AGILE Index: AI Governance Evaluation

• AGILE Index is a composite index that measures national AI governance by balancing technological development, governance environment, policy instruments, and effectiveness.
• It employs a structured framework evolving from 2024 to 2025, expanding coverage and refining dimensions to ensure scientific rigor and cross-country comparability.
• The index integrates multi-source data with normalization and aggregation techniques to diagnose alignment between development and governance outcomes.

The AI Governance InternationaL Evaluation Index (AGILE Index) is a cross-country composite index designed to measure national AI governance under the principle that “the level of governance should match the level of development.” The inaugural evaluation, released in February 2024, assessed 14 representative countries through four pillars, 18 dimensions, and 39 indicators; AGILE Index 2025 expanded the framework to 40 countries, 17 dimensions, and 43 indicators, integrating policy documents, governance practices, research outputs, and risk exposure into a unified comparison architecture (Zeng et al., 21 Feb 2025, Zeng et al., 10 Jul 2025). Its distinctive feature is that it does not treat AI governance as a narrow legal or administrative question: it links technological development, governance environment, governance instruments, and governance effectiveness in a single evaluative structure (Zeng et al., 21 Feb 2025).

1. Emergence and governing rationale

AGILE emerged in response to the rapid expansion of generative AI and the associated acceleration of ethical, legal, social, and policy concerns. The 2024 formulation explicitly framed effective AI governance as a global concern and positioned the index as a data-driven instrument for depicting national governance status, identifying governance stages, and uncovering governance issues across countries (Zeng et al., 21 Feb 2025). The 2025 release described the project as having been launched since 2023, with the 2024 edition serving as an operational and comparable baseline that was subsequently refined to improve scientific rigor, practical adaptability, metric validity, and cross-national comparability (Zeng et al., 10 Jul 2025).

The guiding principle—“the level of governance should match the level of development”—structures the entire index. In AGILE’s own logic, under-governance permits uncontrolled development and social harm, whereas over-governance may impede innovation. The index therefore operationalizes governance not as a standalone capacity, but as a balancing relation among development, institutional readiness, policy instruments, and realized social outcomes (Zeng et al., 21 Feb 2025). A plausible implication is that AGILE should be read less as a pure league table than as an alignment diagnostic: the most important analytical object is often the profile across pillars rather than the headline rank alone.

AGILE also defines itself against adjacent indices. The 2024 study states that Stanford HAI’s AIVT emphasizes development and vibrancy, Tortoise GAII centers on capabilities, Oxford Insights’ GRI on government readiness, and OECD’s GDT on digital transformation, whereas AGILE integrates development, governance environment, instruments, and effectiveness, and adds AI governance research and AI for SDGs (Zeng et al., 21 Feb 2025). The 2025 release preserved that broad governance orientation while expanding source diversity by more than 20 sources and refining keyword sets to include frontier AI such as generative AI (Zeng et al., 10 Jul 2025).

2. Framework architecture

Across both major releases, AGILE is organized around four top-level pillars: AI Development Level, AI Governance Environment, AI Governance Instruments, and AI Governance Effectiveness. What changed between 2024 and 2025 was the internal dimensional structure, the number of countries covered, and the indicator set (Zeng et al., 21 Feb 2025, Zeng et al., 10 Jul 2025).

Release	Coverage	Structure
2024	14 countries	4 pillars, 18 dimensions, 39 indicators
2025	40 countries	4 pillars, 17 dimensions, 43 indicators

In the 2025 architecture, Pillar 1, AI Development Level, contains D1 AI R&D Activity, D2 AI Infrastructure, and D3 AI Industry Vitality. Pillar 2, AI Governance Environment, contains D4 AI Risk Exposure and D5 Overall Governance Readiness. Pillar 3, AI Governance Instruments, contains D6 AI Strategy & Planning, D7 AI Governance Bodies, D8 AI Principles & Norms, D9 AI Impact Assessment, D10 AI Standards & Certification, D11 AI Legislation Status, and D12 Global AI Governance Engagement. Pillar 4, AI Governance Effectiveness, contains D13 Public Understanding of AI, D14 AI Social Acceptance, D15 AI Development Inclusivity, D16 Data & Algorithm Openness, and D17 AI Governance Research Activity (Zeng et al., 10 Jul 2025).

The 2024 version used a closely related but slightly different dimensional decomposition. Its development pillar comprised AI Research and Development Activity, AI Infrastructure, and AI Industry Scale; the governance environment pillar comprised AI Risk Exposure and AI Governance Readiness; the governance instruments pillar included AI Strategy & Planning, AI Governance Bodies, AI Principles & Norms, AI Impact Assessment, AI Standards & Certification, AI Legislation Status, and Global AI Governance Engagement; and the effectiveness pillar included Public Understanding of AI, Public Trust in AI, AI Development Inclusivity, Data & Algorithm Openness, AI Governance Research Activity, and AI for SDGs Activity (Zeng et al., 21 Feb 2025). The 2025 restructuring therefore did not replace the original conceptual logic so much as consolidate and refine it.

At the indicator level, the 2025 edition combined quantitative continuous variables, binary and ternary legal-policy codings, and survey-derived scales. Examples include D1.4 “Number of large-scale AI systems developed & ratio to GDP,” D4.1 “AI-related incidents & ratio to GDP,” D6.1 “AI strategy published (0/100),” D11.1 “Comprehensive AI law enacted/in-process (0/50/100),” D14.3 “Trust in AI applications,” and D17.2 “AI safety/security publications and proportion” (Zeng et al., 10 Jul 2025). This mixture of indicator types is central to AGILE’s design: the index attempts to combine structural capacity, legal institutionalization, and social reception within a single scoring system.

3. Indicators, sources, and coding practice

AGILE’s data model is explicitly multi-source. In the 2025 release, development indicators draw on DBLP, WIPO PATENTSCOPE, Epoch/Our World in Data for large-scale AI systems, Stanford HAI AI Index, QUID, TOP500, Data Center Map, and ITU IDI. Governance context indicators use World Bank WGI, HDI, ITU Global Cybersecurity Index, Global Data Barometer, World Bank GovTech Maturity, UN EGDI/E-Participation, and the SDG Transformation Center. Instrument indicators rely on desk research of official policy and legislation, as well as UNESCO, G20, UN, the AI Safety Summit (UK), the Seoul AI Ministerial, the AI Action Summit (France), REAIM, and ISO/IEC JTC 1/SC 42. Risk exposure is built from OECD AIM, AIID, AIAAIC, and AIGO. Effectiveness indicators use IPSOS AI Monitor, OECD Going Digital Toolkit, and IBM Global AI Adoption Index (Zeng et al., 10 Jul 2025).

The 2024 edition used a similarly heterogeneous evidence base, but with a stronger reliance on Tortoise Media GAII, DBLP, TOP500, Data Center Map, LifeArchitect.AI, Epoch.AI, AIID, AIAAIC, AIGO Observatory, OECD AIM, World Bank WGI, GTMI, SDGDI, OECD PISA, Coursera, IPSOS OEAI, KPMG TAI, IBM AI Adoption Index, Hugging Face, GitHub, Stack Overflow, Springer, IEEE Xplore, ACM Digital Library, and an AI4SDGs casebase, alongside survey work with local expert assistance for strategies, governance bodies, principles, impact assessments, standards, and legislation (Zeng et al., 21 Feb 2025). The continuity across versions is the use of triangulated evidence rather than a single repository.

The legal-policy coding rubrics are unusually explicit. In 2025, the D6 Strategy rubric assigns D6.1 a score of 100 if a strategy exists and 0 otherwise; D6.2 a score of 100 only when implementation is specific, actionable, and quantifiable; D6.3 a score of 100 when training or skills upgrading is mentioned; and D6.4 a score of 100 when an ethical component is explicit. Multiple strategies do not yield extra points; the latest or most representative strategy is used (Zeng et al., 10 Jul 2025). D11 legislation is coded on a ternary or binary basis: D11.1 assigns 100 to a comprehensive national AI law, 50 to an in-process draft or proposal, and 0 otherwise; D11.2 assigns 100 when national vertical AI laws or regulations are published; and D11.3 assigns 100 when national AI-targeted data or information protection laws, or AI-targeted amendments, exist. Executive orders and highest court precedents with equivalent national force are included, while regional or provincial laws are excluded unless region-wide frameworks apply nationally (Zeng et al., 10 Jul 2025).

The sampling and validation procedure is correspondingly documentary. National-level laws, acts, executive orders, decrees, regulations, and amendments are cataloged; institutional establishment is confirmed through official announcements and mandates; standards participation is verified against SC 42 P-member lists; and risk incidents are aggregated across multiple repositories with time-window alignment to March 2025 (Zeng et al., 10 Jul 2025). The index therefore relies less on stated commitments than on publicly verifiable artifacts and continuity checks.

4. Quantification, normalization, and reliability

AGILE uses a composite-indicator methodology with arithmetic aggregation and equal weighting. In the 2025 release, dual-base indicators are constructed by combining an absolute measure with either a per-capita or GDP-normalized measure. If $x$ is the raw quantity, the scale-adjusted forms are

$$x_{\mathrm{pc}} = \frac{x}{\mathrm{population}}, \qquad x_{\mathrm{gdp}} = \frac{x}{\mathrm{GDP}}.$$

After normalization, the dual-base indicator score is

$$s_i = \frac{s_{\mathrm{total}} + s_{\mathrm{scale}}}{2}.$$

The baseline normalization pipeline is z-score standardization followed by linear rescaling and clipping:

$$z = \frac{x - \mu}{\sigma}, \qquad s = 25z + 50, \qquad s' = \min(100,\max(0,s)).$$

For highly skewed data, AGILE 2025 uses percentile-fit normalization, which iteratively trims extreme 1-percent tails after initial standardization and re-standardizes the remaining distribution until quartiles stabilize (Zeng et al., 10 Jul 2025).

Aggregation proceeds in three stages. Indicator scores are averaged into dimensions,

$d_j = \frac{1}{m_j}\sum_{i=1}^{m_j} s_{ij},$

dimensions are averaged into pillars, and pillars are averaged into the overall index:

$I = \frac{1}{4}\sum_{k=1}^{4} p_k.$

A special inversion applies to D4 AI Risk Exposure, since higher incident exposure is treated as worse governance context:

$d_4' = 100 - d_4.$

Pillar 2 is then computed using the inverted risk score together with the governance-readiness dimensions (Zeng et al., 10 Jul 2025). The 2024 edition used the same equal-weight principle, similar arithmetic aggregation, and a related quantile-adapted normalization, described there as percentile-fit or quantile-adapted normalization (Zeng et al., 21 Feb 2025).

Missing data handling changed across releases. In 2024, AGILE used within-indicator mean imputation from correlated sources, and for missing indicator scores it used rank-adjusted mean imputation: temporary imputation at 50, provisional dimension scoring, estimated ranking, and replacement with the mean of countries at similar estimated rank (Zeng et al., 21 Feb 2025). In 2025, this was replaced by a more formal hierarchy. If a 2024 value exists but 2025 is missing, historical trend imputation is

$\hat{x}_t = x_{t-1}(1+\bar r).$

If missingness remains, hierarchical imputation flows from available sub-items to indicator score, from indicator score to dimension score, and from dimension score to pillar score. The 2025 methodology also allows regression-based imputation,

$\hat{x} = \beta_0 + \sum_j \beta_j s_j,$

with predictors chosen from correlated indicators and uncertainty tracked by residual variance (Zeng et al., 10 Jul 2025).

Reliability and robustness are acknowledged but only partially implemented. The 2025 report states that qualitative coding for D6–D12 used dual-coder review, rule-based rubrics, and adjudication of disagreements, but formal inter-rater statistics were not reported in pre-release v1.0.0-pre; Cohen’s $\kappa$ is identified as the intended measure. It also recommends sensitivity checks for weights and normalization choices, and bootstrap confidence intervals, but notes that uncertainty bands were not yet included in the pre-release (Zeng et al., 10 Jul 2025). This is significant: AGILE emphasizes interpretability and transparency, but its own documentation treats formal reliability reporting as an ongoing rather than completed component.

5. Empirical findings and comparative patterns

The 2024 inaugural AGILE ranking placed the United States first at 72.3, followed by China at 68.5, Singapore at 66.4, Canada at 64.9, Germany at 62.6, and the United Kingdom at 61.4; South Africa ranked last among the 14 countries at 34.4 (Zeng et al., 21 Feb 2025). The report divided countries into four score tiers—above 60, 50–60, 40–50, and below 40—and also into three governance types by pillar profile. Type A combined higher development and investment with lower governance environment scores, and included China, the United States, and the United Kingdom. Type B exhibited relatively even pillar scores and included Singapore, Canada, Germany, Japan, and France. Type C combined higher governance environment scores with lower overall AGILE scores, and included India, Brazil, Italy, and the UAE (Zeng et al., 21 Feb 2025).

By 2025, the top 10 were China (70.1), the United States (69.9), Germany (68.8), South Korea (67.8), the United Kingdom (67.6), Singapore (65.7), France (65.6), Canada (63.5), Japan (61.4), and Finland (60.3). The bottom 10 were South Africa (33.9), Colombia (37.1), Argentina (37.2), Peru (40.8), Brazil (40.8), India (41.0), Hungary (41.3), New Zealand (41.8), Poland (43.1), and Mexico (43.8) (Zeng et al., 10 Jul 2025). The 2025 release reports three performance tiers: Tier 1 above 60, Tier 2 between 50 and 60, and Tier 3 below 50. It also identifies four governance types: All-round Leaders (US, Singapore, China, UK), Governance Overachievers (France, South Korea, Canada), Governance Shortfallers (Ireland, Israel, New Zealand), and Foundation Seekers (India, South Africa) (Zeng et al., 10 Jul 2025).

Several structural patterns recur across both editions. First, overall AGILE scores are positively correlated with GDP per capita, though with notable exceptions (Zeng et al., 21 Feb 2025, Zeng et al., 10 Jul 2025). Second, development and governance instruments are the most dispersed pillars, whereas governance environment and effectiveness are more concentrated (Zeng et al., 10 Jul 2025). Third, countries with stronger development often face higher governance pressure because risk exposure rises alongside scale. The 2024 report makes this especially explicit for the United States, which leads in publications, researchers, patents, systems, supercomputing power, data centers, funding, and startups, but also accounts for 67% of incidents among the assessed countries in 2023 (Zeng et al., 21 Feb 2025).

The 2025 edition also emphasizes that lower- and middle-income countries can partially compensate for weaker development resources through lower risk exposure and higher social acceptance. It reports that upper- and lower-middle-income countries show a slight edge in Pillar 2 and Pillar 4, driven by lower risk exposure and higher social acceptance (Zeng et al., 10 Jul 2025). Country examples illustrate this logic. France ranks first in instruments with $$x_{\mathrm{pc}} = \frac{x}{\mathrm{population}}, \qquad x_{\mathrm{gdp}} = \frac{x}{\mathrm{GDP}}.$$0 but only mid-tier in effectiveness; Saudi Arabia performs strongly on environment and instruments but weakly on public understanding and acceptance; Indonesia, Mexico, and Thailand show high social acceptance and lower risk exposure despite lagging development inputs; Denmark and Finland rank highly in governance readiness despite somewhat lower instrument depth (Zeng et al., 10 Jul 2025).

AGILE’s own guidance warns against reading the composite score in isolation. The 2025 report explicitly advises against over-weighting headline rank, instructs readers to examine pillar profiles and dimension drivers, and notes that year-on-year comparisons should be used only where definitions are consistent because new indicators affect comparability (Zeng et al., 10 Jul 2025). A related methodological caveat already appeared in 2024: the index operationalizes the principle that governance should match development through pillar structure and profile analysis, but it does not define a single numeric alignment metric (Zeng et al., 21 Feb 2025).

6. Extensions, reinterpretations, and methodological debates

Subsequent literature has used AGILE not only as a country index but also as a transferable design template. One line of work proposes adapting the transparent composite-indicator framework of the Global AI Vibrancy Tool—its scope, pillars, indicators, weighting, normalization, and aggregation methods—to governance-specific aims, explicitly presenting AGILE as a governance-focused inheritance of that methodology (Fattorini et al., 2024). Another line, centered on general-purpose AI evaluations, argues that an AGILE-style index could score governance regimes on internal validity, external validity, reproducibility, portability, institutional capacity, and provider commitments, thereby shifting the unit of analysis from states alone to evaluation systems under the EU AI Act (Paskov et al., 2024). These proposals suggest a methodological broadening of AGILE from national governance profiling toward evaluation science and regulatory assurance.

A second extension concerns the normative scope of governance. The Sentience Readiness Index treats AGILE as a useful host framework but argues that mainstream AI governance indices underweight moral-status preparedness; it therefore proposes categories such as Policy Environment, Institutional Engagement, Professional Readiness, Public Discourse, and Adaptive Capacity, and recommends partial non-compensability so that strong R&D does not mask ethical or professional deficits (Rost, 2 Mar 2026). The AI Pluralism Index makes a different intervention: it formalizes evidence-based governance scoring through participatory governance, inclusivity and diversity, transparency, and accountability, and introduces explicit handling of “Unknown” evidence by reporting evidence lower-bounds, known-only scores, and coverage (Mushkani, 9 Oct 2025). Both strands suggest that AGILE’s original composite structure can be retained while its normative target changes.

A third cluster of work moves AGILE toward international institutional assessment. Proposals on international institutions for advanced AI map governance functions into models such as a Commission on Frontier AI, an Advanced AI Governance Organization, a Frontier AI Collaborative, and an AI Safety Project, then outline AGILE-like dimensions around readiness, capability, inclusivity, and performance (Ho et al., 2023). Jurisdictional certification proposals centered on an International AI Organization extend this logic further, tying AGILE-style scoring to certification, trade-control alignment, licensing, liability, and enforcement against non-certified jurisdictions (Trager et al., 2023). Historical case-study work on nuclear, chemical, biological, and export-control agreements likewise recasts AGILE around verification robustness, enforcement strength, adaptability to technological change, transparency-security balance, power balance, participation incentives, scope, and export-control coordination (Wasil et al., 2024). These are not components of AGILE 2025 itself; they are subsequent attempts to generalize its measurement logic to international governance regimes.

A fourth debate concerns equity, sovereignty, and representation. Global-majority-oriented work argues that AGILE-like evaluation should prioritize access and infrastructure, capacity and education, inclusion and participation, equity and impact outcomes, corporate power balance and sovereignty, and collaboration through regional strategies, with an explicit equity adjustment factor that discounts high scores when benefits are unequally distributed (Okolo et al., 23 Jan 2026). Relatedly, the AI Index 2026 synthesis proposes enriching AGILE with data on AI Safety Institutes, GIRAI privacy and bias dimensions, FMTI transparency scores, AI incidents from AIID and OECD AIM, data-center concentration, and an “AI sovereignty” lens spanning infrastructure, R&D, legal-regulatory posture, cultural-linguistic inclusion, and international engagement (Sajadieh et al., 14 Apr 2026). These additions reflect a broadening consensus that AI governance measurement must increasingly account for evaluation capacity, international participation, infrastructure dependence, and distributive justice, not merely the existence of domestic policy instruments.

Taken together, these reinterpretations do not displace the original AGILE architecture. Rather, they show that AGILE has become a reference format for composite measurement of AI governance: a four-pillar national index in its baseline form, and a portable methodological scaffold for evaluating governance quality across domains such as evaluation standards, pluralism, certification, sovereignty, and equity (Zeng et al., 10 Jul 2025).