---
title: Can one governance program cover classic ML models and AI agents?
url: https://opencontroller.lyzr.com/blog/govern-ml-models-and-ai-agents-together
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# Can one governance program cover classic ML models and AI agents?

**Yes: one program, one policy owner and one inventory can cover classic ML models, LLM applications and AI agents. But the controls underneath differ. ML models are judged on prediction quality and drift; LLM apps on output quality and data leakage; agents on what they do, with whose authority, through which tools.** Programs that copy the model-risk playbook onto agents miss the action layer entirely.

The call question was broad: "We need governance across data products, ML models and AI agents. Can one approach cover all three?" This post shows what can be shared and what cannot. For how data governance fits in, see [AI governance vs data governance](/blog/ai-governance-vs-data-governance).

## What can be shared across all three?

- **Policy and accountability.** One AI policy, one risk taxonomy, one committee. NIST AI RMF's Govern function and ISO/IEC 42001's management-system clauses apply to all AI systems, not one type.
- **Inventory.** One register, with a type field and type-specific fields.
- **Risk tiering.** The same tiers (for example: internal low-impact, internal action-taking, customer-facing or regulated) drive review depth for every type.
- **Ownership.** Every asset has a named human owner, tied to the identity provider.
- **Evidence store.** One place auditors go, with the same retention rules.

## What differs between ML models, LLM apps and agents?

| Dimension | Classic ML model | LLM application | AI agent |
| --- | --- | --- | --- |
| Main risk | Wrong or biased predictions | Wrong, leaked or harmful text | Wrong or unauthorised actions |
| Key inventory fields | Training data, features, target, performance metrics | Model, prompt version, data sources, output channels | Model, tools, MCP servers, identity, on-behalf-of scope, autonomy level |
| Pre-release tests | Holdout accuracy, fairness metrics, stability | Quality evals, jailbreak tests, PII leakage tests | Multi-turn evals, tool-misuse red-teaming, permission review |
| Change trigger | Retraining, data drift | Prompt or model change | Prompt, model, tool or permission change |
| Production monitoring | Drift, performance decay | Groundedness, toxicity, leakage | Behavioural deviation, tool-call patterns, cost |
| Runtime control | Rollback to previous model | Output filters, redaction | Block, quarantine, kill, tool scoping |
| Who it acts as | Nobody; it returns a score | Usually the app | A service identity or a user, on their behalf |

The bottom three rows are where agent governance is new. A credit model returns a score and a human or system acts on it. An agent may call the refund API itself.

## Why does the model-risk playbook not stretch to agents?

Model risk management grew up around validating a model before use and monitoring its performance. Agents break three assumptions:

1. **The unit changes constantly.** An agent's behaviour changes when someone adds a tool or edits a prompt, with no retraining.
2. **Validation is behavioural, not statistical.** You test sequences of actions under adversarial conditions, not accuracy on a holdout set.
3. **Control must be at runtime.** You need to stop an action in seconds, not schedule a revalidation.

US banking supervisors made this distinction explicit in 2026: as covered in [what AI agent controls banks need](/blog/ai-agent-controls-for-banks-model-risk), the revised model risk guidance excludes agentic AI, leaving firms to define agent controls themselves.

## How do you structure one program with two control layers?

1. **Single register** with a `type` field: model, LLM app, agent. See [what an AI agent registration questionnaire should ask](/blog/ai-agent-registration-questionnaire).
2. **Shared tiering** applied to all types.
3. **Type-specific test suites**: statistical validation for ML; evals and red-teaming for LLM apps and agents.
4. **Type-specific runtime controls**: model rollback for ML; guardrails, gateway policy, block and quarantine for agents.
5. **One evidence view** across types, mapped to the frameworks you report against.

## How Opencontroller does this

Opencontroller focuses on the agent and LLM layer. It auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, connects to platforms such as Databricks, Salesforce and SAP, and keeps a governed registry of agents, versions, tools and MCP servers. Evals, LLM judges, live guardrails, the LLM Gateway and rogue-agent controls cover testing and runtime, with policy mappings to OWASP, MITRE ATLAS, NIST and ISO 42001 as evidence for those frameworks. Start with [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should the same team validate ML models and agents?**
They can sit in one function, but agent testing needs security and red-team skills that classic validators may not have.

**Is an LLM app an agent?**
Not until it can take actions through tools. Once it can, govern it as an agent.

**Does ISO/IEC 42001 cover agents?**
It covers AI systems generally. It does not prescribe agent-specific controls, so you add them in your own control set.

## Key facts

One AI governance program can cover classic ML models, LLM applications and AI agents with a shared policy, inventory, risk tiering, ownership and evidence store. Controls differ: ML models are judged on prediction quality and drift; LLM apps on output quality and leakage; agents on actions, tools, identity and on-behalf-of scope, needing behavioural tests and runtime block, quarantine and kill. NIST AI RMF and ISO/IEC 42001 apply to all AI systems but do not prescribe agent-specific controls. Opencontroller by Lyzr governs the agent and LLM layer with discovery, a governed registry, evals, guardrails, an LLM Gateway and rogue-agent controls. Pricing starts from $250K per deployment per year.

## Related

- [What is the difference between AI governance and data governance?](https://opencontroller.lyzr.com/blog/ai-governance-vs-data-governance.md)
- [What AI agent controls do banks need now that model risk guidance excludes agentic AI?](https://opencontroller.lyzr.com/blog/ai-agent-controls-for-banks-model-risk.md)
- [What should an AI agent registration questionnaire ask?](https://opencontroller.lyzr.com/blog/ai-agent-registration-questionnaire.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
