Implementation Guide

ISO/IEC 42005 Reassessment Triggers: When to Redo an AI Impact Assessment

Redo an AI impact assessment when a change moves the system outside the boundary that was assessed: its models, tools, data, users, intended use, or the legal and organizational context around it. Changes that stay inside that boundary need a documented review, not a new assessment. In both cases, record the criteria applied, the decision and the evidence. The trigger categories in this guide are a proposed implementation approach that fits alongside ISO/IEC 42005 and ISO/IEC 42001.

Setting up a repeatable process

  1. 1

    Record the baseline

    Capture models, tools, data sources, user populations, intended uses, policies and the assessment version as the assessed boundary.

  2. 2

    Connect detection channels

    Feed change management, release records, incident management, regulatory monitoring and runtime signals into one intake.

  3. 3

    Triage against criteria

    Apply the category rules and the materiality test to each change, escalating unclear cases to a named reviewer.

  4. 4

    Decide and record

    Log the criteria applied, the decision, the decider and the date, for immaterial changes as well as material ones.

  5. 5

    Link and retain evidence

    Keep the assessment version, change record, rationale and any runtime evidence consulted together under one reference.

  6. 6

    Update the baseline

    After a reassessment or review, revise the recorded boundary so the next comparison starts from the current state.

Reassessment is a change-detection problem

Most regulated enterprises complete an AI impact assessment once and then keep changing the system. Models are swapped, tools and data sources are added, user populations grow, policies are revised and incidents occur. Without defined criteria, teams either rerun assessments for every minor release or leave them untouched until the document no longer describes what runs in production.

ISO describes ISO/IEC 42001 as making impact assessments a consistent and required part of AI governance, and ISO/IEC 42005 as making those assessments thorough, systematic and actionable. A practical reading is that the trigger process belongs at the management-system level, where changes are already tracked, while the reassessment itself is performed at the system level using the 42005 method.

The working artifact is a recorded baseline of the assessed boundary: the models in use, the tools and data sources the system can reach, the intended uses, the user populations, the applicable policies and the assessment version. A trigger is any observed difference between that baseline and reality. The categories below are an editorial proposal for turning that idea into decision rules. Check them against the standard text your organization holds.

Trigger categories, decision rules and evidence

Each category needs an owner, a detection channel, a rule that separates full reassessment from documented review, and a defined evidence set. The rules below are illustrative starting points to calibrate to your own risk appetite.

  • System and model changes: Full reassessment: a model is replaced, or a change alters behavior in the assessed use. Documented review: a patch within the same model version with unchanged evaluation results. Evidence: release record, before and after evaluation results, assessment version.
  • Intended use and user population: Full reassessment: a new purpose, a new type of decision, or newly affected groups. Documented review: more users of the same type for the same use under the same controls. Evidence: scope request, approval, rationale.
  • Data and tool access: Full reassessment: a tool that can take external or write actions, or a new category of sensitive data. Documented review: a new source of an already assessed data category under the same access controls. Evidence: access change ticket, tool inventory comparison, data classification.
  • Performance and incident signals: Full reassessment: an incident affecting individuals, or sustained degradation relevant to assessed impacts. Documented review: a contained event within tolerance with a recorded root cause. Evidence: incident record, metrics, corrective actions.
  • Regulatory and organizational context: Full reassessment: a new obligation or policy change that alters risk classification, impact criteria or acceptable use. Documented review: a change confirmed not to apply to the system. Evidence: monitoring log, applicability note.

Separating full reassessment from documented review

A single materiality test keeps decisions consistent across categories. Ask whether the change alters who could be affected, what the system can do, what information it acts on, or how well it performs for the assessed purpose. A yes to any of these points to full reassessment. A no points to a documented review, which is a short, dated record that the change was examined against the criteria and found to sit inside the assessed boundary.

The tradeoff is real in both directions. A low threshold produces reassessment fatigue, and reviewers begin to approve changes without reading them. A high threshold produces stale assessments and a gap between the documentation and production behavior. Set the threshold with the people who own the affected systems, write down the reasoning, and revisit it after the first two review cycles using the cases that were disputed.

Combine event-driven triggers with a calendar review. Some drift is gradual and never produces a single change ticket, so a periodic check of the baseline against the running system catches what event channels miss. The EU AI Act describes risk management for high-risk AI systems as a continuous, iterative process run throughout the lifecycle, with regular systematic review and updating. Whether that applies to a given system depends on its classification, but the principle of ongoing review supports a cadence alongside the triggers.

Runtime signals as inputs to the decision

Change tickets show what teams intended to change. Runtime observations show what the system is actually doing. When policy is enforced in the execution path of individual interactions and tool calls, the enforcement layer sees events that can indicate a boundary has moved: a tool being called that was not in the assessed inventory, a rise in blocked or escalated interactions, or a different set of models appearing in use.

Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with actions such as allow, redact, block, rate limit and escalate for human approval, and it generates continuous audit-ready evidence from that enforcement, mapped to frameworks including the EU AI Act and NIST. Its scope is individual interactions and tool calls. Those records can be one input to triage.

These signals support the reassessment decision but do not replace it. A new tool call is a prompt to compare against the baseline, and a person still decides whether the change is material. Runtime evidence also cannot show intended use or user population changes that never appear in an interaction, so those categories still depend on change management and business owners.

Governance and audit readiness

Keep the trigger criteria in the management-system process, not only inside individual assessments. That way the same rules apply to every AI system, and the decision records sit with other 42001 management-system records. Record how each assessment relates to those records and to any NIST AI RMF mapping, so that when a trigger fires, all of them are updated together and remain consistent.

As of October 2026, the European Commission states that the enforcement powers of the AI Office and national competent authorities under the EU AI Act apply from 2 August 2026, with provisions applying progressively. Records showing which changes were evaluated, and why, are the kind of material that makes a program straightforward to explain to an authority or auditor.

A useful test is whether you can list every change evaluated in the last 12 months, including those judged immaterial, and show the rationale and evidence for each. If that takes more than a query against one register, the process needs tightening before the next audit.

From change to recorded decision

Detect

A change is flagged by change management, incident records, regulatory monitoring or runtime signals.

Triage

The change is compared with the recorded assessment boundary.

Decide

Full reassessment or documented review, with a named decider.

Retain

Assessment version, change record, rationale and evidence are linked.

Connect runtime evidence to your reassessment decisions

Talk to an expert about how policy enforcement in the execution path and audit-ready evidence can feed your change-review process.

Talk to an Expert