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.
A regional operations lead receives a vendor recommendation on a Tuesday afternoon. The AI assistant has read the three bids, compared them against last year's contract and produced a four-page analysis recommending Vendor B, with a cost saving of 14 per cent and a summary of risks. The lead has a budget meeting on Thursday. This scenario is illustrative. It is not a client, and the numbers in it are invented.
What the lead does next can be described as one thing, judgment, or as nine. Describing it as one thing gives you nothing to develop, compare or check. AIRS-J/1, the AIR APAC Standard for Judgment in AI-Enabled Work, describes it as nine, and this article walks the Tuesday afternoon through all of them.
| Domain | Behaviour | Level 0 looks like | Level 2 looks like |
|---|---|---|---|
| Structure | S1 Problem Framing | Treats "summarise and recommend" as the decision | States the actual decision, who owns it and what it is not |
| Structure | S2 Scope and Constraint Definition | Treats all information as available and all options open | Names what is fixed, what is in scope and one unknown that must be resolved |
| Test | T1 Evidence Verification | Acts on an unchecked load-bearing claim, or rejects everything unexamined | Checks the load-bearing claim against a source before acting |
| Test | T2 Assumption and Alternative Testing | Takes the first explanation as the explanation | Names an assumption and a rival explanation and tests between them |
| Test | T3 Risk and Failure Anticipation | Names no failure, or only the risk of not acting | Names a failure, how it would happen and who pays |
| Engage | E1 Decision Rights and Escalation | Acts outside their authority, or delegates what cannot be delegated | Decides within authority and escalates what needs another's approval |
| Engage | E2 Stakeholder Alignment and Disclosure | Omits a party entitled to know, or hides AI involvement | Names who to consult before and inform after; discloses AI use |
| Perform | P1 Proportionate Commitment Under Uncertainty | Does not commit, or commits against the evidence | Commits in time, at a cost proportionate to stakes, with a stated basis |
| Perform | P2 Consequence Ownership and Correction | Blames the tool, the data or the brief | Owns the outcome; holds or reverses on the merits when evidence changes |
Source: AIRS-J/1 v2.2, section 7.5, condensed. The Level 0 column is drawn from the anchors and their stated common patterns. The full anchors for all five levels are public.
Problem Framing (S1) asks what decision is being made now. The lead was asked to review a recommendation. The decision on Thursday is whether to commit next year's regional logistics to a new vendor. Those are different, and the second contains a question the AI was never asked: is the option set right? Perhaps the incumbent should have been invited to rebid. A lead who presents "approve Vendor B" as the decision has restated the delegated task. That is Level 1. Level 2 states the decision, who owns it and what it is not, before investigating. Level 3 adds what part of the decision cannot be handed to anyone, including the assistant.
Scope and Constraint Definition (S2) asks what is fixed and what is unknown. The contract start date is fixed. The 14 per cent saving assumes volumes hold, and the lead does not know whether the largest customer renews in March. Level 2 names that unknown as one that must be resolved before deciding. Level 3 does something more useful: it says what information would be sufficient to proceed. "Confirmation from the account manager on the March renewal" is sufficient. "More data on volumes" is a wish.
Evidence Verification (T1) is where most AI-related failures land. The analysis says Vendor B's on-time delivery rate is 97.4 per cent. Where did that come from? If the assistant took it from the bid, it is a claim by the vendor. If it took it from the reference checks, it is retrieved. If it produced the number to complete the table, it is generated. Level 2 checks the load-bearing claim against a source and distinguishes generated content from retrieved content. Level 3 calibrates the effort: the delivery rate is load-bearing for a year-long contract, so it gets a call to a reference; the formatting of the bid summary is not, so it gets a glance.
The surprise in T1 is the inverse case. A lead who says "I don't trust AI analysis" and throws the four pages out unread has not verified anything. The Standard scores wholesale rejection without examination at Level 0, the same as wholesale acceptance. Both skip the check.
Assumption and Alternative Testing (T2) asks what the recommendation rests on. The 14 per cent saving rests on the assumption that service levels are equivalent. The competing explanation for Vendor B's lower price is that it is buying the account and will reprice at renewal. Level 2 names both and takes an action that could tell them apart, such as asking Vendor B for its renewal terms in writing. Level 3 names in advance what would overturn the lead's own preferred position, and then acts on it when it appears. Naming the falsifier and ignoring it when it arrives scores 1, not 3.
Risk and Failure Anticipation (T3) asks how this goes wrong and who pays. Level 1 says "there is a risk to service". Level 2 says the risk is a missed peak-season delivery window, the pathway is Vendor B's untested capacity in the northern depots, and the cost falls on the three largest customers and the regional sales team. Level 3 separates reversible from irreversible failure and says which harm is being accepted rather than mitigated. Naming an accepted exposure is not a weakness. Failing to name it is.
For trainers and facilitators: the three Test behaviours are the easiest to run as a session. Give a group a real AI-drafted analysis from their own work and ask three questions in order: which claim carries the decision, what does the recommendation assume, and how does this fail. Scoring is not the point. Seeing the three as separate acts is.
Decision Rights and Escalation (E1) asks what the lead may decide alone. The regional budget authority covers contracts to a certain value. Vendor B's contract is under it. Level 2 decides within that authority and escalates the one part that needs another signature, which is the change in payment terms, to the finance director by name. The failure that catches careful people is at the other end. A lead who sends the whole recommendation up the line "for visibility" has transferred the decision without transferring the reason. The Standard scores both failure directions, acting outside authority and over-escalation, at Level 1 or 0. Escalation is not a safe default. It is a decision with a cost, and the cost is that the person with the context no longer holds the decision.
Stakeholder Alignment and Disclosure (E2) asks who is entitled to know. The incumbent vendor's account team must be told after. The customer whose renewal is in March should arguably be consulted before, because the vendor switch affects their delivery schedule. And the budget paper for Thursday should say that the analysis was AI-drafted and that the delivery-rate figure came from the vendor's own bid. Level 2 does all of that unprompted. Disclosure of AI involvement is scored at Level 2, not as a bonus; leaving it out where a recipient would reasonably want it is Level 0. Level 3 fits each message to what its recipient has to decide, so the finance director does not need a second round of questions.
For HR and L&D: E1 and E2 are the two behaviours most tied to the organisation's own documents. If a decision-rights list and a disclosure expectation exist, they can be put in front of people. If they do not exist, no training on judgment will supply them.
Proportionate Commitment Under Uncertainty (P1) asks whether the lead commits in proportion to the evidence and the stakes. The lead has until Thursday, the March renewal is unconfirmed, and the saving is real if the volumes hold. Level 2 commits within the window to a position whose cost and reversibility match the evidence: for instance, award Vendor B a one-year contract with a volume-adjustment clause, and state why. Level 3 names in advance the condition that would change the position: if the March account is lost, the volume clause triggers and the award is reviewed.
The surprise in P1 is that caution is scored. A lead who recommends staying with the incumbent "to be safe", when the incumbent's price is higher and the switch is reversible, has not shown judgment. The Standard's note is blunt: recommending the most cautious option regardless of stakes is not judgment, it is the absence of it. Proportionality cuts both ways. Minimal care where harm is severe fails, and so does maximal caution where harm is minor.
Consequence Ownership and Correction (P2) is about what happens once the outcome or new evidence arrives. Suppose Vendor B misses the peak window in November. Level 0 says the AI analysis was wrong, or the vendor lied, or the brief was unclear. Level 2 accepts the outcome, states which new evidence changes the position and acts on it. Level 3 does something harder: it separates the poor decision from the unlucky outcome. If the reference check was skipped, the reasoning was wrong and must be corrected. If the reference check was done and the vendor's capacity collapsed for a reason nobody could have seen, the reasoning was sound and should not be abandoned. Holding a sound position under challenge is Level 2 behaviour. Changing position because someone senior pushed back, with no new evidence, is Level 1.
Rejecting the answer unread and accepting it unread are the same failure. Neither one checked.
Every behaviour has a Level 4, and the scale defines every Level 4 the same way: the person leaves a reusable rule, condition, indicator or check that would govern the next decision of this class. For T1 it is a check that persists, such as "vendor-supplied performance figures are always confirmed with a reference before they enter a budget paper". For E1 it is a condition under which the authority boundary itself should be reset. For P2 it is a statement of what in the person's own process produced the error.
Level 4 is not seniority. The Standard says so in its own text and treats any anchor that would need positional power as defective. A final-year student on the Campus track can reach Level 4 on a behaviour by stating a transferable rule. A director can sit at Level 1 by escalating everything. Nothing about the scale rewards rank, and that is what makes it usable across a team where rank and judgment do not line up.
The Standard reports nine behaviours and states that nine is a reporting hypothesis, not an established structure. The evidence programme watches whether each behaviour separates from the others in its domain. If T1 and T2 turn out to be one thing in practice, or S2 proves a facet of S1, they will be merged in the next major version. The model is built so it can shrink to four domains without breaking the certification rules.
That caveat is printed in the Standard rather than hidden in a technical annex because a standard that cannot say how it might be wrong is not a standard. You are being asked to read nine behaviours that AIR APAC has reason to think are distinct, with a published rule for what happens if they are not.
The Tuesday afternoon was ordinary. What the lead did with it was nine separate things, and each of them can be seen, named and developed.
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 scenario in this article is illustrative and invented. Walking through it is not an assessment and produces no level for anyone. Nine Dimensions is a reporting hypothesis; the Standard says it will reduce the number if the evidence requires it. Certification is by separate assessment on the official instrument, and practice never awards it.
AIR APAC publishes these articles in the institution's voice. Every figure carries its source and its limit. All articles.
If your reviewers approve 99.8 per cent of AI outputs, that number is a risk signal, not proof of control. Singapore's agentic AI framework now suggests that override rates be tracked. The clinical evidence shows why.
Singapore's agentic AI framework asks for human approval checkpoints. Korea's AI Basic Act asks for documented oversight. Both assume you hold a list of decisions a machine, or a junior, may not take. Most departments have the policy and not the list.
"Keep a human in the loop" is the most repeated line in AI governance. The loop breaks when the person rubber-stamps, has lost the skill, has no authority to stop, or has too much to check. Only one of those is fixed by training.