---
title: How do you run an AI impact assessment for an AI agent?
url: https://opencontroller.lyzr.com/blog/ai-impact-assessment-for-ai-agents
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you run an AI impact assessment for an AI agent?

**Use the structure in ISO/IEC 42005:2025, the international guidance on AI system impact assessment, and add the questions that make agents different: which tools they can call, which actions they take without a person, whose identity they use, and which systems they can change. Assess before deployment, and again whenever the agent's tools, data or autonomy change.**

An impact assessment asks what the system could do to people and groups, not just to your organisation. That is different from a security risk assessment. For an agent, the answer depends less on the model and more on what the agent is connected to.

## What is ISO/IEC 42005 and how does it relate to ISO/IEC 42001?

ISO/IEC 42005 was published in May 2025. It is guidance, not a certifiable standard. ISO/IEC 42001 requires organisations to assess AI system impacts (clause 6.1.4, and Annex A control area A.5). ISO/IEC 42005 explains how.

It has two main parts:

- **The process**: documenting it, linking it to other management processes, timing, scope, responsibilities, thresholds for sensitive and restricted uses, performing and analysing the assessment, recording, approval, and monitoring and review.
- **The documentation**: what the assessment record should contain.

## How do you apply each section to an agent?

This table follows the documentation sections of ISO/IEC 42005 and adds what to ask about an agent.

| ISO/IEC 42005 section | Agent-specific questions |
|---|---|
| Scope | Which agent versions, deployments and user groups are covered? |
| System description and capabilities | Which models, tools and MCP servers? Which actions can it take alone? |
| Purpose, intended and unintended uses | What tasks is it for? What could users or attackers make it do instead? |
| Data information and quality | Which data can it read, retrieve, remember and send? Is memory persistent? |
| Algorithm and model information | Which model versions? Who is the provider? How are model changes controlled? |
| Deployment environment | Which countries and languages? Which platforms? Is it customer-facing? |
| Interested parties | Who is directly affected: users, customers, employees, third parties the agent contacts? |
| Actual and potential impacts | Benefits and harms; failures; reasonably foreseeable misuse such as prompt injection or excessive agency |
| Measures | Guardrails, approvals, tool limits, monitoring, kill switch, redress |

## What agent-specific harms should you consider?

Agents create harms that a classifier or chatbot does not:

1. **Wrong action at scale.** A refund, email or record change repeated across thousands of cases before anyone notices.
2. **Acting beyond authority.** Using a broad credential to reach data the requesting user should not see.
3. **Hidden automation.** People not knowing they are dealing with an agent, or not being able to reach a person.
4. **Unfair outcomes.** An agent that triages, prioritises or approves requests differently across groups.
5. **Manipulation by inputs.** Third-party content steering the agent through injected instructions.
6. **Chained effects.** One agent's output becoming another agent's instruction.

For each harm, record likelihood, severity, who is affected, and the measure that reduces it.

## How do you run the assessment, step by step?

1. **Set thresholds.** Decide which uses are sensitive or restricted, and what triggers a full assessment. Your [risk tiers](/blog/how-to-risk-tier-ai-agents) can do this.
2. **Gather facts.** Pull the agent's registry record: owner, versions, tools, data classes, identity model.
3. **Identify affected people.** Include people outside the company who receive the agent's actions.
4. **Assess impacts.** Use red-team and multi-turn evaluation results, not just opinions. Test the misuse cases you listed.
5. **Decide measures.** Map each significant harm to a control: a guardrail, an approval step, a tool restriction.
6. **Approve.** A named person signs off, separate from the builder.
7. **Monitor.** Set the signals that show the assessment is still valid: judge scores, policy violations, complaints.
8. **Reassess on change.** New tools with write access, new data, a new model or wider autonomy should trigger a review.

## How does this relate to the EU AI Act?

Article 27 of the EU AI Act requires certain deployers of high-risk AI systems, such as public bodies and some financial services uses, to carry out a fundamental rights impact assessment before use. Following 2026 amendments, the main obligations for stand-alone high-risk systems are scheduled to apply from 2 December 2027. An ISO/IEC 42005-style assessment covers much of the same ground and is useful evidence, but check Article 27's specific content requirements. See [EU AI Act obligations for AI agents](/blog/eu-ai-act-obligations-for-ai-agents).

## How Opencontroller does this

Opencontroller supplies the facts an assessment needs. The governed registry lists each agent's owner, versions, tools and MCP servers. Posture flags risky permissions and configuration. Multi-turn and red-team evals test misuse cases, and live groundedness and task-adherence guardrails and LLM judges show how the agent behaves in production. Built-in mappings to ISO 42001 and NIST link the evidence to controls. See the [framework mapping guide](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is ISO/IEC 42005 certifiable?**
No. It is guidance. Certification is against ISO/IEC 42001, which requires impact assessments; ISO/IEC 42005 shows how to run them.

**When should an AI agent impact assessment be repeated?**
On material changes: a new tool with write access, a new data class, a model change, wider autonomy or a new user group. Otherwise at least annually.

**Who should perform the assessment?**
The agent owner prepares it, with input from security, privacy and legal. Someone independent of the build team approves it.

## Key facts

ISO/IEC 42005:2025 (published May 2025) gives guidance for AI system impact assessments: a process (when to assess, scope, responsibilities, thresholds for sensitive and restricted uses, performing, analysing, recording, approval, monitoring and review) and documentation covering scope, system information (description, capabilities, purpose, intended and unintended uses), data, algorithms and models, deployment environment, interested parties, actual and potential impacts including failures and foreseeable misuse, and measures. For an AI agent, add agent-specific inputs: the tools and MCP servers it can call, its autonomy level and which actions run without human approval, whose identity it acts under, which systems and data it can change, and the people it interacts with. ISO/IEC 42001 requires AI system impact assessment (clause 6.1.4, Annex A.5); the EU AI Act requires a fundamental rights impact assessment (Article 27) for certain deployers of high-risk systems. Opencontroller by Lyzr provides the registry, posture, eval and guardrail evidence that feeds such assessments.

## Related

- [What evidence does ISO/IEC 42001 expect for AI agents?](https://opencontroller.lyzr.com/blog/iso-42001-evidence-for-ai-agents.md)
- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [What does the EU AI Act require for AI agents, and when?](https://opencontroller.lyzr.com/blog/eu-ai-act-obligations-for-ai-agents.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
