An AI management system under ISO/IEC 42001 becomes four streams of engineering work: a live inventory of models and data, evaluation records, change control for models and prompts, and monitoring in production. A VP of engineering should build these as normal tooling, not as audit paperwork.
What does engineering own?
The standard describes an organizational system, so most clauses will be shared with legal, product and leadership. Engineering still owns the evidence that is hardest to fake: what runs where, what changed, how it was tested and what happened in production. As with any paid standard, read the actual text before you promise anything externally. We did not read it for this article, and what follows is engineering practice, not a clause by clause reading.
Stream one: inventory
You cannot govern what you cannot list. For each AI feature, record the model and version, hosting location, training or fine tuning data sources, input data categories, downstream decisions and the owning team. Include third party APIs. The OWASP Top 10 for LLM applications lists supply chain among its ten risks, so knowing which model and which dependencies sit behind a feature is part of risk management, not trivia.
Stream two: evaluation
Define what good means per feature before launch, then record results. That might be accuracy on a labelled set, refusal behavior, bias checks relevant to the use, or resilience to hostile inputs. OWASP's list names prompt injection, sensitive information disclosure, improper output handling, excessive agency and unbounded consumption among the risks to test against. Keep the test sets and results with a date and a version, so you can show how a decision was made.
Stream three: change control
Models change underneath you when a vendor updates a hosted model, when you retrain, or when someone edits a system prompt. Treat each as a release. Version prompts in the repository, require review for changes that alter behavior, and pin model versions where the provider allows it. Record who approved what.
Constructed example: a support summarizer begins to skip refund amounts after a prompt tweak made for brevity. No code changed, no test failed, and a week later finance notices. With prompt versioning and a small regression set, the tweak would have failed review.
Stream four: monitoring
Log inputs and outputs where privacy law permits, track failure categories, and watch cost and rate for abuse. OWASP's unbounded consumption entry is a reminder that resource abuse is a risk in its own right. Define who is paged for what, and what the rollback looks like for a model or prompt.
How does this relate to regulation?
The EU AI Act's own timeline, as tracked on artificialintelligenceact.eu, shows high risk obligations for Annex III systems from 2 December 2027 and for Annex I systems from 2 August 2028. Those obligations include things engineering will recognize, such as record keeping and human oversight design. A management system is a good place to house the evidence, but verify the legal specifics with counsel.
Who should own it?
Give each AI feature one accountable engineer and one reviewer outside the team. The owner keeps the inventory row, the evaluation results and the monitoring dashboard current. The reviewer checks changes that alter behavior. This small structure covers most of what an auditor will ask about roles, and it works because it sits inside the normal pull request flow. Avoid a separate AI governance committee that meets monthly and approves nothing in time to matter.
How to apply this
- Create a repository file or internal tool that lists every AI feature with owner, model, version, data and hosting.
- For each feature write a short evaluation plan and store results with dates.
- Put prompts and model versions under change control with required review.
- Add monitoring for quality, abuse and cost with named responders.
- Decide where AI evidence will live, and who exports it for an audit.
- Review the inventory every quarter.
Quick check
Does a vendor model update count as a change?
Yes, for governance purposes. It can change behavior without any change in your code, so record the version, test before switching and keep rollback options.
What is the cheapest first step?
An inventory of AI features with owner, model, data and hosting. It supports every other control.
Sources
- OWASP Top 10 for LLM Applications 2025
- AI Act implementation timeline
- NIST: AI Risk Management Framework
Last reviewed: 18 October 2026.
