What agentic AI governance is, and why model governance does not cover agents
✓
The six factors that decide how risky an AI agent is
✓
A four-class model for classifying agents, from assist to high impact
✓
Which controls each risk class needs before release
✓
How to run governance in practice: agent registry, lifecycle gates and roles
✓
How agent risk classes fit with the EU AI Act, NIST AI RMF, ISO/IEC 42001 and Singapore's agentic framework
Most companies already have some form of AI governance: a policy, an approval board, maybe a model inventory. Then the first agents arrive, and the existing process either blocks everything or waves everything through, because it was built to review models, not software that acts.
Agentic AI governance is the set of roles, rules and controls that decide which AI agents an organization may build and run, what each one is allowed to do, and how it is monitored, with oversight proportional to the risk each agent carries. The key word is proportional. A meeting-notes assistant and an agent that approves refunds should not go through the same review.
This guide gives you a practical way to make that distinction: agent risk classes. It covers the factors that make an agent risky, a four-class model, the controls for each class, and how to run it day to day without slowing every team down.
Why model governance is not enough for AI agents
Model governance asks whether a model is accurate, fair and documented. Agent governance has to ask more, because agents take actions. Regulators and standards bodies are converging on the same point. In January 2026, Singapore's IMDA published the first government framework written specifically for agentic AI, which centers on an agent's action space (the tools and systems it can reach), its autonomy, and whether its actions can be reversed. The World Economic Forum's "AI Agents in Action" framework from November 2025 makes similar recommendations, and the OWASP Top 10 for Agentic Applications lists the security risks that follow from getting it wrong.
In the US, banking supervisors went the other way: the April 2026 model risk guidance SR 26-2 explicitly leaves generative and agentic AI out of scope, which means banks have to design their own agent governance. We cover the legal side of that gap in enterprise AI governance is now a legal liability.
The 6 factors that decide how risky an AI agent is
Score every agent on six factors before it is built. Each one is a question you can answer in a design review.
Action space: what can the agent touch? Read-only data, internal records, external messages, money, or production systems.
Autonomy: does a person approve each action, a sample of actions, or none?
Reversibility: can a wrong action be undone easily, with effort, or not at all?
Impact on people: could the agent's output affect someone's access to credit, care, a job, or another significant decision?
Data sensitivity: public data, internal data, personal data, or special categories such as health data.
Scale: how many actions, customers or tenants does one mistake reach before someone notices?
A four-class model for agent risk
The six factors roll up into four classes. The highest factor usually decides the class: an agent that touches payments is at least class 3, however simple it is.
Class
What the agent does
Examples
Typical controls
Class 1: Assist
Reads and suggests; a person acts
Summaries, search, draft replies, analysis
Standard security review, usage logging, quality spot checks
Class 2: Act with review
Prepares actions that a person approves one by one
Draft invoices for approval, proposed ticket updates
Approval step in the UI, audit trail, evaluation before release
Class 3: Act within limits
Acts on its own inside thresholds; exceptions go to people
Auto-match transactions under a limit, tag and route tickets
Limits and thresholds, monitoring and alerts, rollback, sampled review
Class 4: High impact
Affects money, health, legal rights or production at scale
Formal risk assessment, human approval at material steps, regulatory mapping, independent validation
Agents can move between classes. Most of the agents we build start in class 2, prove their accuracy against a baseline, and move to class 3 for the cases where the evidence supports it. That path is safer, and it also builds the trust that adoption depends on.
Every agent gets an entry before it is built: owner, purpose, risk class, tools, data sources, models and versions, approval points, evaluation results and review date. The registry is what lets you answer a regulator, an auditor or your own board in minutes.
2. Put gates in the lifecycle, sized to the class
Design review: classify the agent and agree on controls before development.
Pre-release evaluation: test against a baseline your domain experts accept.
Gradual rollout: limited users or volume first, wider release on evidence.
Monitoring: quality, cost, incidents and denied actions, with alerts.
Periodic re-review: yearly for class 1, quarterly or on every material change for class 4.
3. Name the roles
Each agent needs a business owner who is accountable for its output, a technical owner who runs it, and a reviewer from risk, compliance or security for classes 3 and 4. A domain expert, the person whose work the agent supports, should sign off its evaluation criteria.
How agent risk classes fit with existing frameworks
Framework
What it gives you
How risk classes connect
EU AI Act
Legal risk tiers based on the use case, with high-risk duties from December 2027 and August 2028
A class 4 agent in an Annex III use, such as credit scoring, is also legally high-risk
NIST AI RMF
Govern, Map, Measure, Manage functions for AI risk
Classification is part of Map; class-based controls are Manage
ISO/IEC 42001
A certifiable AI management system
The registry and lifecycle gates are evidence for its controls
The six factors follow its focus on action space, autonomy and reversibility
OWASP Agentic Top 10
Security risks specific to agents
Use it as the threat model for class 3 and 4 reviews
How we apply agent risk classes at Linnify
In our ARC (Agentic Release Control) framework, risk and compliance exposure is scored in the very first phase, alongside value, so that the riskiest ideas are either redesigned or deliberately chosen with the right controls. The agent requirements we write in the Ingest phase record decision boundaries and approval points, and the Deploy phase ships access control, logging and governance sign-offs with the release. Our principle is simple: a named person is always accountable for what an agent does, and production-grade controls are there from day one.
Classify every agent by what it can touch, how freely it acts, how reversible its actions are and who it affects, then scale the controls to the class. Governance built that way lets low-risk agents ship quickly and gives high-impact agents the scrutiny they need.
AI governance
Setting up governance for your first agents?
Talk to Linnify about agent risk classes, registries and release gates that fit how your teams already work.
Agentic AI governance is the set of roles, rules and controls that decide which AI agents an organization may build and run, what each is allowed to do and how it is monitored, with oversight proportional to each agent's risk.
How do you classify the risk of an AI agent?+
Score it on six factors: action space, autonomy, reversibility, impact on people, data sensitivity and scale. The highest factor usually sets the class. A simple four-class model runs from assist, to act with review, to act within limits, to high impact.
What controls does a high-risk AI agent need?+
A formal risk assessment, human approval at material steps, limits on what it can do, a complete audit trail, independent validation before release, monitoring with alerts and a mapping to the regulations that apply, such as the EU AI Act.
What is an AI agent registry?+
An inventory of every agent with its owner, purpose, risk class, tools, data sources, models, approval points, evaluation results and review date. It is the basis for audits and for answering regulators quickly.
How does agent governance relate to the EU AI Act?+
The EU AI Act classifies risk by use case. Agent risk classes add the dimensions the Act does not, such as autonomy and reversibility. A high-impact agent used for credit scoring, for example, is also high-risk under the Act from December 2027.
Is there a government framework for agentic AI?+
Yes. Singapore's IMDA published a Model AI Governance Framework for Agentic AI in January 2026. It asks organizations to bound risks upfront, keep humans meaningfully accountable, apply technical controls and inform end users.
References and further reading
• Davis Wright Tremaine. (2026, January). New governance frameworks offer a roadmap for managing risks unique to agentic AI (Singapore IMDA and WEF). dwt.com
• OWASP GenAI Security Project. (2025, December). OWASP Top 10 for Agentic Applications, reported by TechRepublic. techrepublic.com