When a decision involves AI, many managers choose the most cautious option and call it prudence. The Standard scores that at Level 1. Maximal caution where the cost is low is as much a failure as minimal care where the harm is severe.
A customer writes to complain that a delivery arrived two days late. The assistant drafts a reply: an apology, a credit of a modest sum, and a line saying the depot's routing has been reviewed. The team lead reads it and has three choices. Send it. Send it after checking that the routing was in fact reviewed. Or hold it, send the complaint to legal to confirm the credit creates no precedent, ask the depot for a written account, and reply next week. The third choice is the one many leads make, and the one they describe afterwards as being careful. This scenario is illustrative and not a client.
The symptom is documented. The cause is not named. BCG's 2025 survey of AI at work reports regular use rising among leaders and managers while it stalls among frontline employees, and names the next step as moving from adoption to value. Its agentic enterprise study with MIT Sloan Management Review, drawn from more than 2,100 respondents, reports that agentic AI is spreading across enterprises faster than leaders can redesign processes or assign decision rights. Both describe adoption running ahead of the redesign of work. BCG points to training, leadership support, access to the right tools and the redesign of workflows, and all four are real. What these sources do not say is that a large part of the gap is a pattern of individual decisions, made by people like the team lead above. In each, the most cautious option was chosen because it felt like the one nobody could criticise.
The pattern has a shape you will recognise. An AI use policy written to prohibit rather than to permit, so that any use is a risk and no use is a decision. A review queue in which every AI-assisted output waits for a sign-off that adds nothing because the signer has no time to check. A pilot extended for a third quarter because the evidence is not yet complete, when the evidence will never be complete and the cost of the pilot was never counted. Each is defensible on its own. Together they describe an organisation in which nobody can be blamed and nothing happens.
The cost is paid twice. Once in rework and delay on the decisions that are eventually made, because every one passed through a queue. And once, invisibly, in the decisions that were never made at all: the reply not sent, the process not changed, the product not shipped. The second cost never appears on a dashboard, which is why over-caution survives as a strategy. Its failures are absences.
AIRS-J/1 names the behaviour as dimension P1, Proportionate Commitment Under Uncertainty. The anchor at Level 2, the threshold for competence, reads: commits, unprompted and within the time available, to a position whose cost and reversibility are proportionate to the evidence and the stakes, and states the basis. Level 0 is not committing, committing after the window has closed, or committing to a position the evidence contradicts.
Level 1 is where the careful team lead sits. The anchor describes it as committing, but with a response not proportionate to the stakes: maximal caution where cost is low and harm minor, or minimal response where harm is severe or irreversible. The Standard's note is the sentence this article exists to deliver. Recommending the most cautious option regardless of stakes is not judgment; it is the absence of it. Proportionality cuts both ways. The person who under-reacts to a serious risk and the person who over-reacts to a trivial one are scored in the same band, because neither matched the response to the situation.
That symmetry is the point of the anchor. The Standard scores the failure in both directions, and in doing so it gives a name to a behaviour that has been invisible because it never produced an incident. It produced a queue.
For HR and L&D: P1 is the behaviour most often misread in performance conversations as a virtue. A person described as careful may be at Level 1 on this dimension. The anchor gives a way to say so that is about the decision, not the person.
Take the late-delivery complaint and write both answers. They are illustrative.
The over-cautious reply. The lead holds the draft. She asks legal whether a credit of this size could be read as admission of fault. She asks the depot manager for a written account of the routing review. She asks the assistant to produce three alternative replies with different tones. She schedules a review for the following Tuesday. The customer, who wanted an apology and a credit, waits eight days and receives a reply that has been through four people. The sum at stake was modest. Nothing about the decision was irreversible. The cost of being wrong was a second email. The lead's response was calibrated to a harm that was not present.
The proportionate reply. The lead reads the draft and asks one question: was the routing actually reviewed? The depot confirms in a message that it was. She sends the reply within the hour, with the credit, and makes a note that if three more late-delivery complaints arrive from that depot in the month, she will raise the routing with operations formally. The reply was cheap to undo, the evidence was adequate to the stakes, the basis was stated, and the condition under which the position would change was named in advance. That last element is the Level 3 anchor: a named condition, not a general willingness to reconsider.
The difference is not that the second lead took a risk. It is that she measured the risk and found it small. The first lead did not measure it, which is why her caution was not judgment. She applied the response she would apply to a product recall to a late parcel.
The person who over-reacts to a trivial risk and the person who under-reacts to a serious one are scored in the same band. Neither one measured.
One question separates the two replies, and it works on most decisions. Can this be undone, and at what cost? If a decision is cheap to reverse, it may be made on thin evidence, because the cost of being wrong is the cost of reversing it. If a decision cannot be reversed, the evidence must be proportionately stronger before commitment, and the time to gather it is worth spending. The Standard's T3 anchor at Level 3 asks the person to distinguish reversible from irreversible failure, and its P1 anchor at Level 2 asks for a position whose cost and reversibility are proportionate to the evidence and the stakes. The two work together.
Applied to the patterns above, the test cuts in a predictable direction. The prohibitive AI policy treats every use as irreversible, when most uses are a draft that can be discarded. The universal review queue treats every output as high-stakes, when most are not. The extended pilot treats shipping as irreversible, when a pilot that ships to one team can be withdrawn from one team. In each case the response was calibrated to the worst case in the category rather than the case in front of the person, and the Standard scores that as a failure of proportion.
The test also protects against the opposite error, and you should apply it in that direction too. A public statement, a contract termination, a decision about a named person's employment or health: these cannot be undone. The lead who sends an AI-drafted version within the hour on a single confirmation has failed the same anchor from the other side. The Standard does not favour speed. It favours proportion.
For trainers and facilitators: the two-replies exercise runs in thirty minutes with a group's own cases. Ask each participant to bring one AI-assisted decision they held and one they sent, then apply the reversibility test to both. The pattern that emerges in most groups is that the held decisions were the cheaper ones to undo.
Caution is a response, not a virtue. The Standard scores whether it was the right size, and in most organisations the queue says it was not.
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 complaint scenario and both replies are illustrative and invented. The claim that over-caution accounts for part of the adoption gap is AIR APAC's argument; the sources cited do not make it, and no study that measures it is cited here. Certification comes from a separate assessment on StepSim, its official instrument. Practice never awards it.
AIR APAC publishes these articles in the institution's voice. Every figure carries its source and its limit. All articles.
A tribunal held an airline to what its chatbot said. A consultancy's revised report disclosed its AI use, after errors were found in the first. The question in both was the same: who was entitled to know, and when.
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.
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.