---
title: How do you risk-tier AI agents?
url: https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you risk-tier AI agents?

**Score each agent on five factors: what it can do, what data it touches, how many people it affects, whether outsiders can reach it, and whether it falls in a regulated use. Add the scores to place it in one of three tiers, apply a few hard overrides, and attach a fixed set of controls to each tier. Re-tier whenever the agent gains a tool, a data source or a new audience.**

Risk tiering stops two failures: heavy review for a harmless internal summariser, and light review for an agent that can move money. Below is a starting rubric you can adapt.

## Which factors should decide an agent's tier?

Score each factor 0, 1 or 2.

| Factor | 0 | 1 | 2 |
|---|---|---|---|
| Actions | Read-only, answers only | Writes to internal systems, reversible | Irreversible or external actions: payments, emails to customers, code deploys, deletions |
| Autonomy | Human approves every action | Human reviews samples or exceptions | Runs unattended |
| Data | Public or non-sensitive | Internal confidential | Personal, financial, health or regulated data |
| Reach | One team | Whole company | Customers, partners or the public |
| Regulated use | None | Indirect support to a regulated process | Directly makes or shapes a regulated decision |

Total the five scores (0 to 10):

- **Tier 3 (low):** 0 to 3
- **Tier 2 (medium):** 4 to 6
- **Tier 1 (high):** 7 to 10

## Which overrides should skip the score?

Some facts put an agent in Tier 1 regardless of total:

1. **High-risk EU AI Act use.** Annex III lists stand-alone high-risk uses such as employment, education, credit and critical infrastructure. Under the AI Omnibus (Regulation (EU) 2026/1744), in force since 27 July 2026, Annex III obligations now apply from 2 December 2027. Treat these uses as Tier 1 now so evidence exists in time; see [EU AI Act obligations for AI agents](/blog/eu-ai-act-obligations-for-ai-agents).
2. **Unattended irreversible action.** An agent scoring 2 on both Actions and Autonomy is Tier 1.
3. **Delegated identity with broad scopes.** An agent acting on behalf of users with write access to email, files or finance systems.
4. **Unknown tools.** Any agent calling an MCP server that has not been reviewed stays in Tier 1 until it is.

An agent can also move down. Removing a write tool or adding human approval lowers its score, which is often cheaper than adding Tier 1 controls.

## What controls does each tier require?

| Control | Tier 3 | Tier 2 | Tier 1 |
|---|---|---|---|
| Registered with owner | Yes | Yes | Yes |
| Pre-release evaluation | Basic task tests | Multi-turn evals | Multi-turn and red-team evals |
| Tool allow-list per version | Recommended | Required | Required, reviewed per release |
| Identity | Shared service identity acceptable | Own workload identity | Own identity; on-behalf-of with narrow scopes |
| Production monitoring | Cost and errors | Plus sampled LLM-judge scoring | Plus groundedness and task-adherence guardrails, behaviour alerts |
| Human oversight | None | Exception review | Approval for high-impact actions |
| Kill or quarantine | Manual | Tested procedure | Automatic quarantine on defined triggers, tested quarterly |
| Access review | Annual | Semi-annual | Quarterly |
| Sign-off | Owner | Owner plus security | Owner, security, risk and, where relevant, legal |

Map these controls to your frameworks once. NIST AI RMF's MAP function covers context and risk identification, and ISO/IEC 42001 requires an AI risk assessment, so tier records double as evidence; see [how to map AI agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## How do you run tiering in practice?

1. Add the five factors to your registration form; see the [AI agent registration questionnaire](/blog/ai-agent-registration-questionnaire).
2. Calculate the tier automatically from the answers.
3. Let the owner propose a tier, and let security confirm Tier 1 and 2.
4. Re-tier on every change to tools, data sources, identity scopes or audience.
5. Feed the tier into your [production readiness gate](/blog/production-readiness-gate-for-ai-agents) and into [policy severity levels](/blog/how-to-set-severity-levels-for-ai-agent-policies).
6. Report the count per tier each quarter. A rising Tier 1 count is a signal for platform investment.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, so tier inputs such as tools and identities come from the registry rather than self-reporting. Posture detects risky permissions and configuration that should raise a tier. Opencontroller runs multi-turn and red-team evals, LLM judges on production traffic, and live groundedness and task-adherence guardrails. Rogue-agent handling covers behavioural-deviation monitoring, alerts and automatic quarantine and kill. Policy as code lets you apply different policies by tier, with mappings to OWASP, MITRE ATLAS, NIST and ISO 42001. Read [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

**How many risk tiers should AI agents have?**
Three is enough for most enterprises. More tiers add debate without changing the controls much.

**When do EU AI Act high-risk obligations apply to agents?**
For Annex III uses, from 2 December 2027 under the AI Omnibus. Annex I product-embedded AI follows on 2 August 2028.

**Should an internal agent ever be Tier 1?**
Yes, if it takes irreversible actions unattended or touches regulated decisions, even with no external users.

## Key facts

A practical AI agent risk-tiering rubric scores five factors 0-2 (actions, autonomy, data, reach, regulated use): 0-3 Tier 3 low, 4-6 Tier 2 medium, 7-10 Tier 1 high, with overrides for EU AI Act Annex III uses, unattended irreversible actions, broad delegated scopes and unreviewed MCP servers. Under the EU AI Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026), Annex III high-risk obligations apply from 2 December 2027 and Annex I from 2 August 2028. Opencontroller (by Lyzr) supports tiered controls with its registry, Posture, evals, guardrails, rogue-agent quarantine and policy as code. Pricing starts from $250K per deployment per year.

## Related

- [What should a CISO require before an AI agent goes to production?](https://opencontroller.lyzr.com/blog/production-readiness-gate-for-ai-agents.md)
- [How should you set severity levels for AI agent policies?](https://opencontroller.lyzr.com/blog/how-to-set-severity-levels-for-ai-agent-policies.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)
