When a headline says AI could affect one in four jobs, it is tempting to turn that into a redundancy forecast. Before anyone puts the number in a staffing plan, ask what the researchers counted.
The International Labour Organization and NASK published Generative AI and Jobs: A Refined Global Index of Occupational Exposure on 20 May 2025. It estimates that roughly a quarter of workers worldwide are in occupations with some exposure to generative AI. The study examines tasks that technology could perform. It does not count jobs already lost.
Read the number before using it
“Exposure” means that parts of an occupation overlap with potential AI capabilities. The ILO's data explainer groups exposed occupations into four levels, using both the average task score and the variation between tasks. It then combines those scores with labour-force data from more than 140 countries.
That leaves several decisions between a technical possibility and a change in staffing. Someone must adopt a tool, connect it to usable information, check its work and decide how to organise the remaining tasks. A business with unreliable records may experience a very different result from one with a well-maintained system, even when both employ people under the same job title.
The working paper combines a Polish task survey, expert review and AI-assisted scoring. It describes a snapshot of existing occupations at the start of 2025, excluding future new jobs and later technical advances. Its authors expect transformation to be more common than complete replacement, while warning that partial automation can still reduce demand for workers. This is research in an ILO working paper, not a measured outcome for a Caribbean employer.
Look inside one working day
Consider a fictional freight office in Kingston. A coordinator drafts shipment updates, checks whether documents agree, calls a supplier about missing information and explains a delay to a customer. Calling the whole job “administration” conceals the different demands of those activities.
A useful experiment might begin with a draft update from approved shipment fields. Checking a disputed document requires knowing which record is authoritative. A supplier call may reveal information absent from every file. An upset customer's question may need a decision the coordinator cannot make alone.
None of these observations proves that a particular model can or cannot perform a task reliably. They give the team specific things to test. The same exercise could help a small distributor in Belize or an office serving Spanish-speaking customers in the Dominican Republic. The test should use the actual language and process of the workplace, rather than a translated demonstration with local place names added.
Make a task map with the people doing the work
As an original practical application, ask one willing colleague to list the recurring activities in a typical day. Keep the exercise about work design, not ranking employees by replaceability. For each activity, write down:
- The output someone actually needs.
- The information required and whether an approved AI tool may receive it.
- The mistakes that would matter, and who could detect them.
- The judgement, conversation or physical action still required.
Choose one low-risk activity where those answers are clear. Use fictional or properly de-identified records. Compare the existing process with an AI-assisted draft, including checking and correction time. Record failures as carefully as successes. Agree in advance what would make the team stop the trial.
Do not upload personnel records to ask a chatbot who should lose their job. An occupational estimate offers no reliable basis for judging an individual worker. If the exercise raises staffing or employment questions, involve the appropriate people and processes rather than treating the model's answer as authority.
Bring a better question to the next meeting
Ask which tasks could improve, what evidence would demonstrate that improvement, and what staff would need to learn. Ask employees what would become harder if the easier parts were automated. A role left with only difficult exceptions may need different training, support and workload limits.
For a Caribbean or Latin American business, the global figure is a reason to examine the work closely. A local hiring, training or investment decision needs local evidence. Start with one task you can describe and check, then use what the trial reveals to plan the next step.
Sources & further reading
ILO and NASK: Generative AI and Jobs, Working Paper 140 (20 May 2025) ↗Full working paper: methods, limitations and conclusions ↗ILO data explainer: how occupational exposure is measured ↗Sources checked 2026-10-20. Examples and recommendations are original editorial analysis.