Words for AI work: AI-enabled work, supervision, agent, model-assisted
Four words are used in talk about AI at work and in a candidate's report. Each entry sets the Standard's own wording beside a plain reading, then says where a reader meets the word and what it does not mean. The Standard does not use one of the four.
Judgment is how well a person decides once AI has given the answer.
These four entries cover words that readers meet in news about AI at work and in a candidate's report. They are set beside the AIR APAC Standard for Judgment in AI-Enabled Work (AIRS-J/1, "the Standard"). Each entry gives the Standard's own wording first, then a plain reading, then where a reader meets the word and what it does not mean. A plain reading is AIR APAC's explanation in simpler words. It is not the text of the Standard.
1. AI-enabled work
The Standard's wording. Section 3, Terms, has no row for "AI-enabled work". Its meaning comes from the Standard's title, "AIR APAC Standard for Judgment in AI-Enabled Work", and from the scope sentence of Section 2:
"Assessment and certification of individual human judgment when working with, alongside, or in supervision of AI systems in a professional or pre-professional setting." (Section 2, In scope)
Plain reading. AI-enabled work is work in which an AI system produces or shapes an answer and a person decides what happens next.
Where a reader meets it. In the title of the Standard. It also appears in the scope statement of Section 2.2, which every report and credential carries in plain language (Sections 2.2, 12.5).
What it is not. It is not an amount of AI use, and it is not skill with a particular tool. "Technical proficiency with any specific AI tool, model or vendor" is out of scope for this version (Section 2). AIR APAC's entries on what the Standard means by AI-enabled work explain each part of the phrase.
2. Supervision
The Standard's wording. Section 3 has no row for "supervision". The Standard uses the word in four places, each in a different sense:
- "in supervision of AI systems" (Section 2, In scope).
- "is sat in person under supervision" (Section 9.1). This is a supervised sitting. Section 9.1 requires it for "An Assessment commissioned for a hiring decision", and sets when a remote sitting for hiring may open.
- "A substitute for supervision, professional licensing or regulatory authorisation" (Section 2.1). This is one of the things a credential is not: "Nothing in this credential authorises a holder to act beyond their existing professional or legal authority."
- "bounded authority with a named supervisor to escalate to" (Section 13, Campus track). This is a person inside a Campus scenario, to whom the candidate escalates.
Plain reading. The same word does four jobs in the Standard, so a reader first asks which sense is meant. Only the first, a person overseeing an AI system, is part of the scope sentence. The others are a supervised sitting, the oversight of a holder's own work, and a supervisor in a Campus scenario.
Where a reader meets it. In the scope sentence and in the rules for a sitting. It also appears in the table of what a credential is not, and in the Campus track.
What it is not. Supervising AI is not a separate track or a separate certification. The scope sentence names it beside working with and alongside AI systems, and the individual tracks use the same dimensions, anchors, thresholds and decision rules (Sections 2, 13).
3. Agent
The Standard's wording. The Standard does not use the word "agent". Section 3 has no row for it, and no section of the Standard contains it. The nearest wording is "in supervision of AI systems" (Section 2), and the description of E1 Decision Rights and Escalation:
"Acts within their authority, escalates what requires another party's approval, and does not delegate what cannot be delegated, to a person or to a machine." (Section 5, E1)
Plain reading. In everyday use, an AI agent is an AI system that takes actions as well as giving answers. This reading is AIR APAC's plain explanation, not a term of the Standard. Where an AI system can act, a person still has to decide what may be handed to it, and remains responsible for what follows.
Where a reader meets it. Outside the Standard: in news reports, in product descriptions and in everyday talk about AI at work.
What it is not. It is not a defined term of AIRS-J/1, and it does not name a track or a certification. How each dimension applies when an AI system acts is a separate question, which this entry does not answer.
4. Model-assisted
The Standard's wording. Section 3 has no row for "model-assisted". Section 11.4, "Rule-based, automated and model-assisted scoring", sets the rule:
"A model-generated (for example large-language-model) score is not an observation and never stands alone:
- A model may propose a level with cited transcript evidence.
- A qualified human assessor confirms or overrides, with reasons recorded.
- The pair counts as one observation, attributed to the human assessor.
- Model-assisted proportion is recorded on the result and disclosed in the register entry." (Section 11.4, four of its six points)
Section 12.5 lists "the model-assisted proportion" among the things every candidate receives.
Plain reading. In model-assisted scoring, a model suggests a level and cites the evidence, and a qualified human assessor confirms or overrides it with reasons. The model-assisted proportion on a report shows how much of the scoring was done this way. No model's opinion stands alone as a reading.
Where a reader meets it. On the candidate's report (Section 12.5). Section 11.4 also requires it to be disclosed in the register entry; this entry does not describe what a live register entry shows. In Annex C, the "Scoring mode" field of an observation record can read "model-assisted-human-confirmed".
What it is not. It is not rule-based scoring. Under Section 11.4, an approved instrument "applying published rules (Annex F) to the retained trace, generates observations directly", and no human confirmation is needed.
AIR APAC publishes the standard for judgment in AI-enabled work, certifies people against it through its official instrument, and stands behind every result.
What these entries do not claim: a plain reading adds no rule to the Standard, and nothing here states how much of any result was model-assisted.