---
title: How do you use ISO/IEC 42005 for AI agent impact assessments?
url: https://opencontroller.lyzr.com/blog/iso-42005-ai-impact-assessment-for-agents
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How do you use ISO/IEC 42005 for AI agent impact assessments?

**ISO/IEC 42005:2025 gives guidance on how to run and document an AI system impact assessment: who and what could be affected, the benefits and harms, and what you will do about them. For agents, use it as the method behind ISO/IEC 42001 clause 6.1.4 and Annex A control A.5, and add agent-specific inputs: autonomy level, tools and permissions, delegated identity, and kill or quarantine plans.**

Teams adopting ISO/IEC 42001 hit the same question: the management system standard says to assess impacts, but not how. ISO/IEC 42005 fills that gap.

## What is ISO/IEC 42005?

ISO/IEC 42005 was published in May 2025 by ISO/IEC JTC 1/SC 42. According to [ISO's listing](https://www.iso.org/standard/42005), it provides guidance for organisations conducting AI system impact assessments, focused on how AI systems and their foreseeable applications may affect individuals, groups and society, across the AI lifecycle. ISO describes it as complementing ISO/IEC 42001 (AI management systems) and ISO/IEC 23894 (AI risk management).

It is guidance, not a certifiable requirements standard. You are certified against ISO/IEC 42001; ISO/IEC 42005 helps you do one part of it well.

## How does it connect to ISO/IEC 42001?

| ISO/IEC 42001 reference | Requirement | How ISO/IEC 42005 helps |
|---|---|---|
| Clause 6.1.4 | Perform an AI system impact assessment | Defines a repeatable process and documentation |
| Annex A, A.5.2 | Impact assessment process | Process steps, timing, roles, thresholds |
| Annex A, A.5.3 | Documentation of impact assessments | What the assessment record should contain |
| Annex A, A.5.4 | Impact on individuals or groups | Identifying affected parties and harms |
| Annex A, A.5.5 | Societal impacts | Broader effects beyond direct users |

The risk assessment in 42001 asks what could go wrong for the organisation. The impact assessment asks what could go wrong for people. Auditors look for both. See [ISO 42001 evidence for AI agents](/blog/iso-42001-evidence-for-ai-agents) for the wider control set.

## What agent-specific inputs should the assessment include?

A generic template written for a scoring model misses what makes agents different. Add these inputs for every agent:

1. **Autonomy level.** Does it suggest, act with approval, or act alone?
2. **Tools and actions.** Every API, MCP server and write action it can take, with the most damaging one named.
3. **Identity.** Does it act as itself, or on behalf of a user? With what scopes?
4. **Data reach.** Data classifications it can read, and where outputs go.
5. **Affected parties.** Employees, customers, third parties, and people who never interact with it but are affected by its actions.
6. **Failure modes.** Hallucinated actions, prompt injection via tools, loops, wrong-recipient actions.
7. **Foreseeable misuse.** How a user or attacker could steer it outside its purpose.
8. **Controls.** Guardrails, evaluations, monitoring, and how fast you can stop it.
9. **Model dependency.** Which model and version, and what happens when it changes.

## How do you run it for an agent, step by step?

1. **Set thresholds.** Decide which agents need a full assessment (customer-facing, consequential decisions, write access to systems of record) and which get a short-form screen.
2. **Trigger at the right time.** Before first deployment, and again on material change: new tools, new model, new user group, higher autonomy.
3. **Assign roles.** Agent owner drafts; risk, security, privacy and legal review; an accountable executive approves.
4. **Describe the system.** Purpose, users, deployment context, the inputs above.
5. **Identify impacts.** Benefits and harms for each affected party, including reasonably foreseeable failure and misuse.
6. **Rate and treat.** Severity and likelihood; mitigations; residual impact.
7. **Record and approve.** Store the assessment against the agent version in your registry.
8. **Monitor.** Tie production signals (guardrail hits, eval scores, incidents) back to the assessment and re-open it when they drift.

If the agent processes personal data, run this alongside a [DPIA for AI agents](/blog/dpia-for-ai-agents) instead of duplicating work. For a framework-neutral version, see [AI impact assessment for AI agents](/blog/ai-impact-assessment-for-ai-agents). Our pillar on [mapping agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) shows how one evidence set serves several standards.

## What evidence will an auditor ask for?

- The documented process and thresholds.
- A completed assessment per in-scope agent, linked to the version assessed.
- Proof of re-assessment after material changes.
- Monitoring records that show controls operating.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, so every assessment can be tied to the exact version it covers. Posture flags risky permissions and configurations; multi-turn and red-team evals, LLM judges on production traffic, and live groundedness and task-adherence guardrails supply the monitoring evidence. Policy as code includes built-in mappings to ISO 42001.

## Frequently asked questions

**Can you be certified to ISO/IEC 42005?**
No. It is a guidance standard. Certification is against ISO/IEC 42001, which requires impact assessments; ISO/IEC 42005 helps you perform them.

**How often should an agent's impact assessment be repeated?**
Before deployment and whenever there is a material change, such as new tools, a new model, a new user group or more autonomy, plus a periodic review.

**Is an impact assessment the same as a risk assessment?**
No. A risk assessment focuses on risks to the organisation; an impact assessment focuses on effects on individuals, groups and society.

To tie impact assessments to live agent versions, [book a demo](https://www.lyzr.ai/book-demo/).

## Key facts

ISO/IEC 42005:2025 (published May 2025) is guidance for AI system impact assessments covering effects on individuals, groups and society across the lifecycle; it complements ISO/IEC 42001 and ISO/IEC 23894 and is not certifiable. ISO/IEC 42001 requires an AI system impact assessment in clause 6.1.4 and Annex A controls A.5.2 to A.5.5. For AI agents, assessments should add autonomy level, tools and write actions, delegated identity scopes, data reach, failure modes such as prompt injection and loops, and stop/quarantine controls, and be repeated on material change. Opencontroller ties assessments to agent versions in its registry and maps policy as code to ISO 42001.

## Related

- [How do you run an AI impact assessment for an AI agent?](https://opencontroller.lyzr.com/blog/ai-impact-assessment-for-ai-agents.md)
- [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 run a GDPR DPIA for an AI agent?](https://opencontroller.lyzr.com/blog/dpia-for-ai-agents.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
