How do the dimensions apply to a person who supervises AI agents?
AIRS-J/1 assesses how a person decides when working with, alongside or in supervision of AI systems. This page reads each of the nine dimensions against the supervision of an AI agent, then shows one fictional restock decision step by step.
All nine dimensions of AIRS-J/1 apply to a person who supervises an AI agent. This page gives AIR APAC's plain reading of each dimension, then one fictional decision.
Supervision is in scope
An AI agent acts on its own within limits that a person sets. This page reads section 5 for the person who sets the limits, checks the agent's work and answers for it.
Judgment is how well a person decides once AI has given the answer. AIRS-J/1 assesses how a person decides "when working with, alongside, or in supervision of AI systems" (section 2). The Standard does not use the word "agent". The first sentence of this section is this page's own meaning of "AI agent". The same dimensions, anchors, thresholds and decision rules serve all three uses of the Standard (section 2, "Three uses, one rubric"). The tracks do not differ in them either (section 13).
The nine dimensions, read for supervision
Names are from section 5. Quoted words are the Standard's. The rest is AIR APAC's plain reading.
Structure
- S1 Problem Framing. The person names the decision that the agent's work feeds, and "what part of the responsibility cannot be delegated".
- S2 Scope and Constraint Definition. The person sets what the agent may do, what is fixed and what is unknown. The person also sets "what information would be sufficient to proceed".
Test
- T1 Evidence Verification. The person traces what the agent reports to its source, and tells generated content from verified content. The Standard asks for "verification effort proportionate to consequence".
- T2 Assumption and Alternative Testing. The person names the assumptions the agent's set-up rests on and tests them, "including against their own preferred position".
- T3 Risk and Failure Anticipation. The person names how the agent could fail. The person also names "who bears the cost, and which harms are being accepted rather than mitigated".
Engage
- E1 Decision Rights and Escalation. The person acts within their authority and escalates what needs another party's approval. The person "does not delegate what cannot be delegated, to a person or to a machine".
- E2 Stakeholder Alignment and Disclosure. The person consults those who must be consulted and informs those who must be informed. The person also discloses "AI involvement, uncertainty, limitation and self-interest to parties entitled to know, without prompting".
Perform
- P1 Proportionate Commitment Under Uncertainty. The person commits to a step "at a cost and reversibility proportionate to the evidence and the stakes". The person also states the basis for it.
- P2 Consequence Ownership and Correction. The person accepts the outcome "without transferring it to the AI system". The person also "distinguishes a poor decision from an unlucky outcome".
One fictional decision, worked through
A composite case, written for this page.
1. The situation
Tran Van Minh leads supply planning at a pharmacy retailer. An AI agent, MinivetAI, places restock orders with suppliers within limits that his team set. Minh's weekly report goes to the regional operations meeting at 12:00 on Monday. An order placed in error costs money and shelf space, and stock is hard to send back once it ships.
2. What the AI produced
On Monday morning, MinivetAI's summary for the week reads: "All 38 restock orders placed and confirmed. Each order matches last week's sales. I am 94 percent sure."
3. The check
Minh does not rely on the summary alone. He compares the suppliers' confirmations with the agent's order list. The list shows 38 orders, and the confirmations show 39. One supplier confirmed the same order twice. The agent's log shows why. That supplier's system did not answer, so the agent sent the order again. The summary counted the order once.
Dimensions at work: T1 (the comparison) and T3 (the retry after no answer).
4. The decision, and who was told
Minh cancels the second order before it ships. He does not switch the agent off. He adds one limit to its set-up. Before any retry, the agent checks for an open order that is the same. Every retry is then held for a person to approve. Minh tells the supplier, and the managers of the two stores the order was for. In the weekly report, he states the error and the fix as his own decision.
Dimensions at work: S2 and E1 (the new limit) and E2 (telling the supplier and the store managers). P1 (a fix in proportion to the error) and P2 (reporting the error as his own decision) apply too.
5. Three months later
Retries are rare. One was held, checked and approved. The supplier's account shows no double orders.
The page gives Minh no level, band, composite or confidence. Levels and bands come only from an Assessment (AIRS-J/1, section 20.3).
What this does not claim
- The Standard does not certify skill with any AI system or agent.
- The nine readings on this page are AIR APAC's plain explanation. They add no criteria. The Standard's own wording is in section 5.
- The scene is fictional. Tran Van Minh, the retailer and MinivetAI are invented.
AIR APAC publishes the standard for judgment in AI-enabled work, certifies people against it through its official instrument, and stands behind every result.