready with governed integration and transformation, classification, lineage, and role-based access controls. Permissions defined for people and teams carry from ingestion through consumption, so AI experiences built on top of that data stay inside the same guardrails.

Activation: Domo is an agentic platform for the intelligent enterprise, where AI agents and apps operate on governed data with bounded autonomy. People set objectives and constraints, and human-in-the-loop approval can be required before higher-risk actions run.

Distribution: Domo delivers AI outcomes into the tools teams already use, including embedded analytics, mobile, and AI assistants, with the same role-based access enforced at the point of delivery.

Domo runs on top of an existing cloud data platform and works with a customer's preferred inference models (for example, OpenAI, Google, Anthropic, or an internal model) through Domo's AI services, so teams can activate governed data without replacing their current infrastructure.

If you're ready to turn AI TRiSM from a framework into real, enforceable controls (governed data, traceability, and role-based access included) Watch a demo and see what it looks like in practice.

See AI TRiSM controls in action—govern, inspect, enforce

Watch demo

Start building safer AI with governed data and access controls

Try free
See Domo in action
Watch Demos
Start Domo for free
Free Trial

Frequently asked questions

Does AI TRiSM apply to internal productivity tools or only customer-facing AI?

AI TRiSM applies to both, but with different control intensities. Internal summarization tools might need only basic prompt inspection and audit logging. Customer-facing agents that take actions require stricter data lineage, approval gates, and tool sandboxing. Risk tiering determines which controls apply to which systems.

How does AI TRiSM handle AI systems built on third-party models like GPT-4 or Claude?

The framework applies regardless of whether you built the model or licensed it. You still need to govern what data reaches the model, inspect outputs before they reach users, and enforce boundaries on any actions the system can take. Third-party models actually increase the need for runtime inspection since you have less visibility into how the model behaves.

What metrics indicate whether AI TRiSM is working?

Track incidents involving AI data exposure or unauthorized actions, time to deploy new AI systems with governance in place, audit findings related to AI compliance, and adoption of governed AI tools versus shadow alternatives. Tie metrics to workflows changed and compliance maintained rather than checkbox completion.
No items found.
Explore all
No items found.
AI