Caribbean AI AssociationCAIA Research

Risk Index - Job Loss Model (JLM)

Job Loss
Model

Explore interactive results →

Calculate a workforce scenario from assumptions about AI use, task replacement, transfers to other roles and new positions. The result measures work capacity. It does not forecast national job losses.

Research edition 0.1 • Occupation data and scenario calculation

This example uses full-time equivalents (FTEs), not a count of people.

435gross pressure
−217.5redeployment
−100new positions
117.5net pressure
427occupation groups
3,265task rows in the source file
0.27median score across occupation groups

The calculator uses assumptions. It does not estimate national job losses. Read the scope

Method

How the assessment works

Select a step to read its instructions.

The source has 427 occupation groups and 3,265 task rows. These blocks represent tasks, not workers.

Start with an occupation score.

An exposure score describes the potential for AI to affect tasks. A national estimate also needs employment counts for each occupation.

Step 1 of 4 · Select a step or start playback

Workforce calculation

Calculate a workforce scenario

The example covers three years. One full-time equivalent (FTE) represents the work assigned to one full-time position. Employment pressure is a calculated reduction in required work. The calculator treats the exposure score as a share of work capacity. This conversion needs validation. Adoption, substitution, redeployment and new positions are assumptions.

Loading interactive data…

The default exposure is an assumption of 0.29. The source median is 0.27.

One FTE represents the work assigned to one full-time position.
The model uses this score as the affected share of work capacity.
Share of exposed work for which AI is used.
Share of adopted exposed capacity that replaces labour.
Share of gross pressure absorbed through work in other roles.
New FTEs as a percentage of the initial workforce.

435.0Gross FTE pressure
217.5FTE pressure absorbed through redeployment
100.0New FTE positions assumed

Net employment pressure

117.5 FTE 1.175% of the initial workforce

A negative result means net expansion under the assumptions. FTEs measure work capacity. They do not identify people who will lose jobs.

Effect of adoption

The orange point marks the selected adoption value. The line changes adoption only. It does not represent calendar time.

Effect of adoption and substitution

Select a point to read its values. Use Rotation or Tilt to change the view.

Selected scenarioPositive pressureExpansion
View the values as a table
AdoptionGross FTE pressureRedeployed FTENew FTENet FTE pressure

Occupation groups by exposure category

The bars count occupation groups, not workers. The category names come from the ILO file. The scale is 0–427 groups.

View all occupation scores
ISCO-08OccupationMeanTask-score SDCategory

Related article

Why an AI exposure score does not predict job loss

Read the article

Research by Adrian DunkleyFull studyPurpose, data, method, results and limits.

Study type: Analysis of occupational exposure and a deterministic scenario model. Question: How do adoption, substitution, redeployment and new positions change the calculated workforce requirement?

1. Exposure, employment pressure and job loss

An exposure score describes the technical potential for AI to affect tasks. Employment pressure is a calculated reduction in required work under stated assumptions. Job loss is the observed loss of a job or position over a defined period. JLM links exposure to employment pressure. It does not measure job loss. The ILO also distinguishes exposure from an employment forecast. [ilo-explorer]

The model uses full-time equivalents, abbreviated as FTEs. One FTE is the amount of work assigned to one full-time position. FTEs measure work capacity, not a count of identified people. A result can therefore include a fraction of an FTE.

The calculator shows how each assumption affects the result. Employers, workers and public agencies can use it to identify the evidence needed for a local study. Local measurements can replace the assumed values after validation.

2. What the occupation data show

The source is the supplementary workbook for the ILO’s 2025 occupational-exposure study. It contains 3,265 task rows linked to 427 four-digit ISCO-08 occupation groups. ISCO-08 is the International Standard Classification of Occupations. [ilo25] [ilo-data]

The workbook repeats occupation means and standard deviations across task rows. CAIA checked these repeated values for consistency and kept one record per occupation code. Thus, an occupation with more listed tasks does not receive more weight in the occupation statistics.

The median exposure score is 0.27, and the mean is 0.2965. These statistics give each occupation group equal weight. Thirteen groups are in the source’s highest exposure category. Data Entry Clerks score 0.70. Typists and Word Processing Operators score 0.65. These are exposure scores, not percentages of Caribbean workers who will lose their jobs.

The source classifies 231 groups as Not Exposed and 84 as Minimal Exposure. Another 112 groups fall across Gradients 1–4. The 112 groups represent 26.2% of the 427 occupation groups, with each group counted once. This percentage does not describe the share of workers. This release has no national employment counts for these groups.

The assessment retains the source category and the standard deviation of task scores within each occupation. The standard deviation describes the spread of those task scores. It is not a probability of job loss.

3. Model terms and calculation

The model calculates employment pressure for each occupation group o. Its terms have these meanings:

TermMeaning
NoInitial workforce in FTEs.
eoExposure score used as a substitute for the affected share of work capacity.
aoAdoption: the fraction of exposed capacity for which AI is used during the selected period.
soSubstitution: the fraction of adopted exposed capacity that replaces labour.
roRedeployment: the fraction of gross pressure absorbed through work in other roles.
GNew FTE positions during the same period.

Gross employment pressure is D = Σ Noeoaoso. Pressure after redeployment is U = Σ Noeoaoso(1 − ro). Net employment pressure is L = U − G. The symbol Σ means the sum across occupation groups.

A negative L means net expansion under the selected assumptions. The model retains negative results. The percentage pressure index is 100 × L / ΣNo. The total initial workforce must be greater than zero.

The exposure score is not a measured share of hours that AI can automate. Its use as a work-capacity fraction is an extra model assumption. Local task-time studies must test this conversion. The calculator limits e, a, s and r to values from zero to one. Users enter new positions as a percentage of the same initial workforce.

The default example covers three years and uses 10,000 FTEs. It assumes exposure of 0.29, adoption of 50%, substitution of 30%, redeployment of 50%, and new positions equal to 1%. The exposure assumption differs from the source median of 0.27.

The example gives 435 FTEs of gross pressure. Redeployment absorbs 217.5 FTEs, and the example assumes 100 new FTEs. Net pressure is 117.5 FTEs, or 1.175% of the initial workforce. These values describe a calculation, not observed layoffs.

4. How to read the charts

The animated line chart changes adoption from zero to 100%. It holds the other inputs fixed. The animation tests sensitivity to adoption. It does not predict annual changes.

The 3D chart changes adoption and substitution. Its height shows net FTE pressure at the selected exposure, redeployment and new-position values. Orange shows positive pressure. Blue shows net expansion. Rotation changes the view, not the calculation.

At zero adoption, gross AI-related pressure is zero. Separately assumed new positions can still produce net expansion. At 100% redeployment, pressure after redeployment is zero. If no new positions are assumed, net pressure cannot be negative. Calculation tests check these limits.

The model treats G as a separate input. It does not require new positions to increase with adoption. This assumption matters when a user changes the adoption rate.

For one occupation group, the sensitivity to adoption is N × e × s × (1 − r). A larger workforce or substitution rate increases this sensitivity. More redeployment reduces it. The equation gives this relationship directly.

The model does not assign probabilities to the input values. A chart surface or slider range is therefore not a confidence interval. JRC guidance supports the explicit examination of these design choices. [jrc]

5. Evidence and limits

The World Bank and ILO study identifies potential for both AI assistance and automation in Latin America and the Caribbean. Digital access affects who can use that assistance. Some labour microdata in its reproduction package are restricted. JLM does not reproduce its national employment estimates or replace missing records with invented data. [wb-lac] [wb-repro]

Task-based economic theory separates displacement, productivity gains and new tasks. JLM uses that distinction, but its equation is not a fitted model of the whole economy. It excludes prices, trade responses, migration, fiscal transfers, job-search delays and feedback between demand and adoption. [automation]

Workplace studies do not supply a universal conversion from time saved to layoffs. Brynjolfsson, Li and Raymond report higher productivity in their support-agent setting. Becker and colleagues report slower completion in their early-2025 developer trial. These studies support task-specific measurement, not one effect for all Caribbean sectors. [work] [metr]

6. Evidence needed for a workforce estimate

Before deployment, an employer study needs a task list, paid hours, output quality, vacancies, new hires, departures and internal transfers. A phased introduction can support a comparison. The study must examine selection into adoption and whether comparison groups had similar earlier trends.

The study must measure both task-time changes and the use of any released work capacity. Redeployment requires a defined paid role and evidence of work in that role. Attendance at training does not establish successful redeployment.

A national estimate needs employment counts by occupation and industry, consistent FTE conversions, survey weights and adoption measures for the same period. Reports must identify missing occupations. Income and participation data provide context, but they do not replace employment counts or establish causation. [participation] [income] [api]

Before the analysis, the plan must fix the outcome, period, sampling frame and comparison method. Later periods or separate organisations must provide validation data. A future probabilistic model must report forecast error, probability calibration and coverage of worker groups.

The composite-indicator handbook can inform reporting. Governance and worker participation require a suitable local process. [composite] [ram] [ethics]

JLM currently helps users examine assumptions and plan data collection. The existing Jamaica economic study is a separate scenario analysis. Users must reconcile populations, periods and mechanisms before they combine outputs from different models. They must not add the losses from both models without that check.

Data

Data and calculation files

Download the data or repeat the calculations.

Data retrieval date: 10 November 2026. The study uses the default settings. The controls change only the assessment results.

View the default data table
Source categoryOccupational groups
Not Exposed231
Minimal Exposure84
Exposed: Gradient 244
Exposed: Gradient 338
Exposed: Gradient 117
Exposed: Gradient 413
Sources and references 15 sources

Sources checked on 10 November 2026. The notes state which material CAIA checked and how the study uses it.

  1. Gmyrek, P., Berg, J., Kamiński, K., et al. (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper 140.Abstract, methods overview and supplementary occupational file; DOI 10.54394/HETP0387.
  2. Gmyrek, P., et al. (2025). Supplementary ISCO-08 scores.Downloaded Final_Scores_ISCO08_Gmyrek_et_al_2025.xlsx; Sheet1, 3,265 rows. Numeric fields extracted; source workbook remains with its publisher.
  3. Gmyrek, P. (2026). AI Exposure Explorer.Interpretation and data coverage; updated 7 October 2026.
  4. World Bank and ILO (2024). Generative AI and Jobs in Latin America and the Caribbean: Is the Digital Divide a Buffer or Bottleneck?At a glance and key messages. Publication summary; regional ranges are not country forecasts.
  5. Gmyrek, P., Winkler, H., and Garganta, S. (2024). Buffer or Bottleneck? Reproducibility package.Data access restrictions, source datasets and reproduction requirements. Study DOI 10.1596/1813-9450-10863.
  6. Acemoglu, D., and Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2), 3–30.Task-based framework and displacement/reinstatement mechanisms, pp. 3–10. DOI 10.1257/jep.33.2.3.
  7. Brynjolfsson, E., Li, D., and Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889–942.Abstract and introduction: staggered deployment, 5,172 support agents, 15% average productivity increase in that setting.
  8. Becker, J., Rush, N., Barnes, B., and Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.Historical randomised study, 16 developers and 246 tasks; not a statement about current tools or all developers.
  9. European Commission, Joint Research Centre (2025). Composite Indicators: 10 Step Guide.Steps 1–10: construct, missingness, weighting, aggregation, sensitivity and external comparison.
  10. OECD, European Union and EC-JRC (2008). Handbook on Constructing Composite Indicators.Publication metadata and methodological scope; used with the JRC implementation guide.
  11. World Bank WDI / ILO. Ratio of female to male labour force participation (%), SL.TLF.CACT.FM.ZS.Modelled ILO estimate; participation includes people employed and people seeking work. It does not measure equal pay.
  12. World Bank WDI. GDP per capita, PPP, constant 2021 international dollars, NY.GDP.PCAP.PP.KD.Used only for the income-dependence diagnostic, not in the foundation score.
  13. World Bank. Indicators API documentation.Version 2 API; query URLs, response dates and hashes retained with the analysis.
  14. UNESCO. AI Readiness Assessment Methodology.What is RAM and how it works; current overview checked 10 November 2026.
  15. UNESCO (2021). Recommendation on the Ethics of Artificial Intelligence.Values, oversight and ethical impact assessment, especially paragraphs 49–53. Normative guidance, not a regional statute.

Cite this research

Dunkley, A. (2026). Risk Index - Job Loss Model (JLM). Caribbean AI Association. Version 0.1, 10 November 2026. https://caribbeanaiassociation.com/research/observatory/jlm/

Download BibTeX

Use the project title and version in your citation. Credit the original data sources. State any changes that you make. Third-party materials retain their original rights. This page does not claim peer review or provide a DOI.

BibTeX
@misc{dunkley2026jlm,
  author = {Dunkley, Adrian},
  title = {Risk Index - Job Loss Model (JLM)},
  year = {2026},
  publisher = {Caribbean AI Association},
  note = {Version 0.1; 10 November 2026},
  url = {https://caribbeanaiassociation.com/research/observatory/jlm/}
}