How we score AI exposure

Every role gets broken into 9–15 concrete tasks across three buckets: tasks AI handles end-to-end, tasks where you and AI collaborate, and tasks that stay human. This task decomposition is inspired by task-based and occupation-based exposure studies — notably Eloundou et al. (2024), "GPTs are GPTs" in Science, and the Felten, Raj & Seamans (2021) AIOE framework — but the specific task list and rubric below are ours, not theirs.

• Routinization (R)

  How repeatable and rule-based is the task? Filing the same expense report each week scores high; negotiating a contract scores low.

• Data availability (D)

  Does AI have the inputs it needs — text, code, images, structured data? A task that hinges on tacit context the model can't see scores low.

• Judgment required (J)

  How much human judgment, ethics, or accountability does it take? This is an inverse signal — high judgment lowers automation risk.

• Presence required (P)

  Does it need a body in a room or a real human relationship? Also inverse — physical / interpersonal presence resists automation.

automation(task) = (R + D) / 8 × (1 − J/4) × (1 − P/4)
augmentation(task) = D/4 × max(J/4, P/4) × (1 − automation)

Both scores are bounded 0–1. A task scores high on automation only when it's repeatable and AI has the data and human judgment / presence are low. Augmentation is highest when AI has the data but a human still has to own the call.

These are design choices, not literature outputs. By construction, automation falls to exactly zero whenever J = 4 or P = 4 — meaning maximum required judgment or in-person presence zeroes out automation regardless of how routine the task is. That's a strong normative bet on what makes work resist AI; reasonable people could weight these factors differently.

automation  = (4 + 4) / 8 × (1 − 1/4) × (1 − 0/4)
            = 1.00 × 0.75 × 1.00
            = 0.75   →  75%

augmentation = 4/4 × max(1/4, 0/4) × (1 − 0.75)
            = 1.00 × 0.25 × 0.25
            = 0.06   →  6%

automation  = (1 + 1) / 8 × (1 − 4/4) × (1 − 4/4)
            = 0.25 × 0.00 × 0.00
            = 0.00   →  0%

augmentation = 1/4 × max(4/4, 4/4) × (1 − 0.00)
            = 0.25 × 1.00 × 1.00
            = 0.25   →  25%

The model also estimates each task's share of the role's time (a kind of "where does the day actually go"). Weights are normalized to sum to 100%, and the role-level scores are time-weighted sums:

automation_risk = Σ time_share_i × automation_score_i
augmentation_potential = Σ time_share_i × augmentation_score_i

We compute a time-weighted standard deviation of per-task automation scores and present it as a ± range (clamped to ±3 to ±18 percentage points). A role with consistent task ratings gets a tight band; a role that mixes very automatable and very human work gets a wider one.

This is not a statistical confidence interval in the repeated-sampling sense — it doesn't quantify uncertainty around a population parameter. It's a dispersion measure that tells you how internally mixed the task profile is. Read it as: "how much do the tasks inside this role disagree with each other?"

Each map is mapped to its closest O*NET-SOC code — the U.S. Department of Labor's standard occupation taxonomy — so your role connects to a real labor-market reference, not a free-form string.

Note that our time-share weighting is a product choice that goes beyond O*NET's native core-vs-supplemental task convention (which is based on relevance and importance thresholds, not estimated time). It's defensible as a modeling decision, but it is additional structure on top of O*NET, not a feature of O*NET itself.

Methodological foundations:

• Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702). Preprint: arXiv:2303.10130 (we rely on the preprint appendix for rubric-level detail).
• Brynjolfsson, E., Mitchell, T., & Rock, D. (2018). What can machines learn, and what does it imply for the workforce? AEA Papers and Proceedings — the Suitability for Machine Learning (SML) framework, the closer methodological ancestor of the GPTs task-exposure rubric.
• Felten, E., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence (AIOE). Strategic Management Journal. The authors explicitly state AIOE is agnostic about whether AI substitutes for or complements labor — citing it as direct support for an "automation risk" score would overclaim.
• U.S. Department of Labor / ETA, O*NET OnLine (onetonline.org) — task, ability, and SOC taxonomy.

Career-guidance reference (not a methodological foundation):

• U.S. Bureau of Labor Statistics, Occupational Outlook Handbook — used for role descriptions and outlook context, not for the exposure scores themselves.

• Per-task ratings (R, D, J, P) and the task list itself are generated by a language model against our rubric. So the end-to-end pipeline is "model-generated task judgments + deterministic aggregation," not deterministic from first principles.
• The headline number is best read as technical automation exposure — how much of the work AI could do — not a labor-substitution forecast about hiring, wages, or whether a specific employer will adopt AI.
• The "2–3 year horizon" is our near-term forecast window, not a claim made by any of the cited papers. The GPTs paper in particular explicitly does not predict adoption timelines.
• The O*NET code is the model's best mapping — not yet validated through O*NET Web Services or the O*NET Code Connector.
• The ± band is task dispersion, not a statistical confidence interval (see Step 5).