What we ask, and what you get back

E = mean(evidence dimensions)
A = mean(attitude dimensions)

composite = round( min( 0.6·E + 0.4·A ,  E + 18 ) )

The right-hand term is a hard bound: the composite cannot exceed the evidence mean by more than 18 points at any attitude score. A response with maximum attitude and zero evidence yields a composite of 18.

people = round( min( 0.6·fluency + 0.4·mean(mindset, vision) ,  fluency + 18 ) )

A second reading of the same response, not a second score. No composite or tier changes, and the People number carries no tier. It is shown beside the composite, and a difference of 12 points or more is stated in either direction. It does not use the outside-work item.

Eloundou, Manning, Mishkin & Rock — “GPTs are GPTs” (Science, 2024)

Occupation-level exposure scores. MIT licensed.

O*NET 30.2 Database, USDOL/ETA

Job-title → occupation mapping, and the occupational task statements behind the personal game plan. Used under the CC BY 4.0 license; we filtered it to scored occupations, normalized the titles, kept only Core tasks, and truncated each occupation's task list to the eight with the most incumbents responding.

Bick, Blandin & Deming, “The Rapid Adoption of Generative AI”, Federal Reserve Bank of St. Louis Working Paper 2024-027 (revised October 2025)

“We find an average time savings of 5.2% of work hours by genAI users in the November 2024 wave, or about 2.1 hours for a 40-weekly-hour worker.”

US Real-Time Population Survey, November 2024: employed adults who used generative AI for work in the previous week (n = 925), self-reported. “Less than an hour” counts as 0 and “more than four” as 4, so the authors call it a lower bound.

Bick, Blandin & Deming (as above), Figure 8b: time saved by how often people used it

“Among workers who used genAI every day last week, 36.8% reported saving four hours or more, compared to 10.4% of respondents who used it one day last week.”

Same survey. Weighting Figure 8b’s shares for daily users (19.6% two hours, 17.5% three, 36.8% four or more) gives about 2.4 hours a week, again a lower bound.

Humlum & Vestergaard, “Large Language Models, Small Labor Market Effects”, NBER Working Paper 33777 (2025)

“Across our 11 occupations and employer policies, the average time savings amount to 2.8% of the total work hours of users.”

Danish survey of workers in 11 AI-exposed occupations, linked to employer records. Users of any frequency, including monthly, self-reported.

OpenAI, “The State of Enterprise AI”, 2025 Report

“Workers who save more than 10 hours per week are not just using more intelligence, they are also using multiple models, engaging with more tools, and using AI across a wider range of tasks.”

OpenAI’s own survey of 9,000 workers across almost 100 enterprises, matched to usage data, self-reported. Vendor research about the vendor’s product; the top survey answer was “>10” hours a week.

8.1Exposure is measured, but not by this instrument

Readiness (how able a respondent is to work with AI) and exposure (how much AI will change their role) are independent. This instrument scores only readiness. Exposure is now reported alongside it as a separate measure, derived from published occupation-level research rather than from any response given here (§6). Consequence: role exposure still loads onto the vision dimension, because a respondent in a low-exposure role who accurately reports that AI will not reshape their work presents as low vision. The rubric's protections for skeptical and accurate-low responses remain load-bearing, and the two scores are reported side by side rather than combined.

8.2Exposure resolves to an occupation, not to a respondent

An exposure score is the published figure for an entire occupation, assigned by matching the respondent's typed role. Every respondent matched to an occupation receives its score regardless of their actual task mix. The underlying ratings were produced against a 2023-vintage model and O*NET's task lists, and the occupation taxonomy has no entry for some contemporary titles — a software product manager resolves to Marketing Managers, the nearest available. Consequence: exposure is interpretable at band resolution for a class of work, not as a fact about an individual, and it ages independently of this instrument's versioning.

8.3Roles that name no occupation carry no exposure

Titles that do not identify the work — "manager", "team lead", "associate" — are left unmatched rather than assigned to a generic occupation that would nominally accept them. Affected respondents are reported as exposure unknown, excluded from role averages, and counted explicitly. Consequence: role coverage is partial by design, and the uncovered share is stated wherever exposure is shown.

8.4Scores are not psychometric measurements

Dimension scores blend a language model applying a fixed rubric to four free-text answers (temperature 0.2) with deterministic scores on the structured items. Repeat scorings of identical prose can differ by a few points. No inter-rater reliability coefficient, test-retest coefficient, or factor analysis has been computed. Consequence: scores are interpretable at band and tier resolution, not at point resolution. A 63 and a 66 are the same reading.

8.5No normative sample

Tier boundaries are fixed against the rubric, not derived from a reference population. "Builder" denotes meeting a defined bar, not a percentile within an industry. Consequence: cross-company comparison is not supported, and no benchmark claim is made.

8.6Single-rater

Each submission is scored once, by one engine, from one sitting. There is no second rater. Consequence: individual scores carry more uncertainty than aggregate ones. Organization-level readings are more stable than any single person's.

8.7A retake sees the same knowledge items

Respondents may take the instrument again, and a second reading is compared to the first only when the two are at least 30 days apart and were produced by the same version. The five objective knowledge items are identical on every administration: the instrument has no parallel forms, because alternate items would have to be shown to be equally difficult before a score built on them could be compared to an earlier one. Consequence: a repeat respondent has seen those items before, which biases the fluency dimension slightly upward on retake. The effect is bounded — fluency blends the objective items evenly with a free-text answer that is not repeated — but it is a real, one-directional artifact of retaking, not a change in the respondent.

8.8Self-report moves when a standard moves

Several items ask a respondent to characterize their own practice and confidence. Someone who has spent weeks learning how much they did not know about checking AI output can answer those items more conservatively than they did before, having revised what they consider adequate. Consequence: a lower second reading is not evidence that a respondent got worse, and a decline over time is not by itself a finding about a program. Movement is reported as counts of respondents who rose, held, or fell rather than as a single figure, and a change within the instrument's resolution is reported as held.