The most common ISO 42001 audit findings are an incomplete AI system inventory, a risk assessment that ignores AI specific issues such as bias and human oversight, missing impact assessments, and weak evidence of management review. Most come from policy that was written but not practiced. Fix those first.
This post explains what auditors tend to flag. It does not say any product certifies you, and Safeguard holds no ISO 42001 certification.
What is ISO 42001 and who audits it?
ISO/IEC 42001 is a standard for an AI management system, a set of policies and processes for governing how an organization develops, provides or uses AI. An accredited certification body audits you against it. Findings are graded, and the grade decides whether you certify.
One certification body's published guidance describes a major nonconformity as the absence or complete failure of a required element, which must be resolved before certification. A minor one is a single lapse that does not point to a systemic failure. It allows certification but needs a corrective action plan with evidence it was closed. That body does not grade each finding in advance, because the grade depends on what the auditor sees.
Which findings come up most?
The list below follows a published article from a certification body, with its clause references. Treat the clause numbers as that article's, and check them against your copy of the standard.
1. Incomplete AI system inventory (clauses 4.3 and 8.1). Third party services and AI embedded in other products are the usual omissions. A support tool with a built in model counts.
2. Generic risk assessment (6.1.2). The team reused its IT risk method and never asked about bias, explainability or human oversight.
3. Missing impact assessments (6.1.4 and 8.4). Some systems have none, or the ones that exist ignore effects on the people the system touches.
4. An AI policy with no responsible AI commitments (5.2). An IT policy with the letters AI sprinkled in.
5. No evidence of competence (7.2). The people running AI governance have no recorded training or experience.
6. Statement of Applicability gaps (6.1.3). Annex A controls are not all addressed, or exclusions are not justified.
7. No management review evidence (9.3). The meeting happened, or did not, but nothing shows required inputs and outputs.
8. Weak internal audit (9.2). Wrong criteria, no independence, or findings left open.
9. Third party AI not governed (8.1 and Annex A). Vendors are outside the risk assessment and supplier process.
10. Nonconformity process not followed (10.2). Incidents were not recorded, or fixes were never checked for effect.
Why do the same ones repeat?
Look at the pattern. Items 1, 2, 3 and 9 are about knowing what AI you have and what it can do to people. Items 5, 7, 8 and 10 are about proof that the system runs.
The inventory is where it starts. If the inventory is wrong, the risk assessment, the impact assessment and the supplier review are all scoped wrong.
How do you close the gaps early?
Build the inventory from several angles, not one survey. Check purchasing records, single sign on app lists, browser extension lists, code dependencies and model API keys. Shadow use shows up in expense reports before it shows up in a form.
Write the risk method for AI on purpose. Add questions on data provenance, bias, explainability, human override and what happens when the model is wrong.
Keep dated records. A management review with an agenda, attendees, decisions and follow up actions is worth more than a polished slide deck.
What to do
- Build an AI inventory from purchasing, identity, code and network sources, and name an owner for each system.
- Include third party and embedded AI, then tie each vendor to your supplier review.
- Write an AI specific risk method and run it on every system in scope.
- Complete an impact assessment per system, covering affected people.
- Walk every Annex A control for the Statement of Applicability and justify each exclusion in writing.
- Hold and minute a management review, then schedule an internal audit against ISO 42001 criteria with an independent auditor.
- Log AI incidents and near misses, and check that fixes worked.
Quick check
Can a major finding stop certification?
Yes. Major nonconformities generally must be resolved before certification is granted, while minor ones need a corrective action plan.
Is the AI inventory really the biggest risk?
It is the most frequently cited finding in the source used here, and the scope of several other documents depends on it.
Sources
- Common ISO 42001 nonconformities, Glocert International, fetched 18 October 2026
The source is a certification body's article, not the standard. Clause numbers and grading practice vary by body, so confirm them with your own certification body.
Last reviewed: 18 October 2026.
