Skip to content
All articles
Audit practice 8 min read· by Lars Zimmermann

Sampling AI systems: how many cases are enough?

A sample that draws once audits chance along with the system. Why sampling AI has two directions and where the sample size really comes from.

In short

With a non-deterministic system the sample has two directions: breadth, that is how many cases you draw, and depth, that is how often you repeat the same case with identical input. Without repetition you are measuring chance as well. The size follows from the acceptable error rate that the organisation has defined in advance, not from the auditor’s gut feeling.

Auf Deutsch lesen: deutsche Fassung

In the classroom I like to hand this question back. Someone has just explained that he audited the AI-assisted quotation review: twenty cases drawn, all in order, box ticked. Then I ask how often he drew each of these twenty cases. Usually there is a pause. Once, of course. That is how it is done, after all.

For a booking entry, an invoice or an inspection record, once is indeed right. The record is fixed. It does not change while you are looking. With a language model it does change. The same input does not necessarily produce the same output. Anyone who draws once has not audited the system, but one draw from it.

I build AI systems myself and I audit them. So here is the reasoning behind the sample size, and above all the question you have to ask first, before any number can come out at all.

With AI the sample has two directions

With a classic audit object the sample has one direction: how many elements do you draw from the population? Twenty of a thousand invoices, and each invoice is one observation. With a non-deterministic system this one direction splits into two.

  • Breadth: how many different cases do you draw? It tells you whether the system holds up across the range of real inputs, including the rare special case that nobody provided for.
  • Depth: how often do you draw the same case with identical input? It tells you how far the results diverge for the same input. That is the quantity that simply does not exist for an invoice.

Twenty cases with one draw each give twenty observations of twenty different things. Twenty cases with five draws each give a hundred observations and in addition something that the single draw cannot show at all: the variation within a single case. That is exactly where the risk sits that later hits the people affected. The one applicant whose pre-selection would have turned out differently on the second run never shows up in a sample with one draw.

A sample that draws only once audits chance along with the system and then calls the result conformity.

Where the number comes from, and why it does not start with you

The most frequent follow-up question is: so how many? And the honest answer is that this number does not belong to the auditor. It belongs to the audited organisation, and it must already be there before anyone comes from outside.

ISO/IEC 42001 requires in Annex A under A.6.2.4, AI system verification and validation, that the organisation defines and documents verification and validation measures, together with criteria for their use. The guidance in Annex B.6.2.4 mentions test methodologies, the selection of test data, release criteria and acceptable error rates in this context. The associated guidance in Annex B expressly mentions reliability requirements including acceptable error rates as well as operational factors that can prevent a minimum performance level.

The size follows from this error rate. Anyone who says that one per cent wrong pre-selection is tolerable needs a considerably larger sample than someone who accepts ten per cent because a human reads every case anyway. Both can be right. What does not work is choosing the number afterwards.

And that brings us to the finding that is much more frequent in practice than any question about size: there are test results, but no release criteria defined in advance and no acceptable error rate. The result is assessed as sufficient after the fact. If you encounter that, your sample is not too small. It has merely been given a different purpose: it now proves that the yardstick is missing.

What ISO 19011 provides, and what it does not

The audit guideline ISO 19011:2026 deals with sampling in Annex A, clause A.6. It is the basis on which I plan, and it provides more than most expect. Clause A.6.1 describes the steps of sampling in order.

  • Establish the objective of the sampling.
  • Define the population from which the sample is drawn.
  • Select the method.
  • Determine the size.
  • Carry out the draw.
  • Evaluate, report and document.

The size is therefore only one step of six, and it comes fourth. Anyone who starts with it has skipped three decisions. The second in particular is regularly wasted with AI: defining the population. Very often I receive an export file, and only on enquiry does it turn out that it contains only the cases in which a human accepted the suggestion. The rejected and the aborted ones are missing. Anyone who draws from that is auditing the agreement and not the system. ISO 19011 expressly points out in A.6.1 that a sample from insufficient or inaccurate data does not provide a useful result; that also applies when the number of cases drawn looks impressive.

After that, ISO 19011 distinguishes two families: judgement-based sampling in A.6.2 and statistical sampling in A.6.3.

Judgement-based sampling relies on the competence and experience of the audit team and takes into account, among other things, the complexity of the requirements, the degree of technical and personnel change and previously identified risks. Its drawback is in the guideline itself and is readily overlooked: with this approach the uncertainty of the findings cannot be quantified statistically. You may sample this way. You just may not report the result as if it had been calculated.

The point about the degree of change deserves particular attention with AI. With a machine you know when something was last changed, because somebody used a spanner. With a bought-in language model the provider can swap the model version behind the same product name without anyone in the business signing anything. So if you use the intensity of change as a criterion for your sample size, you must not rely on the audited organisation’s change list. You have to ask how this organisation finds out at all that something has changed outside, and as a rule exactly that is the more interesting finding.

Statistical sampling distinguishes between attribute-based sampling, when there are exactly two outcomes per case, and variable-based sampling, when the result lies in a continuous range. This is also where the confidence level appears: a sampling risk of five per cent corresponds to a confidence level of 95 per cent, that is the deliberate acceptance that about one in twenty samples does not show the picture of the population.

And here lies the boundary that you have to close yourself. ISO 19011 thinks of the population as a set of existing records. An element is drawn and then is what it is. That a single element answers differently on a second look does not occur in this way of thinking, because with records it does not occur either. The depth of the sample is therefore not in any audit guideline. It has to go into your audit plan.

Attribute-based sampling fits until the boundary blurs

Attribute-based sampling presupposes that every case falls unambiguously into two outcomes: right or wrong, passed or not. For a form check that is trivial. For a language model exactly that is the real work, and it happens before the first draw.

An example from the pre-selection of job applications. The question must not be whether I as the auditor would have made the same decision. That is my opinion and not a criterion. It must be: does the result stay within the limits that the organisation itself has drawn, is it reasoned, and does the reasoning support the decision? I write this sentence down before the draw and have it confirmed by the business side. Otherwise we end up discussing every single answer afterwards instead of the system.

In vision inspection at the machine the boundary is sharper, because there is a physical good part against which you can measure. In quotation or document processing in the ERP it lies in between: an amount is right or wrong, an allocation often only more or less plausible. The blurrier the boundary, the more important depth becomes, because then the variation itself is the result.

How I actually do it in an audit

I never start with a number. I start with the question of which error rate the organisation has laid down for this system and who approved it. If a number with a date and signature comes back, the rest is craft. If none comes, I have my finding and no longer need the large sample at all to substantiate it.

After that I define breadth and depth separately and write both justifications into the audit plan. Breadth follows the variety of real inputs, not the convenience of the export function. Depth follows the question of how strongly the system varies at all: for a tightly constrained extraction with a fixed output format a few repetitions are enough, for freely worded assessment texts I need more. With every draw the model version, the state of the system prompt and the time are recorded, otherwise the repetition cannot be traced later.

If the variation within a single case is greater than the defined tolerance, the finding is not that one answer was wrong. It is that the system does not deliver a result within its own limits for identical input. That is a different finding, it concerns the effectiveness of the control and not the individual case, and it is considerably more uncomfortable for the organisation.

With bought-in systems, repeated draws often fail in practice: every call costs money, the production environment must not be burdened with test cases, or the provider simply does not let you. Then depth shifts from the draw to the record. You do not measure the variation yourself, you read it from ongoing operation.

For that you need logs that keep input and output together with the model version. ISO/IEC 42001 requires in A.6.2.8 that it be determined in which phases of the life cycle event logs are kept, and for high-risk systems Article 12 of the EU AI Act requires logging over the lifetime of the system, applicable from 2 December 2027 for systems under Annex III (Regulation (EU) 2026/1744). If these logs are missing, that too is a finding, and one that cannot be cured by a larger sample.

What does not work under any circumstances: silently leaving out depth and writing the report as if you had had it.

The sentence that has to be in the report at the end

An audit report that only says twenty cases were audited without nonconformity is not verifiable. It names neither the population nor the method, neither breadth nor depth, and it conceals the sampling risk that ISO 19011 specifically describes in Annex A.6.

I therefore write into every report what the sample was drawn from, how many different cases there were, how often each case was drawn, against which criterion defined in advance the assessment was made and with which model version at what time. Plus one sentence on what this sample cannot show. That sounds like more effort than it is, and it is the difference between evidence and a claim with a date.

If you want to know whether your AI systems have a defined error rate and the matching logs at all today, that is the first question a readiness check answers.

To the AI readiness check

Planning a sample for a non-deterministic AI system

  1. Ask for the acceptable error rate

    Have someone show you the release criteria defined in advance and the acceptable error rate under ISO/IEC 42001 A.6.2.4, with date and approver. If they do not exist, record that as a finding before you talk about size.

  2. Define the assessment criterion in writing

    Before the first draw, formulate when a result counts as correct, and have that confirmed by the business side. Without this sentence you end up discussing individual answers afterwards instead of the system.

  3. Define the population

    Define what the sample is drawn from, and be guided by the variety of real inputs, not by what the export function conveniently provides. Rare special cases expressly belong in it.

  4. Justify breadth and depth separately

    Determine how many different cases you draw and how often you repeat each case with identical input. Write both justifications into the audit plan; depth is not in any audit guideline and has to come from you.

  5. Draw with the model version recorded

    With every draw, record the model version, the state of the system prompt and the time. Without this information the repetition cannot be traced later and the evidence is worthless.

  6. Assess variation against the tolerance

    If the variation within a single case is above the defined tolerance, the finding is that the system does not deliver a result within its own limits for identical input. Whether an individual answer was wrong is irrelevant to that.

  7. Write the limits of the sample into the report

    Name the population, method, breadth, depth, criterion, model version and time, plus one sentence on what this sample cannot show. ISO 19011 mentions in A.6.1 sample size, selection method and confidence level as content that the report on a sample can take into account.

Share: LinkedIn E-Mail

Frequently asked questions

Are twenty cases enough for an AI audit?+

The question cannot be answered without the organisation’s acceptable error rate. The size follows from it, not the other way round. And with a non-deterministic system the number of cases is only half the answer anyway: without stating how often each case was drawn with identical input, it says nothing about the stability of the system.

Do I have to sample statistically or may I sample by judgement?+

Both are permissible under ISO 19011. Judgement-based sampling under A.6.2 relies on the competence and experience of the audit team. Its drawback is in the guideline itself: the uncertainty of the findings cannot then be quantified statistically. You may sample this way, you just may not present the result as calculated.

What do I do if the provider does not allow repetition?+

Then depth shifts from your own draw to the record of ongoing operation. You need logs that keep input and output together with the model version, as can be based on ISO/IEC 42001 A.6.2.8 and, for high-risk systems, Article 12 of the EU AI Act. If they are missing, that is a finding in its own right.

Where do I get the acceptable error rate from if the organisation has none?+

Nowhere. The auditor does not set it, otherwise they audit against their own yardstick. The absence is the finding: ISO/IEC 42001 A.6.2.4 requires criteria for verification and validation defined in advance, and the guidance on it mentions release criteria and acceptable error rates. If test results are only assessed as sufficient afterwards, the control is not effective.

Author & expert review: Lars Zimmermann · ISO/IEC 42001 Senior Lead Auditor & Senior Lead Implementer · ISO/IEC 27001 Lead Auditor & Lead Implementer (PECB)

Last updated: 10 November 2026. Researched and reviewed to the best of our knowledge; not a substitute for individual legal advice.

Questions about your own case?

In a free 15-minute intro call we assess where you stand on ISO 42001, ISO 27001 and the EU AI Act, honestly and without a sales pitch.

Continue reading