AI Adoption Index: Metrics & Methodologies

• AI Adoption Indexes are measurement frameworks that quantify AI uptake using metrics like usage intensity, startup exposure, and readiness scores.
• They employ diverse methodologies including real-time usage logs, survey data, and composite benchmarking to assess economic and operational environments.
• The indices inform policy and labor market strategies by revealing regional disparities, commercialization incentives, and critical infrastructure requirements.

Searching arXiv for relevant papers on AI adoption indices and related measurement frameworks. Artificial intelligence adoption indices are measurement frameworks intended to quantify the uptake, deployment, or readiness of AI across occupations, regions, sectors, governments, firms, or sociotechnical environments. The term does not denote a single canonical construct. Instead, recent work operationalizes adoption in several distinct ways: as realized usage intensity relative to population, as startup-targeted occupational exposure, as public-sector readiness, as composite national capability-and-adoption scores, as firm-level survey-based deployment rates, or as multidimensional diagnostics of deployability and trust under infrastructure constraints. A central methodological distinction is between indices that estimate what AI could do and indices that track what AI is doing or is plausibly poised to do in practice (Fenoaltea et al., 2024).

1. Conceptual scope and definitional variants

The contemporary literature uses “AI Adoption Index” in at least three analytically distinct senses.

First, some frameworks measure observed use. The Anthropic AI Usage Index (AUI) asks whether usage in a region is higher or lower than expected given its working-age population, using Claude.ai conversations as the core signal (Appel et al., 19 Nov 2025). Malik’s BTOS-based index similarly measures the weighted share of firms that report having “used AI in the last two weeks,” with a parallel measure for planned use within six months (Malik, 19 May 2025).

Second, some frameworks measure adoption-oriented exposure rather than realized usage. The Occupational AI Startup Exposure (AISE) index links O*NET occupation descriptions to Y Combinator–funded startup applications and is explicitly designed to proxy near-future AI adoption through venture-backed innovations (Fenoaltea et al., 2024). This differs from purely technical exposure metrics because it incorporates economic viability, technical maturity, and social appetite for automation.

Third, some frameworks measure readiness or enabling conditions. The Government AI Readiness Index (GARI) evaluates whether states are prepared to harness AI in public goods and services via technological infrastructure, human resources, policy frameworks, and innovation capacity (Alalaq, 26 Mar 2025). GCC-specific work defines a public-sector AI Adoption Index as a composite 0–100 measure built from Infrastructure & Resources, Organizational Readiness, and Policy & Regulatory Environment (Albous et al., 5 Sep 2025). The Next Billion AI Index (nexbax) extends this logic further by treating economic viability, operational deployability, and governance alignment as co-equal determinants of AI utility in “next-billion” contexts (Rawat et al., 29 May 2026).

A related but separate category consists of broad composite AI indices. Some reports assemble usage, investment, talent, research, policy, and public-opinion metrics into a multidimensional adoption picture, but several major AI Index reports do not publish a single formal “AI Adoption Index” formula (Maslej et al., 2023, Zhang et al., 2022, Maslej et al., 2024). By contrast, one 2025 report explicitly describes how such indicators can be normalized and aggregated into a composite measure, though it notes that this construct is not presented as a single labeled metric in the source report itself (Maslej et al., 8 Apr 2025).

This heterogeneity suggests that “AI Adoption Index” is best treated as a family of measurement designs rather than a standardized indicator.

2. Adoption versus capability and exposure

A major fault line in the literature concerns whether an index should prioritize technical feasibility or actual diffusion. The AISE framework is explicit on this point: traditional occupational exposure measures derived from expert assessments, patent-text analyses, or benchmark-based evaluations capture “what AI could do” rather than “what AI is doing” (Fenoaltea et al., 2024). AISE instead uses venture-backed startup activity as a signal of revealed market direction.

In AISE, the occupational universe is O*NET’s 1,016 occupations, represented by concise short descriptions of essential tasks. The startup universe comprises Y Combinator firms financed between 2005 and March 2024, with filtering on AI-related thematic tags producing 958 AI-tagged startups (Fenoaltea et al., 2024). For each occupation–startup pair, a binary indicator $\delta_{o,s}$ is set to 1 if the startup’s application can substitute one or more essential tasks of occupation $o$, and 0 otherwise. The occupational score is then

$AISE_o = \tfrac{1}{|S|}\sum_{s \in S}\delta_{o,s}.$

This construction produces a ratio between 0 and 1 and directly encodes whether venture-backed AI applications target any essential task of a given occupation (Fenoaltea et al., 2024).

The classification step is operationalized using Meta’s Llama 3 8B via a yes/no prompting scheme. The robustness checks reported for prompt variants and short versus detailed startup descriptions yield Pearson correlations above 0.9, supporting stability of the binary mapping procedure (Fenoaltea et al., 2024).

The substantive contrast with theoretical exposure is central. AISE and the Ability-based AIOE index are broadly coherent at low exposure but diverge sharply at high AIOE: occupations requiring advanced degrees, notably Job Zone 5 roles, can show high theoretical exposure yet low startup targeting (Fenoaltea et al., 2024). Judges and pediatric surgeons are cited as low-AISE cases not because tasks are infeasible in the abstract, but because ethical, high-stakes, and physical considerations deter startup entry. Conversely, office clerks, data scientists, and market research analysts rank highly because structured information-processing and organizational tasks align with current startup activity (Fenoaltea et al., 2024).

A plausible implication is that adoption-oriented indices encode not only capability but also commercialization incentives and social acceptability. This makes them particularly relevant where the policy question concerns near-term labor-market pressure rather than long-run technical possibility.

3. Formal constructions and aggregation schemes

The formal structure of AI adoption indices varies considerably across settings. Several recurrent designs appear in the literature.

The usage-share normalization approach is exemplified by the AUI. For region $i$, with usage $U_i$ and working-age population $N_i$, let usage share be $s_i = U_i / \sum_j U_j$ and population share be $p_i = N_i / \sum_j N_j$. Then

$AUI_i = \frac{s_i}{p_i}.$

An $AUI > 1$ indicates higher usage than population alone would imply; $o$0 indicates under-use (Appel et al., 19 Nov 2025). This design is ratio-based, comparative, and population-adjusted rather than absolute.

The weighted prevalence approach is used in survey-based firm adoption measurement. Malik defines, for any group $o$1 and survey wave $o$2,

$o$3

where $o$4 if firm $o$5 reports having used AI in the past two weeks, and $o$6 is the survey weight correcting for stratified sampling and non-response (Malik, 19 May 2025). A parallel Intent_Index is constructed for planned use within six months.

The pillar-weighted composite approach dominates readiness and national benchmarking frameworks. GARI computes normalized pillar scores $o$7 and aggregates them with approximately equal weights:

$o$8

Indicators are min–max scaled and averaged within pillars before the top-level aggregation (Alalaq, 26 Mar 2025).

The GCC public-sector AI Adoption Index instead derives weights from PLS-SEM path coefficients. Using normalized weights from $o$9, $AISE_o = \tfrac{1}{|S|}\sum_{s \in S}\delta_{o,s}.$0, and $AISE_o = \tfrac{1}{|S|}\sum_{s \in S}\delta_{o,s}.$1, the rescaled 0–100 index is

$AISE_o = \tfrac{1}{|S|}\sum_{s \in S}\delta_{o,s}.$2

Here Infrastructure & Resources dominates the index, followed by Policy & Regulatory Environment, while Organizational Readiness receives a small weight due to its nonsignificant independent path in the estimated model (Albous et al., 5 Sep 2025).

The theme-average diagnostic approach is used by nexbax. Each of 10 dimensions receives a rubric score $AISE_o = \tfrac{1}{|S|}\sum_{s \in S}\delta_{o,s}.$3, theme scores are averaged within Effective Efficiency, Operational Practicality, and Societal Integrity, and the overall score is the equal-weight average of the three themes by default (Rawat et al., 29 May 2026).

The composite national benchmarking approach can also be hierarchical. In the United States–China regional comparison, the AI Index is

$AISE_o = \tfrac{1}{|S|}\sum_{s \in S}\delta_{o,s}.$4

with seven top-level dimensions and indicator normalization based on fixed anchor points rather than sample-dependent extrema (Li et al., 22 Oct 2025).

These examples show that adoption-index design choices are inseparable from the object being measured: usage intensity, occupational targeting, readiness, or broad ecosystem maturity.

4. Data regimes and operationalization

The empirical content of an AI adoption index depends heavily on its data-generating process.

AISE is built from O*NET short descriptions of essential tasks and Y Combinator startup descriptions harvested through web scraping, with AI startup selection based on thematic tags such as “AI,” “machine learning,” “generative AI,” and “computer vision” (Fenoaltea et al., 2024). The framework also uses O*NET skills with importance scores and Job Zone classifications for downstream analysis.

The AUI uses privacy-preserving samples of 1 million Claude.ai consumer conversations and 1 million API transcripts, geolocated by IP and filtered with suppression rules for cells with fewer than 15 conversations or fewer than 5 unique users (Appel et al., 19 Nov 2025). This makes the index behaviorally grounded in observed usage logs rather than survey reports.

Malik’s index relies on the U.S. Census Bureau’s Business Trends and Outlook Survey, fielded biweekly since September 2023, with roughly 200,000 active firms contacted every two weeks and an average response rate on AI questions of approximately 16% (Malik, 19 May 2025). Adoption is captured via direct self-report of recent usage.

GARI combines World Bank ICT Statistics, UNESCO education data, the World Economic Forum’s Global Competitiveness Report, national AI strategies, and Oxford Insights’ own surveys, with nearest-neighbor or time-series interpolation for missing values and winsorization of outliers at the 1st and 99th percentiles (Alalaq, 26 Mar 2025). The GCC index uses a purposive survey of 203 mid- and senior-level government employees across GCC states, analyzed by K-Means, PCA, and PLS-SEM (Albous et al., 5 Sep 2025).

Nexbax uses rubrics rather than a single observational dataset. Its operationalization draws on pricing disclosures, MLPerf Inference scores, AI Energy Score, benchmark suites such as HELM and DecodingTrust, multilingual benchmarks including FLORES-101 and MTEB/MMTEB, and documentation or ecosystem audits (Rawat et al., 29 May 2026). This makes it diagnostic and evaluative rather than purely descriptive.

These divergent data regimes produce different inferential properties. Usage-log and survey-based measures may better reflect realized adoption; startup-linked measures may better capture near-future commercialization pressure; readiness and rubric-based measures may better reflect enabling conditions. This suggests that cross-index comparisons are meaningful only when the underlying adoption construct is made explicit.

5. Empirical patterns across occupations, regions, sectors, and governments

Despite their heterogeneity, recent adoption indices identify several recurring empirical regularities.

AISE finds that startup targeting is concentrated in routine organizational and information-processing roles. Office Clerks (General) lead the occupational ranking with $AISE_o = \tfrac{1}{|S|}\sum_{s \in S}\delta_{o,s}.$5, followed by Data Scientists, Computer and Information Systems Managers, Market Research Analysts, and Executive Assistants (Fenoaltea et al., 2024). At the opposite end, Athletes, Clergy, Magistrate Judges, and Pediatric Surgeons register near-zero exposure. Service-oriented, information-intensive industries such as Professional, Scientific, and Technical Services, Finance, and Information show high sectoral AISE, while Agriculture, Construction, Educational Services, and Health Care show low sectoral AISE (Fenoaltea et al., 2024). Geographically, the highest regional exposure is concentrated in San Jose, San Francisco, Seattle, Austin, and Boulder.

The AUI identifies a strongly unequal geography of use. Global usage share is led by the United States at 21.6%, followed by India at 7.2%, Brazil and Japan at 3.7%, and South Korea at 3.2% (Appel et al., 19 Nov 2025). But population-adjusted adoption intensity is far more concentrated: Israel leads with $AISE_o = \tfrac{1}{|S|}\sum_{s \in S}\delta_{o,s}.$6, followed by Singapore at 4.6, Australia at 4.1, New Zealand at 4.05, South Korea at 3.73, and the United States at 3.62 (Appel et al., 19 Nov 2025). Emerging-market examples such as Indonesia at 0.36, India at 0.27, and Nigeria at 0.20 indicate substantial under-use relative to working-age population. The report further finds a global relation $AISE_o = \tfrac{1}{|S|}\sum_{s \in S}\delta_{o,s}.$7 with $AISE_o = \tfrac{1}{|S|}\sum_{s \in S}\delta_{o,s}.$8 (Appel et al., 19 Nov 2025).

At the enterprise layer, the Anthropic report shows that API deployment is disproportionately coding- and automation-heavy. Computer & Mathematical tasks account for 44% of API transcripts versus 36% on Claude.ai, and 77% of API transcripts show automation modes compared with about 50% on Claude.ai (Appel et al., 19 Nov 2025). This suggests that consumer and enterprise adoption should not be conflated.

Survey-based U.S. firm measurement yields substantially lower absolute adoption rates than conversation-based usage systems. Malik reports that only about 7% of U.S. firms currently use AI, while roughly 11% plan to do so within six months, with a linear trend explaining 86% of the variance in national use over September 2023 to April 2025 (Malik, 19 May 2025). Information Technology leads industry adoption at 15.59% current use and 20.15% intent, followed by Professional, Scientific & Technical Services at 11.01% and 16.22%; Construction and Accommodation & Food Services are at 1.39% and 1.55% current use, respectively (Malik, 19 May 2025). Larger firms adopt at roughly three times the rate of the smallest microbusinesses.

Public-sector readiness and regional benchmarking indices likewise reveal strong concentration. GARI ranks the United States first in 2023 at 84.80 and Yemen last at 19.89, with a global mean around 49.6 (Alalaq, 26 Mar 2025). The U.S.–China comparative index reports composite scores of 68.1 for the United States and 59.4 for China, with large intra-China disparities: East China scores 40.0, North China 35.9, and South China 28.4, while Central, Southwest, Northeast, and Northwest China remain much lower (Li et al., 22 Oct 2025).

A plausible synthesis is that AI adoption is highly uneven across space, sector, and institutional setting, regardless of whether adoption is measured through usage, startup exposure, or readiness.

6. Methodological controversies, robustness, and limitations

The literature repeatedly highlights problems of subjectivity, weighting, and interpretability.

Campello et al. criticize standard linear additive indices for ignoring interactions among indicators, embedding deterministic and potentially subjective weights, and failing to quantify ranking stability (Campello et al., 2024). Their proposed alternative uses the Choquet integral to model correlations among dimensions and Stochastic Multicriteria Acceptability Analysis (SMAA) to derive probabilistic rankings. In their computational analysis on Tortoise Global AI Index data, Pearson correlations above 0.75 among Research, Development, and Commercial Ventures reveal substantial redundancy, and the resulting Condorcet-derived rankings are more robust than conventional weighted sums (Campello et al., 2024).

This criticism applies broadly to composite adoption indices. GARI uses approximately equal pillar weights (Alalaq, 26 Mar 2025); the GCC public-sector index uses PLS-SEM-derived weights that heavily privilege infrastructure (Albous et al., 5 Sep 2025); nexbax defaults to equal weighting across themes but allows custom weights (Rawat et al., 29 May 2026). None of these weighting decisions is value-neutral.

Data coverage is another recurring limitation. AISE may undercount innovations outside Y Combinator or non-English contexts and abstracts away from within-occupation heterogeneity (Fenoaltea et al., 2024). The AUI is tied to Claude usage and therefore reflects one platform’s user base rather than the total AI ecosystem, even though it spans more than 150 countries (Appel et al., 19 Nov 2025). Survey-based firm measures inherit nonresponse and self-report bias (Malik, 19 May 2025). GARI relies on interpolation and indicator harmonization across heterogeneous sources (Alalaq, 26 Mar 2025).

There is also a conceptual controversy over whether broad ecosystem composites should be called “adoption indices” at all. Several AI Index annual reports explicitly track adoption through multiple indicators without publishing a single composite adoption score (Maslej et al., 2023, Zhang et al., 2022), and the 2024 edition is described as lacking any formally defined AI Adoption Index (Maslej et al., 2024). By contrast, a 2025 synthesis specifies how usage, investment, talent, research, policy, and public-opinion pillars can be normalized into a composite measure, but this construction is an interpretive synthesis of report components rather than a labeled official index in the underlying report (Maslej et al., 8 Apr 2025).

This suggests that terminological slippage is common: “AI adoption index” may refer either to a formally defined measure in a primary study or to an editorially assembled composite from multiple adoption indicators. For rigorous use, the underlying operational definition must be stated explicitly.

7. Policy and research significance

AI adoption indices are primarily valuable because they transform diffuse, multidimensional evidence into structured signals for labor, industrial, and governance analysis.

AISE is framed as a near-real-time gauge of where venture capital is directing automation efforts, thereby supporting targeted reskilling, adjustments to occupational licensing or certification, and infrastructure investment in lagging regions (Fenoaltea et al., 2024). Its main policy contribution is to challenge blanket claims that “high-skill” jobs uniformly face imminent AI disruption.

The AUI offers a different lens: it reveals concentrated early adoption in high-income, innovation-ready regions and identifies task- and context-bottlenecks affecting enterprise deployment (Appel et al., 19 Nov 2025). Its open-sourced task-level, geographic, and collaboration-mode data are intended to support independent research on macroeconomic and labor effects. The finding that high-AUI regions favor collaborative augmentation while low-AUI regions favor automation after controlling for task mix further suggests that local economic structure shapes not only how much AI is used but how it is used (Appel et al., 19 Nov 2025).

Public-sector and national indices are used for benchmarking and strategic planning. GARI is designed to assess whether governments can harness AI in public service delivery (Alalaq, 26 Mar 2025). The GCC index identifies infrastructure and policy clarity as the strongest determinants of AI outcomes in a resource-rich, top-down environment, with the combined model explaining 70% of the variance in AI outcomes (Albous et al., 5 Sep 2025). Nexbax is proposed not as a universal score of social value but as a modular diagnostic for “artificial useful intelligence,” especially where intermittent connectivity, low-end devices, tight budgets, and multilingual settings constrain deployment (Rawat et al., 29 May 2026).

Across these strands, a common theme emerges: adoption is not reducible to model capability. Economic viability, deployability, governance, local infrastructure, trust, and institutional arrangements all condition diffusion. This suggests that future adoption-index research will likely move toward hybrid designs that integrate observed usage, commercialization signals, readiness conditions, and robustness analysis rather than relying on any single proxy.