Methodology 3.3.0
What the rating means—and what it does not
The AI Delegation Potential rating measures how favorable an occupation's meaningful work is for useful delegation to AI agents. It does not estimate the percentage of a job that can be automated, the likelihood that a job will disappear, or the probability that a worker will be replaced.
From technical score to qualitative rating
Every occupation retains an underlying 0–100 methodology score. The public rating is derived from that score using fixed thresholds. The same thresholds apply to all 923 scored occupations, and occupations with identical scores remain in the same category.
The underlying calculation
Each candidate workflow receives five 0–100 component scores, weighted equally. Before scoring, every candidate passes a workflow eligibility gate. Only agentic workflows are eligible for selection and scoring; up to three are aggregated using the strength of their O*NET task evidence, then adjusted by the share of meaningful work their eligible evidence addresses.
Coverage modifier = 0.80 + (0.20 × coverage share)
Qualitative judgment modifier = 0.90 artistic/editorial, 0.92 normative/policy, otherwise 1.00
Underlying methodology score = round(base score × coverage modifier × qualitative judgment modifier)
Coverage still reduces scores when agentic opportunities address only part of an occupation, but the adjustment is capped at 20% so strong, useful workflows remain visible even in highly human roles.
The qualitative judgment constraint applies when at least 35% of the selected workflow evidence depends on artistic intent, editorial taste, political interpretation, or normative policy recommendations. Quantitative analysis and rule-based software testing are not penalized merely because they involve professional judgment.
Workflow eligibility comes before scoring
Agentic workflow
AI owns a bounded digital outcome with monitoring, action, verification, and escalation. This can include documentation when the agent collects evidence, reconciles records, acts in systems, verifies completion, and routes exceptions. These workflows may be selected and scored.
Documentation support
AI may help with records, checklists, scheduling, coordination, or handoffs around human work. When the work is only drafting, transcription, or a single documentation step, it does not affect the score. If fewer than three credible agent workflows are available, one of these administrative ideas may fill an open recommendation slot instead of overstating what an agent could do. It is clearly labeled as documentation support.
Physical execution
Operating, moving, installing, repairing, cutting, cleaning, or otherwise handling physical things remains human work and is excluded from the published recommendations. The gate applies to the actions inside every workflow, regardless of occupation.
Protected clinical and veterinary actions
Physical examinations, medication administration, diagnosis, prescribing, vaccinations, treatment changes, and patient or animal disposition remain with licensed professionals. AI may prepare evidence or route exceptions, but it does not own those decisions or perform the physical care.
An occupation may therefore publish fewer than three delegable workflows—or none—rather than filling the list with physical or documentation-only work.
Five interpretable dimensions
Importance & frequency 20%
How central and frequent the supporting O*NET tasks are. Importance, relevance, and frequency are normalized and combined.
AI capability 20%
How feasible the reasoning and information-processing portions are for current AI.
Digital actionability 20%
How much of the workflow happens through software, data, browsers, APIs, or communication systems.
End-to-end leverage 20%
Whether an agent can own a useful loop—monitoring, deciding, acting, checking, and escalating—rather than one isolated task.
Safety & reversibility 20%
Whether bounded delegation is appropriate, errors are detectable, and actions can be reversed.
From government data to workflows
- 1. Normalize O*NET. Preserve O*NET-SOC codes, occupation titles, reported titles, task statements, task ratings, and task-to-DWA mappings from release 31.0.
- 2. Generate without scoring. Create exactly eight distinct outcome-level workflows for every eligible occupation, grounded in cited O*NET task evidence.
- 3. Evaluate independently. A separate pass scores the frozen workflows, estimates whole-occupation coverage, and applies conservative physical-work and safety constraints.
- 4. Select and rank. Classify all eight candidates, exclude physical execution and documentation-only support from scoring, and publish up to three eligible agentic workflows. Closed-loop documentation may qualify when the agent owns a verifiable digital outcome; only task evidence for its digital steps contributes to its score. Retain every candidate, its classification, task evidence, model and prompt versions, and methodology version for audit.
Limitations
Capability, coverage, and oversight values are modeled estimates and require human judgment. The current release scores the 923 O*NET occupations with task statements; 93 O*NET codes without task statements are excluded. Wage data are not used in the score, and physical-work and safety effects are modeled from workflow evidence rather than full ORS detail.
Independence and attribution
The Agentic Opportunity Score is an independent analysis of government occupational data. It is not a rating or endorsement by O*NET, the U.S. Department of Labor, or the Bureau of Labor Statistics.
O*NET 31.0 data are used under CC BY 4.0. O*NET® is a trademark of the U.S. Department of Labor, Employment and Training Administration. BLS Occupational Employment and Wage Statistics are used only as separate economic context.