DEV Community

Do Nghia
Do Nghia

Posted on

What ISO 42001 audit prep actually takes: 100+ AI projects from the provider side

Most teams treat ISO 42001 prep as a policy writing exercise. Assign someone
to draft an AI policy, collect signatures, wait for the auditor. I sat on the
other side of that table, and the audit does not only test your documents. It
tests whether you know your own systems.

I walked a certification body through a system with 100+ operating AI
projects. I did it as the AI Provider Representative for ISO 42001:2023, with
a team of 20. I standardized 10+ AI project documents and led 3 teams through
audit prep. Here is what that work demanded.

The first question is a count

Before any control, before any policy, an auditor needs one number: how many
AI systems do you operate?

We ran 100+. Listing them sounds trivial. It is not. Projects get built by
different teams, under different names, with overlapping parts. A
recommendation model here, an LLM feature there, an internal automation
nobody calls AI. By the time you finish counting, the number has moved.

That count carries the whole standard. Every control you implement points at
a system: impact assessments, documented responsibilities, monitoring. If
your inventory is stale, every control sits on sand.

My stance: audit readiness is an inventory problem before it is a compliance
problem. If you cannot list your AI systems in one meeting, you are not ready
to write a single policy.

Someone must own the system

ISO 42001 asks the provider to operate an AI management system. That system
needs a person. In our case that was me, the AI Provider Representative, with
a team of 20 as the spine of the effort.

The role is not ceremonial. The representative decides which documents count
as evidence, which template each project fills, and who signs what. When two
teams disagree about scope, the representative makes the call and records it.
An audit goes badly fast when nobody owns the management system. The auditor
asks who to talk to, and a room of shrugs ends the conversation.

One person with authority beats a committee with a shared inbox. That is not
a management preference. It is what the audit needs: a named point of
accountability.

Standardize the documents before you collect them

We had 100+ projects. Left alone, each team documents its work its own way.
One project keeps a design doc in a wiki. Another has a risk note in a slide
deck. A third has nothing written because the code is the documentation.

You cannot audit that. The audit compares evidence across projects. If every
project formats its documents differently, comparison is impossible, and prep
turns into archaeology.

So we standardized. 10+ AI project documents, one set of templates across the
whole system. Each template maps to something the auditor wants to see: what
the system does, what data it uses, who owns it, which risks were accepted,
what monitoring runs.

This changed the shape of the work. Instead of a bespoke document per
project, teams filled 10+ shared templates. Prep stopped being a writing
project and became a filling project.

Templates beat heroics. Ten documents, consistently filled, read better than
a hundred essays nobody can cross check.

Rehearse the walkthrough

The defining moment of audit prep is not the document review. It is the
walkthrough: an auditor picks a system, and a person who built it explains it
live.

I led 3 teams through audit prep, and every rehearsal asked the same
questions. What does this system do? What data does it use? Who owns it? What
can it not do? What stops it when it misbehaves?

One question mattered more than the rest. Can you describe what your system
does in one sentence, without opening the code? Engineers who could, walked
in with the story straight. Engineers who could not, went back and wrote the
answer down. That exercise surfaced gaps no document review would catch,
because the gap was not in the documents. It was in the understanding.

Build the governance into the product

Policies written after the fact read like defenses. Constraints designed up
front read like engineering.

I saw this from both directions. On an AI cheating detection engine I led,
privacy and legal constraints were designed up front, before launch, because
the product judges people in real time. On an agentic e-commerce concierge,
the governance rule was one line: the agent may act when the action is
reversible and cheap to verify. The agent finalizes purchases inside user
constraints, and that rule is the guardrail behind every action it takes.

Rules like that shorten audit conversations. When a control matches how the
system actually behaves, evidence is easy to produce. When it does not, you
spend the audit explaining the distance between the policy and the product.

If you are starting ISO 42001 prep now, my sequence is short:

  1. Name one representative with authority. One person, named in writing.
  2. Count your AI systems. Write the number down and date it.
  3. Standardize your documents first. 10 or more templates, each mapped to a control.
  4. Rehearse the walkthrough. If an engineer cannot explain their system in one sentence, the prep is not done.

None of this needs a consultant to invent it. It needs someone who has
carried the pager. The certificate is a side effect of knowing your own
systems, not a paperwork trophy.

Links

Do Trung Nghia (Đỗ Trung Nghĩa) is an AI consultant and technical leader in
Vietnam. 6+ years shipping 40+ AI/ML solutions across telecom, finance,
blockchain, and cloud. Remote for global teams.

Top comments (0)