The company trained the tool and not the judgment
Meta description: AI training that ends at features produces prompting without verification — competence begins with supervised mistakes.
The training numbers are excellent, and I have stopped being reassured by them. Completion rates above ninety percent. Prompting certifications across the management layers. A curriculum that vendors describe as comprehensive. Then I look at the decisions the trained organization makes on Monday morning, and they are identical to last quarter’s: same routing, same review depth, same judgment calls made the same way. The company trained a skill. The skill that was missing was a different one, and it was never in the curriculum, because the curriculum was built around what courses can teach.
My approach to AI education begins with a simple observation: a new technology often creates excitement before it creates understanding. Many people encounter AI through dramatic headlines or impressive demonstrations. They then try a tool, receive a plausible answer, and assume that the system will behave like an all-knowing colleague. The first disappointment can be disproportionate because the original expectation was unrealistic.
What does the current curriculum treat as the skill?
The curricula I review treat tool operation as the skill: how to prompt, how to iterate, how to use the features, which tool for which task. This is the purchasable curriculum — vendors teach it, platforms certify it, progress is measurable in modules. I do not dismiss it; tool operation is genuinely necessary. I notice what the curriculum’s shape assumes: that the gap between the workforce and AI value is an operating gap. In the organizations I visit, the operating gap closes quickly — people learn the tools in weeks — and the value still does not arrive. The gap was never where the curriculum aimed.
What skill does the failed decision actually require?
The skill courses skip is failure recognition: the ability to check answers, recognize failure modes, and inspect whether the output meets the purpose. I encourage people to begin with small tasks, test the system carefully, and compare the results with what they already know. A good early exercise is to ask AI for help in an area where the user has genuine expertise. The user can then inspect the answer: what is accurate, what is missing, what is confidently wrong?
That kind of supervised experience builds a more durable understanding than either blind trust or instant rejection. Confidence becomes competence through practice, not through a feature tour.
I can locate the missing skill by its absences. It is missing when a model’s confident output is accepted without the two-minute check the person’s expertise made possible. It is missing when a model’s uncertainty is accepted as fact because the output looked polished. It is missing when a decision that should have been escalated — because the stakes outgrew the tool — was processed routinely, because processing it routinely was easier. The missing skill is a judgment about the tool’s outputs: when to trust, when to verify, when to stop the line. It is exercised in seconds, dozens of times a day, and no course I have reviewed teaches it, because it cannot be taught in the abstract. It lives in the person’s domain expertise meeting the tool’s behavior.
Why is the wrong skill easier to purchase?
The false explanation is vendor quality: find a better training partner, a more rigorous program. I have watched companies cycle through vendors and repeat the outcome, because the purchasing logic itself selects the wrong skill. Procurement needs definable scope, measurable outcomes, scalable delivery — and tool operation supplies all three. Judgment training supplies none: it is specific to each function’s work, its outcomes resist module-level measurement, and it does not scale through a platform. The market sells what can be specified. The specification excluded the skill that mattered before any vendor was shortlisted.
What pattern marks mistargeted training?
The marker I use is the location of the post-training failure. After tool training, failures concentrate in judgment moments: over-trust, under-verification, inappropriate delegation to the tool. After well-targeted development, failures concentrate elsewhere — in genuine edge cases, in novel situations, in places where the failure teaches something. I walk through trained organizations reading incident reports, and the reports keep describing judgment failures in a workforce certified in operation. The certificate and the failure coexist in the same person, and neither the person nor the program finds that strange.
I compare learning AI to learning a new language. A person does not become fluent by memorizing a few impressive phrases. Fluency develops through repeated practice, correction, and gradually more difficult tasks. A training program that measures completion of a slide deck has measured attendance, not fluency. The organization should be able to see what learners can now do: check an output, revise an instruction, recognize when a result should not be trusted.
My motivation comes partly from my personality. I am at heart a helper and teacher. Long before AI became my daily work, I was interested in psychology, self-development, and the ways people make decisions. Over more than a decade, I spent considerable time learning about those subjects. AI did not replace that interest. It gave me a new context in which the same human questions became more important.
Miklos Roth
What question should define the curriculum?
Curricula are built from the question “what should people know about AI?” — a content question with a content answer. The question that would rebuild the curriculum starts from the decisions: which decisions, made daily by trained people, are currently made worse with the tool than without it? That question has answers, and the answers are specific, uncomfortable, and un-purchasable off the shelf. They are also the curriculum.
The question a training budget should be measured against: after the course, can each person check an AI answer in their own field and say what is accurate, what is missing, and what is confidently wrong?
PR excerpt (80–120 words, Miklos’s own words): Many people encounter AI through dramatic headlines, try a tool, receive a plausible answer, and assume the system will behave like an all-knowing colleague. I encourage people to begin with small tasks, test the system carefully, and compare results with what they already know. A good early exercise is to ask AI for help in an area where the user has genuine expertise — then inspect what is accurate, what is missing, and what is confidently wrong. Learning AI is like learning a new language: fluency develops through practice and correction, not memorized phrases.
By Miklos Roth — AI marketing consultant, Budapest. More: Miklos Roth — AI marketing consultant.
