AI and Pay: What the IDB's Wage Scenarios Mean for T&T

The short version

• The IDB's September 2026 scenarios put regional wages 2.3% to 5.3% higher over a decade with worker mobility, or 13.5% to 20.9% lower without it.
• Wholesale and retail, plus manufacturing including mining and quarrying, accounted for about 155,000 T&T jobs in January to September 2025. That is the size of those sectors, not a count of jobs AI can replace.
• Employers can start by asking workers which tasks need help, funding practice time and explaining how any saved time would affect duties and pay.
• A small, supervised trial can reveal problems with a tool. It cannot establish that a job is safe or that wages will rise.

What the IDB has published

The IDB Group brought government officials and technology executives together in New York during UN General Assembly week. Its 21 September 2026 release sets out scenarios from a forthcoming regional report. The full report, From Digitalization to Artificial Intelligence: Turning Promises into Productivity, is scheduled for November.

The release also links a public key-findings attachment. We could read the release during this review but could not retrieve that attachment. The account here therefore uses the release's stated ranges and conditions; it does not claim to have audited the model.

With broad adoption and large productivity effects, the IDB puts the regional GDP gain after a decade at 5.1%. With limited adoption and small productivity effects, it gives 0.3%. The wage scenarios turn on workers' ability to move into growing job areas: gains of 2.3% to 5.3% with mobility, against falls of 13.5% to 20.9% without it. These are conditional regional results. They are not probabilities, sector targets or estimates of an individual worker's future salary.

A person moving from handwritten stock records to checking digital orders is one possible example of a change in duties. Moving to another employer or occupation may require different skills, transport or qualifications. A business cannot settle those questions simply by buying software.

What the local figures actually measure

The Central Bank's Annual Economic Survey 2025 provides a useful starting point. Table 5, using Central Statistical Office data for January to September 2025, records 113,800 people in wholesale and retail and 41,200 in manufacturing, including mining and quarrying. Adding them gives 155,000, or about 27.4% of the table's 565,100 total. Figures are rounded.

Those are broad industry categories. The people counted may sell goods, move stock, repair equipment, supervise production or keep accounts. Neither the table nor our addition measures their exposure to AI. It would be misleading to label all 155,000 jobs at risk.

Table 6 also needs care. Seven manufacturing wage agreements analysed by the Bank had an average yearly increase of 2.4% for 2025; five trade and repairs agreements averaged 3.0%. The median across all the agreements analysed was 2.7%. These are negotiated increases in a small set of registered agreements, not average pay growth for every worker in either industry. They do not show what AI has done to wages.

The Bank's newspaper vacancy indicator fell 11.8% during 2025. It counts advertisements in the Daily Express, Newsday and Guardian. That suggests weaker demand in the jobs advertised there, but it misses hiring through other routes. It cannot tell a trainee whether a suitable vacancy will be available in their town.

Training needs a place and time to happen

The joint ILO and World Bank paper Buffer or Bottleneck?, published in July 2024, estimates that 26% to 38% of regional jobs could be exposed to generative AI. Its narrower ranges are 8% to 14% for potential productivity gains and 2% to 5% for possible full automation. Exposure is an assessment of tasks and capabilities, not an observed job loss.

The same study estimates that gaps in digital access could prevent roughly 17 million workers in potentially benefiting jobs from realising those gains. That regional estimate does not measure T&T's current connectivity gap. For a local employer, the practical question is whether staff have a suitable device, a reliable connection and paid time to learn. Asking people to practise at home may shift the cost onto them.

The World Bank's October 2026 regional economic review reports an AI preparedness score of 0.44 for T&T, compared with 0.69 for its non-LAC OECD comparison group, using 2023 data. This broad index is another reason to examine skills and infrastructure. It cannot predict a factory's return on an AI tool or fill in the missing country wage estimate. Our review of that report explains the measures in more detail.

Start with the people doing the work

Adrian Dunkley's method People First, Problem Second, AI Last, as described in this network's methods library, starts with the human need before choosing technology. Applied here, the first discussion is with the worker who knows where orders get stuck, where records are unclear and where checking already takes too long.

For example, suppose a shop wants clearer replies to routine delivery enquiries. Write down the problem, choose a few made-up enquiries and compare an AI draft with the current reply. Ask the staff member to identify missing details and misleading promises. A shared template may solve the problem more cheaply. A trial should allow that answer.

Before changing duties, explain the proposed training, who pays for it and what alternative work might be available. Give staff and their representatives a chance to respond. These are editorial suggestions for a fairer trial, not steps claimed by the IDB or a promise that redeployment will always be possible.

Where a supervised trial could help

Drafting a supplier email or summarising a non-sensitive internal note may be a manageable starting point. The benefit depends on how much correction the draft needs. Time the whole task, including preparing inputs, checking the answer and fixing errors. A response produced in seconds may still take longer to use than the existing process.

Experienced staff can help decide what to test and what to leave alone. A cashier may spot a misleading returns promise; a production supervisor may recognise an omitted exception in a shift summary. That knowledge deserves time and recognition in the trial. An AI tool should not silently turn checking into extra unpaid work.

Even if a business saves time, higher wages do not follow automatically. Managers and workers still need to discuss how the gain will affect pay, hours, workloads and service. The IDB's regional range cannot answer those workplace questions.

Keep the trial away from high-stakes decisions

Use synthetic examples first. Do not paste customer details, employee records or confidential supplier terms into an unapproved service. Before using real business information, establish who may access it, where it is stored and whether the provider may retain or reuse it. Someone with authority to approve the data use should make that decision.

Leave machine settings, safety instructions, quality release decisions and employment decisions out of an informal chatbot trial. A plausible mistake there can have consequences far beyond a slow email. Existing qualified review and safety procedures still apply.

The labour data has gaps too. The Central Bank reports retrenchment notices covering 224 people in 2025, down from 265 in 2024. It cautions that this measure excludes some small redundancies and other separations, including non-renewed temporary contracts. A low notice count does not establish that workers face little disruption, and this series does not attribute losses to AI.

A practical first month

The dates below are suggested planning dates for a small workplace exercise. They are not official deadlines, and the exercise is not a test of system reliability.

1. By 16 October, ask the people doing the work to list recurring tasks that are slow or frustrating. Pick one low-risk problem, and check whether a clearer form or template would solve it first.
2. Write down the training time, approved tools, data restrictions and named reviewer. Ask what happens to the worker's duties if the task changes; keep the answer available to everyone involved.
3. Over five working days, record the current task time. Over another five, use synthetic or approved examples with an AI draft and check every result. Log corrections, serious errors, total time and costs. Try to compare similar work, while noting that two short periods cannot isolate the tool's effect.
4. By 6 November, discuss the results with staff. A useful time saving with acceptable reviewed outputs may justify a longer supervised trial. Any material error, privacy concern or extra checking burden needs attention before expansion. A small sample with no errors is not proof of safety.

If an adjacent skill would help, agree a realistic practice schedule. Two paid hours a week for six weeks is one possible starting proposal, not an evidence-based minimum or a guarantee of a new job. Check actual vacancies and entry requirements before asking someone to spend money on a course.

What remains unanswered

We have not verified a T&T wage estimate from the IDB model. The regional scenarios, the older ILO exposure study and the Central Bank's local labour statistics answer different questions. Combining them does not produce a forecast. The November report may provide more detail about the assumptions, but local task studies and evidence from workplaces will still matter.

The immediate question for a worker is more concrete: who will pay for the training, and what work will be available afterwards? An employer should be able to discuss that before asking staff to welcome an AI rollout.

Questions readers ask