Methodology

Your AI Exposure Score is transparent and explainable by design - here's exactly how it's built.

What the score means

The AI Exposure Score is a 0–100 estimate of how much your role's tasks overlap with what AI can do today and in the near future. A higher number means more of your current tasks are exposed - it is not a prediction that your job will disappear, and every score is paired with a constructive next step.

The bands

Recalibrated July 2026: the earlier thresholds assumed scores would spread across close to the full 0–100 range. In practice, real scores across all 968 occupations cluster more tightly (roughly 23–84), so we adjusted the boundaries so each of the five tiers reflects a meaningful share of real occupations. No occupation's underlying 0–100 score changed - only which band label it falls into.

How the percentile works

“More exposed than N% of roles” means exactly that: the percentage of tracked occupations with a strictly lower exposure score. Roles with the same score always receive the same percentile. We round down rather than up, so the displayed percentage never overstates the share of roles below it.

How the composite is weighted

The score uses six occupation-specific components plus a fixed neutral-context baseline. The weights are published so you can see what drives a number:

The 5% context term is currently fixed at a neutral value of 50 for every occupation. It keeps the published formula stable but does not personalize the precomputed index score; scanner answers do not change that number. A paid report can personalize guidance, reframes, and the action plan around your situation; it does not personalize or recalculate the exposure score.

Task-level exposure

Beyond the single number, your role is split into its real tasks, each tagged:

• Automatable - likely to be largely done by AI.
• Augmentable - AI speeds it up; you stay in the loop.
• Durable - resists automation (judgment, relationships, accountability).

This is the actionable part: shift your time toward the durable and augmentable core.

How each task gets its own score: rather than guessing from the words in a task description, we look at how automatable that same real-world activity tends to be across every occupation that performs it (via O*NET's own activity taxonomy) - not a guess specific to your role. One honest limitation this creates: a task is scored by what kind of work it is, not who's doing it - so a relationship- or judgment-heavy task can still read as more exposed than expected if the same underlying activity is shared with more automatable work elsewhere. We disclose that rather than force artificial variety into a role's breakdown that the data doesn't actually support.

How specialized roles are mapped

Some job titles people use every day are specializations inside a broader O*NET/BLS occupation, not standalone government occupations. For those roles, we keep the authored role-specific tasks first and add only a reviewed subset of the parent O*NET occupation's tasks when they fit the specialization.

When an everyday title can point to multiple O*NET occupations, we do not let the first alternate-title match decide the task list. Ambiguous titles need a reviewed mapping, or they stay on their authored task set until one exists.

Tasks are included or excluded on occupational relevance only, never on their AI-exposure score. Each reviewed bridge records the source occupation, included task IDs, excluded task IDs, and exclusion reasons in version-controlled data so the mapping can be audited.

Example: Data Analyst. O*NET lists Data Analyst as a reported job title under Data Scientists (SOC 15-2051.00), but that title appears under several O*NET occupations. We preserve the four authored generalist analyst tasks and add seven reviewed tasks covering visualization, presentation, problem framing, trend analysis, business decisions, stakeholder recommendations, and analysis programming. Predictive-model, survey, scientific-theory, and research-monitoring tasks are excluded as non-universal analyst work.

Where the data comes from

O*NET-linked occupations start from O*NET 29.1 descriptors: work activities, work context, skills, job zone, related occupations, public-facing and interpersonal work, and physical-presence requirements. Those O*NET fields also produce the occupational-resilience component; BLS does not. The automation and augmentation components are then empirically grounded where matched by three public research datasets. Two independent measures set the exposure magnitude: the OpenAI/Penn "GPTs are GPTs" task-exposure study (~88% of roles) and the Felten, Raj & Seamans AI Occupational Exposure (Language-Modeling variant; ~78% of roles) - one task-based, one ability-based, so they corroborate rather than echo each other. The Anthropic Economic Index of observed Claude usage then nudges whether real-world use of a role leans toward automation or augmentation (~69% of roles). Roles a dataset does not cover keep the transparent O*NET-derived estimate. The 116 curated roles without a matched O*NET SOC retain a labeled, lower-confidence modeled estimate; we never claim empirical coverage they do not have. BLS wage, employment, growth, and openings appear separately as U.S. labor-market context and are not score inputs. These are exposure measures (what AI could affect), not predictions of job loss; the Anthropic signal reflects Claude.ai usage specifically, so it is a proxy, not a census. See the full data sources & attribution.

Exposure is not destiny

An exposure score measures what AI could affect about a role - a theoretical ceiling - not a prediction that the job disappears. The early real-world evidence fits that distinction. Stanford's "Canaries in the Coal Mine" study finds a roughly 16% relative employment decline for early-career workers ages 22-25 in the most AI-exposed U.S. occupations, after firm-level controls, relative to workers in less-exposed fields and more experienced workers in the same occupations. This is a group-level association, not proof that AI alone caused the change. Broader analyses (the Yale Budget Lab and BLS) find overall employment has so far held up; PwC's 2026 AI Jobs Barometer even finds the industries most exposed to AI have seen faster wage growth and productivity, and that workers with AI skills now command a roughly 62% wage premium - exposure is reshaping work, not simply erasing it. The strain is showing up first in entry-level cognitive work, not across the whole economy.

Revision history: the study originally reported about 13% with data through July 2025, then reported 16% after adding firm-level controls and data through October 2025 in the November 2025 revision. The authors' February 2026 follow-up says the relationship becomes significant from 2024 under the broadest controls and cannot establish that AI alone caused it.

The constructive read: the roles that hold up pair durable human judgment, relationships, and domain depth with AI-augmented work - so the move is to build those earlier, whatever your exposure band. And because the published exposure measures disagree in magnitude (which is exactly why we blend several rather than trust one), we never present a score as a certainty. See the corroborating research behind this.