1. What the score measures
EAEI is a proposed framework for research on the economic effects of AI. The intended outcomes are sustained productivity, job quality, wider participation and the distribution of gains. The current pilot does not measure these outcomes. It calculates a foundation score from selected conditions that could support AI use.
The full framework has five evidence areas: access and infrastructure, skills and adoption, productivity and quality, distribution and job changes, and accountable governance. The pilot covers selected measures of access and participation. It does not score skills, adoption, measured productivity, wages, business ownership, safety or the distribution of benefits. Missing evidence does not receive a zero score.
The IMF AI Preparedness Index examines conditions for adoption. UNESCO RAM examines the capacity to govern AI. EAEI proposes a later test of the link between country conditions and measured benefits. This pilot does not establish that link. [imf] [ram]
2. Countries, data and missing values
The study includes the fourteen sovereign CARICOM members and the Dominican Republic. It does not cover all of the Caribbean or Latin America. [caricom] CAIA retrieved the World Bank Indicators API data on 10 November 2026. The reference snapshot is 2024. [api]
For each input, the snapshot uses the latest available value from 2022 through 2024. A country receives no overall score if any required input is missing. The historical view requires every input from the exact year shown. It does not carry an earlier value into a later year.
| Input | Measure | Important limit |
|---|---|---|
| Internet use | Internet users as a percentage of the population. | It does not measure AI use or connection quality. |
| Fixed broadband | Fixed subscriptions per 100 people. | A subscription is not a unique person. Mobile access is a separate measure. |
| Electricity access | People with electricity access as a percentage of the population. | Access does not measure outages, reliability or cost. |
| Participation ratio | The female labour-force participation rate divided by the male rate, multiplied by 100. | The ratio does not measure equal pay, job quality or overall participation. |
The input file records indicator codes and observation years. The participation series uses ILO modelled estimates. The inputs are not all direct survey observations. [internet] [broadband] [electricity] [participation]
GDP per person at purchasing-power parity provides a separate comparison with income. Purchasing-power parity adjusts for differences in price levels. This income measure does not enter the foundation score. [income]
3. How to calculate the foundation score
The calculation first converts each input to a value from zero to one. I is internet use divided by 100. B is fixed broadband subscriptions per 100 people divided by 40. E is electricity access divided by 100. P is the participation ratio divided by 100. Each converted value has a lower limit of zero and an upper limit of one.
The digital component is D = √(I × B). The default foundation score is 100 × (D × E × P)1/3. This geometric mean multiplies the three components and takes the cube root. The converted components have no physical units.
CAIA selected 40 fixed subscriptions per 100 people as the broadband reference value. This value is a scaling choice. It is not an internationally validated target or a household coverage target. It does not imply that each person needs a fixed line.
Internet use and fixed broadband each supply half of the digital component’s influence. This limits, but does not remove, overlap between the two measures. The default formula gives equal weight to the digital, electricity and participation components.
A low component reduces a geometric mean more than an arithmetic mean. A zero input produces a zero score. A missing input produces no score. These are design choices, not estimated economic relationships.
The assessment lets you change the component weights and broadband reference value. Each score then updates from the published inputs. The OECD handbook and JRC guide explain why an index should disclose these choices. [composite] [jrc]
4. Results from the published data
Eleven countries have complete inputs under the snapshot rule. Antigua and Barbuda, Dominica, Grenada, and Saint Kitts and Nevis lack the required participation value for 2022–2024. They receive no overall score. The median score of the eleven complete countries is 76.8 out of 100. Each country receives equal weight in this median. It is not a population-weighted regional estimate or a target.
Jamaica’s default score is 78.5 out of 100. Its 2024 inputs are 90.1% internet use, 16.0 fixed subscriptions per 100 people, 97.7% electricity access and an 82.6% participation ratio. The digital component is 60.0 after the broadband conversion. This result depends partly on how the index defines digital access.
Barbados scores 89.7. Its broadband value is from 2022, and its other inputs are from 2024. Haiti scores 30.1. Its fixed-broadband value is particularly low under the selected formula. These results do not measure the ability of an individual business to use AI.
A business that uses mobile connections can operate where fixed broadband is limited. A user must check whether the fixed-broadband measure represents the local problem. The score alone cannot justify a purchase or investment decision.
5. How design choices change the result
The analysis tests 10,000 combinations of three nonnegative component weights that sum to one. The test samples the possible combinations uniformly. The random seed is 20261110. The inputs and broadband reference value remain fixed.
For Jamaica, the 5th–95th percentile score range is 65.3–91.9 across these weight choices. The corresponding percentile range of rank positions is 3–6. These ranges show sensitivity to the selected weights. They are not confidence intervals or probabilities about Jamaica.
Across the eleven complete countries, the Pearson correlation between the default score and the natural logarithm of PPP GDP per person is 0.721. The small, selected sample does not support a causal conclusion. Shared development conditions can affect both measures. Future tests must establish whether the score adds information beyond income differences.
Exact-year coverage decreases from ten countries in 2020 to five in 2024. The historical animation therefore changes both the values and the set of countries with complete data. A regional trend analysis needs the same country set in each year. A comparison for one country needs complete inputs in each selected year.
Source revisions can change values for earlier years. The release retains the retrieved files and file hashes so users can identify the exact data used.
6. What the next study must test
The World Bank and ILO identify digital access as a constraint on potential AI assistance in Latin America and the Caribbean. This supports a study of access. It does not validate the EAEI formula. [wb-lac]
Workplace findings also differ by setting. Brynjolfsson, Li and Raymond studied support agents. Becker and colleagues studied experienced developers in an early-2025 experiment. Neither study supplies an effect that applies to every Caribbean sector. [work] [metr]
The proposed next study uses earlier foundation measures and a sample of establishments selected with known probabilities. It must record adoption dates, verified task time, output quality, costs, paid hours and wages. The design must define sector groups and account for business size and previous productivity.
Measurements before and after adoption need a suitable comparison group. Tests must examine whether the score adds information beyond income. Separate countries or later data must test whether the results apply beyond the design sample.
The current pilot can help public agencies identify missing statistics and questions about access. Employers can combine country data with local task measurements and LAC Eval tests. Economic analysis must separate substitution, productivity and new work. [automation] Claims about economic benefits require evidence of those effects. Users must credit the original data providers and follow each dataset’s licence terms. [licence]