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.
In July 2025 the Australian Department of Employment and Workplace Relations published a 237-page report it had commissioned from Deloitte under a contract worth about A$440,000. The report carried a quote attributed to a Federal Court judgment that the judgment did not contain, and it cited academic works that do not exist. Chris Rudge, a researcher at the University of Sydney, found the problems from outside. Neither the firm nor the department had caught them before publication. Deloitte issued a revised version, which disclosed that a generative AI system had been used in creating the report. It agreed to repay the final instalment of the fee. The department said the substance of the review was retained and the recommendations did not change.
Every part of that story is ordinary. A capable firm used a capable tool, the tool produced fluent text, and a person signed it. The consequence landed on the firm, in money and in public, and on the department, which now has to defend a report whose errors it did not catch. The tool did not sign the report. It could not.
AI systems now produce the answer: a draft, an analysis, a recommendation, a ranked list of options. The answer is often good, and it is sometimes wrong in ways that read exactly like being right. In neither case does the system carry what follows from acting on it.
That is a different situation from the one most management practice was built for. For most of the last century the scarce contribution of a skilled person was to produce the work. Judgment was exercised inside that production, so it was never named as a separate thing, and you could tell a good analyst from a weak one by the analysis.
Now the analysis arrives finished, from a system that writes with the same confidence whether it is inside or outside what it can do. The scarce and consequential human contribution has moved. It sits after the answer, not inside it.
The best evidence for this is still a 2023 field experiment with 758 consultants at Boston Consulting Group. Researchers from Harvard, Wharton, MIT and Warwick, with BCG's own researchers, ran it with the firm's staff. The consultants were split into two groups that did not overlap. In the first, 385 consultants did eighteen realistic consulting tasks that sat within what the model could do. Those using AI were faster and completed more. Their work was graded more than 40 per cent higher in quality. Those below the average benefited most.
The researchers also built one task to sit just outside the model's range, and a second group of 373 consultants did it. It was a realistic business case: it used interview notes and a spreadsheet and asked for a recommendation. On that task, consultants working without AI were right about 84.5 per cent of the time. Consultants with AI were right 60 and 70 per cent of the time, depending on the condition. Averaged, they were 19 percentage points less likely to reach the correct answer than colleagues working alone.
The experiment is often read as a warning about the tool. It is better read as a finding about the person. The consultants whose results fell when they used AI, in the authors' words, "tended to blindly adopt its output and interrogate it less". Consultants were assigned to the groups at random, so the gap is not explained by weaker people. On that task they did not test where the answer stopped being reliable, and that is the judgment that failed.
Three cautions apply. This was one firm, one model from April 2023, and one deliberately built task. The frontier has moved since and keeps moving. That is part of the problem, because a person cannot learn its location once and be done.
The firm in the first story and the firm in the second are both capable. Capability in the people and in the tool did not prevent either result. The task that caught people out was built to be as realistic as the others, and your team will not get a warning that the next one is different.
For HR and L&D: nothing in this result points to a shortage of tool skills. The missing skill was telling when to stop trusting the answer, and no attendance record shows whether a person has it.
For trainers and facilitators: a tool demonstration does not show this skill. The skill shows only when a person meets a task that looks like the others and has to decide whether to trust the answer.
If the answer is now produced, what is left for a person to do? AIR APAC's answer, published as a Standard, is that four decisions remain. They are decisions, not tasks, and the difference between doing each well or badly is observable.
Is this fit to act on? Not "is it well written" or "does it sound right", but fit for the decision in front of you at the stakes in front of you. That means knowing which decision that is, which is harder than it sounds, and testing the evidence in proportion to what rides on it. The Deloitte report was fit to read. It was not fit to publish under a professional firm's name.
Who is entitled to be told? Who must be consulted before a decision holds, and who must be informed after. Whether the recipient is entitled to know that an AI system was involved, and how certain the work is. Disclosure in the revised Deloitte report came after the problem was found. The question was live before the first version was sent.
What may not be handed over? To a system, to a junior person, or to a vendor. Every organisation has a list of decisions that must stay with a named human. Few have written it down, and fewer have asked whether the people holding those decisions can tell when they are about to hand one over without noticing.
What happens when it is wrong? Who owns the outcome, whether the position is held, corrected or reversed on the merits, and whether the person can tell a poor decision from an unlucky one. The firm that says "the tool did it" has already answered this question, badly.
The Standard measures this contribution through four domains of judgment: Structure, Test, Engage and Perform. Under them sit nine observable behaviours, each with a level from 0 to 4, and the levels are not seniority. Level 2 means the behaviour appeared without prompting and was adequate to the situation. Level 3 means it was calibrated to the stakes and recorded or stated so another person could check the reasoning. Level 4 means the person left a rule the next person could use.
These four questions are what accountability for AI-assisted work means in practice. If you cannot say who answers them for a piece of work that leaves your department, nobody is accountable for it. Each question is also something you can watch for in a decision review. "Was the framing right?" is a better question than "Did they use the tool well?"
For HR and L&D: the anchors are the development path. They describe what each behaviour looks like at each level, so you and the person can say what the next step would look like.
For trainers and facilitators: the four questions are a session structure. Each can be put to a group with a real piece of their own work on the table.
The tool did not change. The judgment about the tool was what failed.
None of this is a secret. BCG, one of the large consultancies, has pointed to judgment as the human skill that matters in two recent reports.
In November 2025, MIT Sloan Management Review and Boston Consulting Group published a global study of agentic AI drawn from more than 2,100 respondents. It describes workflows "blending human judgment and machine autonomy" and advises organisations to prepare people for "guidance- and judgment-oriented roles". In April 2026 BCG wrote that the roles that remain in an AI-reshaped workforce "require contextual judgment, oversight, and coordination skills that are typically built through on-the-job experience".
Both are fair statements, and both leave the reader where they started. Neither passage says what judgment is made of or what a competent level of it looks like. Neither says how a board would know whether two people in two countries have it to the same degree. Judgment is named as the scarce thing and then left as a word.
There is a reason it has stayed that way. Judgment was, until recently, bundled inside the work, and what you cannot see on its own you cannot describe on its own. Now the work arrives from a machine and judgment is all that is left. It is visible for the first time, and it still has no agreed description.
That matters for a practical reason that BCG's own line points to. If judgment is "built through on-the-job experience", and AI is now doing the junior work that was the experience, then the way organisations have always grown judgment is closing. You cannot manage a shortage of something you cannot describe.
A published definition does three things that a slogan cannot.
It gives a shared language. A manager can say "the framing matched the Level 1 anchor: she named the delegated task, not the underlying decision", and a colleague in another office knows what was meant. Today the equivalent sentence is "I'm not sure about her judgment", which tells nobody anything and cannot be acted on.
It gives a level. Not a trait or a personality, but a description of what the behaviour looks like when it is absent, partial, adequate, calibrated or transferable. A level can be developed towards. A reputation cannot.
It gives comparability. Two people assessed in different organisations, countries or industries can be read against the same anchors. That is what makes the result portable for the person and usable for the organisation. It is also what current records do not give. Training attendance, tool certificates and self-reported confidence are counts of exposure and feeling. They do not measure whether a person decides well with AI in the room.
AIR APAC has published that definition as AIRS-J/1, the AIR APAC Standard for Judgment in AI-Enabled Work. The whole rubric is public: the nine Dimensions, the anchors for each level, the evidence rules and the decision rules. Nothing applied in a certifying assessment exists outside the published Standard and the instrument specification it incorporates, and that is deliberate. A standard that a candidate, an employer or a regulator cannot inspect is not a standard. It is an opinion.
AIRS-J/1 certification is open. Sittings are arranged through your trainer or your organisation's learning and development team, who confirm your date. No sitting has run yet. Certification comes from a separate assessment on StepSim, its official instrument. Practice never awards it. StepSim is an AIR APAC initiative, built and operated by Exeter Labs, which the owner of AIR APAC also owns (Standard, section 4.9). Certification is not the first step, though, and this article is not asking you to take it. The first step is to read the definition and hold it against the decisions your people made last week with an AI answer in front of them.
The answer will keep arriving. The decision stays with a person, and the Standard is how you can see whether it was made well.
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. It is not a substitute for supervision, professional licensing or regulatory authorisation. The research cited here is about people overseeing AI output on tasks. None of it measures a person's judgment against the Standard directly, and the link between the two is an argument, not a finding. Nine Dimensions is a reporting hypothesis; the Standard says it will reduce the number if the evidence requires it.
AIR APAC publishes these articles in the institution's voice. Every figure carries its source and its limit. All articles.
As far as AIR APAC can establish, no regulator in the region has yet fined an organisation for acting on unverified AI output. The courts moved first. In Singapore the penalty landed on named lawyers, and the duty to check was held to be non-delegable. Courts in Victoria and Hong Kong criticised unverified AI material too.
Attendance, tool certificates and self-rated confidence are the three records most organisations hold on AI capability. None of them shows whether a person decides well with AI in the room.
Judgment sounds like a trait. In AIRS-J/1 it is nine observable behaviours, each with a level from 0 to 4. One ordinary scenario, walked through all nine, shows what Level 2 looks like and why Level 3 is different.