Methodology

The AI Overlap Index (AOI) is a single 0-100 score per US occupation that combines four weighted signals. It is a directional descriptor of how much of the current task mix overlaps with what today's AI can do well. It is not a probability of job loss. This page documents how every number gets computed, what goes into it, and what the system does and does not claim.

1. Source data

The universe of 342 occupations comes from the Bureau of Labor Statistics Occupational Outlook Handbook. For each one we pull: median pay, entry education, total employment (2024 baseline), projected employment for 2034, the outlook description, and the free-text duty and task inventory.

BLS data vintage for this release: 2024-34 projections (BLS August 2025 release).

The next BLS release of the Occupational Outlook Handbook, covering the 2025-2035 projections, is scheduled for August 27, 2026. We will refresh the AOI scores and narratives against the new BLS vintage by early September 2026.

2. Task extraction

The BLS duties text is parsed by Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) into a structured list of 8-12 tasks per role. Each task is tagged with an importance label - Core, Important, or Supporting - and an evidence string quoting the BLS source line that justified extraction.

The task-extraction prompt hash is b3a4449f4b58690a, pinned for reproducibility.

3. Per-task automation scoring

Each task is then scored 0-100 by the same model for how feasible current-generation AI is at substituting for it. The scoring prompt asks for three things: a score, a rationale, and a statement of what remains human. The scoring prompt hash is 82c4dbc321aca043.

Task scores roll up into the Task Automation Impact signal using importance weights:

Core: 3x weight - defines the job
Important: 2x weight - supports the core work
Supporting: 1x weight - peripheral

4. The four signals

The AOI is a weighted sum of four signals. Each one is either per-occupation and traceable, or deterministic from a single BLS field. Every weight below can be linked back to the data that produced it.

60%

Task Automation Impact

Importance-weighted mean of per-task automation scores. Each score is produced by Claude Sonnet 4.5 with a printed rationale and is visible on the occupation page. This is the strongest per-occupation signal in the index.

10%

Automation Potential

Karpathy/BLS Digital AI Exposure score (0-10 scale rescaled to 0-100). Used directly as-is.

15%

Market Pressure

BLS 2024-2034 outlook translated deterministically into a pressure score via a published mapping table. Shrinking roles score high; growing ones low. Read this as "how tight is the job market underneath the AI question", not as an AI-caused signal.

15%

Entry Barrier Erosion

BLS typical entry-level education mapped deterministically to an erosion score. Lower credentials translate to higher erosion because AI-augmented competition genuinely floods easier-entry roles faster.

5. Seniority curves

Each category gets a trio of multipliers applied to the base AOI to express entry-, mid-, and senior-level exposure. Entry multipliers run 1.15 to 1.35 - entry roles carry the brunt because they concentrate automatable subtasks. Mid is the anchor at 1.0. Senior multipliers run 0.65 to 0.85 - judgment, accountability, and relationships insulate the top of the ladder.

6. Exposure tiers

Highly Exposed (70+): task mix is already within AI's reach today
Mostly Exposed (55-70): most of the workflow is automatable
Partially Exposed (40-55): clear pressure on routine tasks
Selectively Exposed (30-40): physical, social, or oversight-heavy
Insulated (below 30): embodied skill, frontline presence, deep institutional judgment

7. Premium narrative generation

Premium content (pivot paths, wage impact, tool recommendations) is written by Claude Opus 4.7 (). Task scoring stays on Sonnet 4.5 for cost efficiency across 3,000+ scorings; the narrative bar justifies Opus. The premium prompt hash is .

8. Audit trail and reproducibility commitment

The block below lists SHA-256 fingerprints of the exact prompts and data files that produced the scores you see on this site. These hashes are not proof on their own - without the underlying files, a reader cannot independently verify them. They serve a narrower purpose:

Internally, they let us confirm that nothing has drifted between the pipeline run and what is live.
Externally, they are a commitment. We will republish these fingerprints on every pipeline rerun, and we will publish the underlying prompt files and data artifacts themselves in a future release so anyone can check the hashes against real files.
If you spot a discrepancy or want to ask about a specific score, we want to know.

LLM-based scoring is also not bit-for-bit reproducible even with the same inputs. Model providers update their models, and responses vary between runs. A rerun of this pipeline would produce scores clustered around the current ones, not identical to them. The hashes establish what was used this time; they do not promise the same numbers next time.

generated_at2026-04-16T04:48:42

task_modelclaude-sonnet-4-5-20250929

premium_model

n_occupations342

task_extraction_promptb3a4449f4b58690a

task_scoring_prompt82c4dbc321aca043

premium_content_prompt

tasks_jsond017cc74a36773ba

task_scores_json090491aa95f4f8d5

signals_json67ee07c360441a52

9. What this is not

The AOI is a research output, not a verdict. It does not account for geography, employer, tenure, union coverage, or individual skill differentiation. It is directional: an AOI of 70 means more of the task mix overlaps with current AI capability than an AOI of 40 does. It does not mean a 70 percent chance of job loss. Nothing on this site is career advice.

10. Limitations

LLM task scoring has a known ceiling bias toward current-generation feasibility and undervalues embodied or social tasks.
BLS projections are model-based and incorporate their own assumptions about AI.
The signals are weighted from judgment, not fitted. Reasonable people can disagree on the 60/10/15/15 split.

← Back to index