How AI TRiSM Can Be Applied to the Public Sector

What Is AI TriSM, and Why Is It Critical for Government AI?

Gartner describes AI TRiSM as a framework for managing AI model governance, trustworthiness, fairness, reliability, security and data protection.

The framework has gained traction as agencies recognize that AI systems require continuous oversight rather than one-time governance reviews.

“The security space likes frameworks,” Rosiek says. “They like ways to define what needs to be done and define all the different components so they can figure out how to mitigate the risk.”

He says AI TRiSM is intended to help organizations get ahead of AI’s rapid evolution rather than react after problems emerge.

“AI is moving so fast,” Rosiek says. “Frameworks and compliance can’t keep up with technology.”

That challenge becomes even more complicated as agencies begin experimenting with agentic AI systems that can autonomously pass information between applications, models and workflows.

“You may have good boundaries and controls on the first stage of that conversation,” Rosiek says, “but then it goes two or three other levels where you have no visibility.”

READ MORE: AI sandboxes allow governments to experiment safely.

How To Operationalize AI TRiSM: A Three-Step Framework

While AI governance can seem overwhelming, agencies can begin operationalizing AI TRiSM through several foundational steps.

Step 1: Establish AI oversight policy aligned to existing governance

Rosiek says agencies should begin with visibility into their data and clear internal governance policies.

“The foundational part of AI within your organization comes down to your IT infrastructure, and understanding and having visibility into your data,” he says.

That means identifying what data exists, where it resides, who has access to it and what types of information should never be exposed to public AI systems.

Rosiek also warns agencies about “shadow AI,” where employees may already be using commercial AI tools without oversight.

“Making sure you have policies and training within the organization” is critical, he says.

Rather than creating entirely separate AI governance programs, agencies can often build on existing cybersecurity, compliance and data governance structures. AI TRiSM becomes a framework for extending those controls into AI environments.

“If you really have supercritical data, you absolutely need to figure out this plan and start rolling that out,” Rosiek says.

Step 2: Implement model monitoring to detect bias, drift and accuracy degradation

Continuous monitoring is another core component of AI TRiSM.

AI systems can drift over time as data changes, and models may begin producing inaccurate or biased outputs without obvious warning signs.

Rosiek says monitoring becomes even more important because of the speed at which AI systems operate.

“The biggest benefit of AI is the speed at which it can do stuff,” he says. “It can analyze data, conduct actions, and it never gets tired.”

That speed also creates risk. A small mistake or misconfiguration can quickly scale into a larger operational problem.

“AI in some ways is unpredictable,” Rosiek says. “It will make a decision, and sometimes you may not know why.”

He compares AI agents with both insider threats and overly enthusiastic interns.

“You don’t just turn the summer intern loose in your organization,” Rosiek says. “You’ve got lots of training, mentoring and controls around them.”

In early deployments, agencies may still need humans reviewing outputs and approving actions. Over time, however, organizations will increasingly rely on “humans on the loop” rather than “humans in the loop,” where staff monitor systems rather than approve every individual action.

“But the only way you can really get the benefit of AI is in that scenario,” Rosiek says.

Step 3: Secure the data pipelines that train and run government AI systems

Rosiek says protecting AI systems ultimately comes down to protecting data pipelines and understanding how information flows between systems.

“Do we know what data we have? Do we know who has access to it? Where is sensitive data?” he says.

That becomes especially important when AI agents or large language models interact with regulated information such as personally identifiable information or healthcare data.

Agencies need to understand where data travels, how it is governed and whether models are authorized to access it.

“Is the data compromised? Is it already poisoned?” Rosiek says. “If you feed poisoned data into an AI system, it could change how it performs.”

He also recommends that agencies prepare for AI failures before deploying systems into production environments.

“If the model does something it shouldn’t have done,” Rosiek says, “how do you undo that mistake?”

That includes maintaining backups, preserving immutable copies of training environments and developing contingency plans for unexpected AI behavior.

LEARN MORE: Governments can embrace Backup as a Service for peace of mind.

What Is Required for Responsible AI in Citizen-Facing Applications?

As governments deploy AI in citizen-facing systems, explainability and accountability become essential.

Residents will increasingly expect agencies to explain how automated decisions are made, especially in areas involving benefits eligibility, healthcare or public safety.

Rosiek says agencies should approach AI cautiously when systems involve sensitive constituent data.

“If you just turn it loose to the wild, it’s not going to end well,” he says.

Agencies also need to understand how AI systems interact with public data, third-party platforms and external models. Without strong governance, sensitive information could inadvertently be exposed or mishandled.

The challenge becomes even more significant as AI agents begin making decisions or initiating actions autonomously.

“You can’t protect what you don’t know you have,” Rosiek says.

How Do Agencies Build a Scalable AI Governance Program?

Rosiek says many of the foundational elements of AI TRiSM overlap with broader zero trust and cyber resilience initiatives already underway across government.

“To achieve some level of maturity with zero-trust architecture, a prerequisite is knowing what your data is and who should have access,” he says.

That visibility challenge is complicated by decades of accumulated government data.

“Organizations are data hoarders,” Rosiek says. “They’ve been storing data for 10 years, 20 years, 30 years.”

Still, he says, smaller local governments may ultimately have an advantage because they often manage less complex environments.

“If they align the stars, they could be incredibly effective and free up a lot of resources,” Rosiek says.

For agencies beginning their AI journeys, he recommends focusing less on rushing deployments and more on building foundational governance, monitoring and resilience capabilities first.

“Starting with that plan first and walking through this methodical process is really the key to success,” Rosiek says.