Judgment was always built on the job, by doing junior work badly and being corrected. AI now does the junior work. Organisations are removing the apprenticeship and keeping the expectation.
A first-year analyst used to spend her first six months building the model wrong. She would bring it to a manager, who would find the error in the assumptions, send her back, and in the second version she would find it herself. By the end of the year she could tell a plausible number from a right one, and nobody had called that judgment. It was just what happened. Now the model arrives in four minutes from an assistant, the assumptions are already in it, and the manager who would have found the error is reviewing twelve such models before lunch. The analyst is not learning to build the model wrong. She is learning to approve it.
BCG stated it plainly in April 2026. In its analysis of how AI reshapes rather than replaces jobs, it wrote that the remaining roles "require contextual judgment, oversight, and coordination skills that are typically built through on-the-job experience". It added that higher skill thresholds can create barriers to entry. The sentence contains its own problem. If judgment is built on the job and AI is now doing the parts of the job that built it, the ladder is being removed from the bottom.
McKinsey's July 2026 piece is the closest mainstream treatment. "Building expertise in the age of AI: Who trains the next generation?" proposes what it calls an "answer key" model: the employee attempts the work first, the AI grades the attempt, and a manager discusses the differences. A narrowing gap, it says, is evidence that judgment is forming. Its main worked example is Bank of America, which is redesigning roles around AI. The practice is sound and the evidence below points the same way. What a narrowing gap cannot give is a common scale. It is a signal inside one team. It does not tell a second team, or an employer in another market, what level the person reached.
Gartner predicts the cost. In January 2026 it forecast that by 2030, 30 per cent of organisations will see worse decision-making because of overreliance on AI. The figure is a prediction, not a measurement, and should be read as a direction.
One randomised trial shows which kind of practice protects the skill. In 2025 a randomised trial of nearly a thousand secondary-school students in Turkey was published in PNAS. Some students were given an unguarded chatbot for maths practice, some a chatbot built to give hints rather than answers, and some no AI. Compared with students who had no AI, practice performance rose 48 per cent with the unguarded tool and 127 per cent with the guarded one. Then the tools were removed and everyone sat the same test. The unguarded group scored 17 per cent lower than students who had never had AI. The guarded group was not measurably different.
Students are not analysts and maths is not a market model, so the transfer is an argument. The argument is this: practice in which the tool supplies the answer produced a short-term gain and a later loss. Practice in which the tool withheld the answer produced a large gain during practice and no measurable loss afterwards. That is the same kind of difference as the one between an assistant that drafts the model and an answer key arrangement in which the analyst drafts it and the assistant grades it.
The practical implication for a development programme is a single design rule. The participant commits before the answer appears. Everything else, the content, the tool, the length, matters less than that one ordering. A programme that demonstrates what the tool can do builds familiarity, and familiarity is what the students with the unguarded chatbot had.
For trainers and facilitators: check every exercise in your programme against the ordering. If the participant sees the AI's answer before producing their own, the exercise is a demonstration. Reversing the order is usually a small change to the session and a large change to what it builds.
AIR APAC adds an observation that the Western literature does not contain, and labels it as an observation. In many Asia Pacific workplaces, a junior does not openly challenge a senior. That norm has costs and benefits and is not the subject here. What matters is how it meets a fluent AI output. An answer written in the register of a senior analyst, confident, structured, with a recommendation at the end, carries the tone of authority. A junior who has learned not to challenge that tone in a person has no separate training for challenging it in a tool.
This is not a research finding. AIR APAC has not measured it and no study reviewed for this article tests it. It is stated here because the organisations AIR APAC works with recognise it, and because a development programme in the region that ignores it is designing for a workforce it does not have. The practical response is the same as the design rule above, applied with more weight: the junior commits first, in writing, and the comparison with the AI's answer is the lesson. Deference to the tool cannot operate on a position the junior has already taken.
For HR and L&D: if the organisation's values statement rewards challenge and its practice does not, the AI will inherit the practice. The ordering rule is a way to make challenge structural rather than personal.
What on-the-job learning used to provide, by accident, can be provided on purpose. Scenarios in which the person decides, with consequence revealed afterwards. Feedback against a described standard rather than a manager's mood. A record of how the person decided, so the next scenario can be chosen to stretch the behaviour that is weakest. That is a designed apprenticeship, and the components already exist in good training practice. What has been missing is the described standard.
AIRS-J/1 supplies it, and its Campus track applies it at entry level. The Standard is single: the nine behaviours, the anchors and the certification thresholds are the same on every track. Tracks differ only in scenario complexity, stakes density and stakeholder count. The Campus track, for final-year students of institutes of higher learning and for graduates in their first two years of work, uses workplace scenarios at entry level, with bounded authority and bounded consequence. A Campus certification is recorded as such and is not convertible to a Professional one. To an employer it means the holder met every certification rule on decisions of the kind an entry-level employee makes with AI in the work.
Level 4 does not require authority, and the Standard says so in its own text. Level 4 requires only that the person leave a rule that would carry forward: a check, a condition, an indicator. It does not require the authority to impose it. The Standard treats any anchor that would need positional power as defective. A graduate who writes, after a scenario, "vendor-supplied performance figures are confirmed with a reference before they enter a recommendation" has written the kind of lasting check that Level 4 asks for on Evidence Verification, once the Level 2 and Level 3 requirements are also met. The transcript of it is hers to obtain.
The ladder was removed from the bottom. It can be rebuilt on purpose, and this time the rungs can be named.
AIR APAC publishes the Standard in full, operates the certification and the register, and stands behind the result. Trainers and delivery partners build and run programmes that develop the behaviours, against a rubric they did not write and may not vary. Certification outcomes are decided by the Standard's rules alone, never by what a candidate or employer bought. That separation is what makes the result worth something to an employer: the programme a person attended does not decide their result.
The credential belongs to the person. It survives a change of employer, and no sponsor or delivery partner may withhold or revoke it, including where they paid for it. Verification is free, by credential code, with no account. For a trainer, that means a participant who goes on to certify holds a public, portable record that the trainer does not control and does not need to. A trainer may tell participants that certification is open to them and where to book it, and may not imply that attendance makes a result more likely.
For trainers and facilitators: the Standard's anchors are public and are written to be used. A programme that maps its exercises to the nine behaviours and gives feedback against the published anchors may describe itself as aligned to AIRS-J/1 only if it meets the conditions in Section 20.1, and the trainer who delivers it need not hold an AIRS-J/1 certification. What a programme may not do is award a certifying result. Practice never does; a separate assessment on the official instrument does.
The junior work is being done by the tool. The judgment it used to build has to be built somewhere else, on purpose, and shown on a scale that outlasts the first job.
AIRS-J/1 assesses how a person decides, not what they know. A result is not a measure of intelligence, personality, AI technical skill or domain expertise, and it is not a substitute for supervision, professional licensing or regulatory authorisation. The observation about deference in Asia Pacific workplaces is AIR APAC's and is not a research finding. The PNAS trial concerns school students in one subject at one school; its application to workplace development is an argument. A Campus certification is not equivalent to a Professional one and the Standard records them separately. Level 1 is a course and ends with a completion record; the record is not an AIRS-J/1 credential, carries no result against the Standard and counts for nothing toward certification.
AIR APAC publishes these articles in the institution's voice. Every figure carries its source and its limit. All articles.
Two managers, the same AI output, two different decisions. Which was right, and would the Jakarta team have decided as the Seoul team did? Until judgment is scored on one published scale under like stakes, nobody can answer either question.
When an AI-assisted decision goes wrong, the first reflex is to blame the tool. The Standard scores that at Level 0. The second is to abandon every decision made the same way. That is Level 1. Good correction tells a poor decision from an unlucky one.
AI now produces the answer. It does not carry the consequence of acting on it. What is left for your people is a short list of decisions, and organisations have no common language for it.