The AI has been rolled out, and still nobody uses it
Training done, instruction issued, tool in place. Nobody uses it. Where the chain breaks, what the standard requires and what evidence an auditor wants to see.
In short
If nobody uses an AI tool despite training and instruction, one link in the chain has usually broken: channel, example, conviction, habit or follow-up. Training alone does not satisfy the standard. ISO 9001 and ISO/IEC 42001 require in clause 7.2 that the effectiveness of the actions is evaluated. That evaluation is exactly what an auditor wants to see.
Auf Deutsch lesen: deutsche Fassung
The training has taken place. The instruction has gone out. The tool is rolled out and paid for. And still nobody does it. I hear this sentence in companies more often than any other, and it is almost always followed by the wrong diagnosis: people do not want to.
Mostly they do want to. The chain has simply broken at one point, and nobody has looked to see where.
The chain that nobody wrote
There is a formula for this that hangs in almost every leadership training course: said is not heard. Heard is not understood. Understood is not agreed. Agreed is not applied. Applied is not retained.
It is usually attributed to Konrad Lorenz. That is not correct. The chain is first documented in 1986 in an Austrian trade journal, without an author, as a rule for managers. It is only attributed to Lorenz around twelve years later, without any source reference. Before and after that, a Viennese physicist (1991), a sales trainer (1996) and most recently Paul Watzlawick (2019) were also named as the originator, depending on who happened to be quoting.
I do not mention this to be a know-all. A statement about communication that has been passed on unchecked for forty years and changes its sender three times in the process proves its own point better than any example. Many have heard it. Almost nobody has looked it up.
The chain itself remains useful all the same. Not as a quotation, but as a checklist. It has five links, and each breaks for a different reason.
Which link breaks with AI
Take a perfectly normal introduction: new AI tool, a policy to go with it, training for everyone. Here are the five breaking points, each with the sentence by which you recognise it.
- Said is not heard. The email went to two hundred addresses, eleven opened it. The information was sent. It was not received. Recognisable by: but we did communicate that.
- Heard is not understood. Everyone knows there is an AI policy. Nobody knows whether their case falls under it. Recognisable by the question: am I allowed to copy the customer’s quotation in there now or not?
- Understood is not agreed. They have understood it and think it is wrong. Whoever takes the rule for distrust bypasses it cleanly and quietly. Recognisable by: I will carry on as before, that worked too.
- Agreed is not applied. They think it is right and still do not do it, because the moment is missing in everyday work. Recognisable by the old route being three clicks closer than the new one.
- Applied is not retained. It runs for four weeks, then it slips back. Recognisable by the fact that nobody ever asked again after the introduction.
Link three deserves particular attention with AI, because it breaks more often here than with other tools. With a new machine nobody argues about the point of it. With an AI tool a question resonates that is rarely asked openly: is my work being measured here? Whoever has this question in mind and gets no answer is formally in agreement and in practice not. That is why every introduction needs a clear sentence on what the usage data is used for and what it is explicitly not used for. Without this sentence you are working against a suspicion you will never get to hear.
The difference is not academic. Something different helps at each link. At link one a different channel helps, at link two an example from people’s own working day, at link three a conversation about the why, at link four a shorter route in the tool, at link five simply asking. Whoever trains again at link four has missed the diagnosis and pays twice.
A practical case, anonymised. A manufacturing company introduces an AI tool for quotation texts. Training of ninety minutes, eleven people, all present, good mood. Eight weeks later two are using it. The assumption in the company: scepticism towards AI. The actual cause: the tool sat behind a second login, and the old template in the ERP was quicker to reach. Link four, not link three. After the switch to a call directly from the ERP, nine of eleven used it. Not a single further training session was held.
What the standard requires is not the training
Now the part that counts in the audit, and that maps the chain surprisingly precisely. Both standards separate two things that tend to merge in a company.
Competence is in ISO 9001 under 7.2, in the 2015 edition as in the new 2026 edition, and in ISO/IEC 42001 under the same number. The requirement is to determine the necessary competence, to ensure that the persons are competent on the basis of education, training or experience, where applicable to take actions and evaluate their effectiveness, plus documented evidence. 42001 puts it in almost the same words, relating to persons whose work affects the organisation’s AI performance.
42001 attaches a further note to its competence clause that is easily overlooked: the note refers, for implementation in human resources, to Annex B.4.6, including the question of which expertise is required at all. That is the difference between one training course for everyone and the question of who has to be able to do what for which role. Whoever operates an AI system needs different competence from whoever checks its results, and both need different competence from whoever purchased it.
Read the third point again. The standard does not require that training took place. It requires that someone checked whether the action worked. That is exactly the question that sits at links four and five of the chain.
Awareness is in 7.3 and means something different from competence. Persons must be aware of the policy, of their own contribution to the effectiveness of the management system and of the implications of not conforming, plus the relevant quality objectives, and in the 2026 edition also the quality culture and ethical behaviour. In 42001 the AI policy from 5.2 takes the place of the quality policy.
That is the normative version of links two and three. Whoever does not know their own contribution has not understood. Whoever does not know the consequences does not agree, because they take the rule for bureaucracy. So the standard named both breaking points long ago, only without the neat mnemonic.
An attendance list proves presence. Presence is the weakest link in the whole chain.
What an auditor wants to see instead
In practice it mostly goes like this: the auditor asks for the evidence, the folder with the signature lists arrives, and the topic is closed for both sides. But the list only satisfies the evidence of competence under 7.2 by half. It shows that an action took place. It does not show whether it worked.
Two answers to the same audit question, to make the difference visible:
- Weak: all employees were trained on the AI policy in March, here are the signatures.
- Robust: everyone was trained in March. In May we asked four people from three areas what they do when a customer sends design data along with an enquiry. Three of four could answer, the fourth could not. We then added the case to the quick guide and discussed it in the team meeting. Since June we have been asking this in the quarterly round.
The second answer costs half an hour and fulfils 7.2 c) completely, including the evaluation of effectiveness. It also ends the discussion about whether the training achieved anything, because the answer comes with a number.
With AI tools there is one additional piece of evidence that does not exist elsewhere and that reveals a surprising amount: the number of cases in which someone rejected or corrected an AI result. If this number is zero, either the tool is perfect or nobody is checking any more. The second explanation is the more likely one, and it is at the same time the evidence that human oversight exists only on paper.
Usable evidence of effectiveness is as a rule already there, you only have to treat it as such: usage figures of the tool, a sample of work results, three conversations at three workplaces, the number of queries to IT, and particularly meaningful for AI, the number of reported cases in which someone rejected a result. Whoever never rejects anything is not checking.
Where the chain breaks in your company, and what that costs
The diagnosis is cheaper than most people think. Pick three people who are supposed to use the tool and ask exactly one question: what is stopping you? Experience shows that the answers fall into the five links, and mostly they cluster at one.
- I did not know, it probably came by email. Link one, a channel problem.
- I do not know whether that applies to my case. Link two, an example problem.
- I think that is excessive. Link three, and that is the only one where a conversation is needed.
- I forget it in day-to-day business. Link four, a route problem.
- I had been doing it, then the mountain of orders came. Link five, a reminder problem.
And now the flip side, because the chain can also be read in the wrong direction. Non-use is not automatically a finding about the people. It can be a finding about the tool. If an AI tool costs more time in the real workflow than it saves, not applying it is the commercially correct decision, and the workforce has done the sums faster than the project.
That belongs in the audit just like everything else. A management system that lists an action as effective which nobody in the company applies is listing a claim. Whoever instead documents that effectiveness was checked, that it did not materialise and why, is in a better position than anyone with the thicker training folder. That is improvement in practice, and that is exactly what the standard wants to see.
The practical conclusion is unspectacular. Plan the effectiveness check at the same time as you plan the training. An appointment six to eight weeks later, three questions at three workplaces, ten lines of notes. That costs less than an hour, fulfils a requirement of the standard that otherwise remains open, and prevents the most expensive variant: training a second time where the chain had not broken at all.
Primary sources
Frequently asked questions
Is an attendance list enough as evidence of training?+
For half the evidence, yes. ISO 9001 requires in 7.2 documented information as evidence of competence, under item d) in the 2015 edition and as a separate sentence without a letter in the 2026 edition, and an attendance list shows that an action took place. But 7.2 c) additionally states that the effectiveness of the actions taken is to be evaluated. ISO/IEC 42001 puts it the same way in its clause 7.2. Without such an evaluation half is missing, and that is one of the most common findings in this clause.
How do I know which link of the chain is the problem?+
Ask three people who are supposed to use the tool: what is stopping you? The answers sort themselves almost automatically. I did not know means a channel problem. I do not know whether it applies to me means a missing example. I think it is excessive means a question of conviction. I forget means the old route is shorter. I did it, then stress came means nobody follows up. Something different helps for each of these answers, and further training helps with only one of them.
Do I have to evaluate the effectiveness of every single training course?+
The standard does not tie this to every event, but to the actions taken to acquire a missing competence. Proportionality is allowed: for a mandatory briefing a simple check is enough, whereas for the introduction of an AI tool that intervenes in quotations or customer communication the check should be more serious. What matters is that there is one at all and that it is documented.
What if employees deliberately do not use an AI tool?+
Then check the tool first, not the people. If it costs more time in the real workflow than it saves, non-use is a correct decision and a finding about the introduction. For the management system this is not a problem as long as you document it: checked, not effective, cause, decision. A system that lists an action as effective which nobody applies is considerably more open to challenge than one that contains an honest nil return.
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.