TL;DR
- On 6 October 2026 the World Bank said the Caribbean and Latin America are held back less by access to AI than by absorptive capacity: the skills, management and data a firm needs to use it. Regional growth is projected at 2.2% in 2026.
- Guyana's 23.7% growth estimate is a national figure. It does not establish that every firm has cash, suitable data or trained staff for AI.
- We apply Automation Fragility, one of Adrian Dunkley's published methods: AI amplifies the weaknesses a process already has.
- Try 20 synthetic or safely de-identified examples from one low-risk process. Count checking time and mistakes, then use a separate set for evaluation. This is a diagnostic exercise, not deployment approval.
What happened on 6 October
The World Bank release of 6 October 2026 projects regional growth of 2.2%, after 2.4% in 2025. It says firms are adopting AI, but often without integrating it into their central processes. The barriers include management capability, workforce skills and the ability to change how work is organised.
The report calls this absorptive capacity: the ability to make productive use of a technology. Access to a model and the ability to use it well are separate questions. The Bank's regional finding is a useful starting point for a firm-level review, not a diagnosis of every business.
The report’s growth table lists Guyana at 23.7% for 2026, the highest estimate among the countries shown. The table uses data available by 23 September and labels 2026 as an estimate. That timing matters: these are not final annual results.
The numbers behind the finding
Start with how many people use the tools. The World Bank reports a median of 17% of working-age adults across the region using generative AI, roughly half the rate in the United States and Canada. The release gives no Guyana figure.
Next, who the tools touch. About 8% of the workforce holds high-skill, knowledge-intensive jobs that AI could enhance, around 10% work in routine cognitive occupations, and about a quarter hold routine manual jobs that may be exposed. These are regional shares. We would not apply them to a Georgetown payroll without counting.
The same table records Guyana's growth at 33.8% in 2023, 43.8% in 2024 and 19.3% in 2025, followed by forecasts of 18.7% in 2027 and 15.3% in 2028. These national rates cannot tell us which firms can afford training or whether a particular team has usable records.
| Indicator | Figure | Source |
|---|---|---|
| Regional growth, 2026 estimate | 2.2% (2.4% in 2025) | World Bank press release |
| Adults using generative AI, regional median | 17%, about half the US and Canada | World Bank press release |
| Guyana real GDP growth, 2026 estimate | 23.7% | World Bank October 2026 report, growth table |
| Guyana internet users, 2024 | 83% (Americas: 86%) | Internet Society Pulse |
| Guyana broadband against mobile download speed, 2026 | 148.55 Mbps and 22.27 Mbps | Internet Society Pulse |
A business may use a chatbot for routine questions while leaving its stock records and invoices untouched. That could still save time in customer service; it does not tell us whether the rest of the business has become more productive. Measure the task that changed, including correction and supervision time, before generalising.
Connectivity also needs a local check. Internet Society Pulse reports 83% internet use in 2024 and a low-consumption basket costing 1.16% of average national income in 2025. Its 2026 speed figures are 148.55 Mbps for broadband and 22.27 Mbps for mobile, based on an average of 300 users. These figures do not describe every household, field location or budget. Test the actual connection and an offline fallback where staff will work.
Automation Fragility, applied to a boom economy
Automation Fragility is the term in Adrian Dunkley's published FAQ for AI amplifying existing weaknesses in data and processes. We apply that idea here: a written procedure can reveal missing checks before software repeats the same error at scale. Clean records help, but they do not guarantee that a model is accurate or that automation is worthwhile.
Consider a hypothetical company taking on 40 staff in a quarter. Procedures that once lived in two experienced people's heads may need to be written down. Ask a new hire to follow one low-risk process, with a supervisor present, and note where the instructions need explanation. The question is about the process before it is about AI.
Picture a hypothetical Georgetown distributor with 30 staff and a fast-growing order book. Orders arrive by WhatsApp, prices sit in one manager's phone, and delivery notes are written by hand. An assistant that drafts replies to customers will read a price that is out of date and state it with total confidence. A useful first step is a controlled price list with a named owner and a check before a quote is sent. Maintaining it takes staff time. An assistant should then be tested against that source, with uncertain answers held for review.
Shadow Literacy, another term from the same FAQ, means recognising when a plausible answer is wrong. Staff need time and clear source material to do that checking. Counting only the seconds a model takes to draft an answer would miss much of the work.
What this means in Guyana
Guyana already treats AI as national business. At a Georgetown seminar on 23 April 2026 with Mount Sinai, President Irfaan Ali argued that AI should play a part in developing the health system, according to the Mount Sinai newsroom. The seminar, on turning promise into practice by integrating AI into health care, covered foundations, ethics and decision-making, and that theme is a fair summary of the World Bank's question.
For teachers, UNESCO’s competency framework offers 15 competencies across five dimensions and three levels: Acquire, Deepen and Create. A school can use it to plan training, then adapt examples to its curriculum and students. It is a framework for professional development, not a finding that a particular classroom product works.
High-stakes settings need a different standard of review. A short office exercise cannot validate clinical decisions, policing, credit eligibility or access to public services. Those uses need qualified oversight, appropriate legal review and evidence matched to the consequences of an error.
Language deserves its own test. If customers use Guyanese Creole as well as English, include examples of both in a safely designed evaluation. Ask local staff and users to review whether the meaning was preserved. A model-directory search cannot establish performance, and a few fluent examples cannot establish reliability.
Disclosure: Maestro AI Labs is part of the StarApple AI network and develops Caribbean data and AI tools. We have a commercial interest in this field. Readers can find the network's own descriptions at StarApple AI and regional discussion at AI Guyana.
What to do this week
- Choose one low-risk repeated task by 14 October, such as drafting an answer from a public product catalogue. Write its steps, sources and escalation rules. Keep medical, financial eligibility and other consequential decisions out of this exercise.
- Test 20 synthetic or properly de-identified examples. Define correct answers and important failure types before testing. Time the current process, including review and correction, and record the range as well as the median.
- Try an assistant offline from production, using an approved tool and a named checker. Do not send its outputs to customers or let them change records. Record all checking time and errors, including those that would matter even if they are rare.
- Use those examples to improve the process, then evaluate a separate unseen set. Reusing the same cases can make results look better through familiarity. Include realistic English and Guyanese Creole examples where relevant, without personal identifiers.
- Review what was learned by 23 October if the work is ready. Decide whether the next step is better source data, clearer instructions, staff training or a more formal trial. Confirm data access, retention and legal requirements before any use of real personal records.
These dates are a suggested schedule, not an official deadline. The exercise costs staff time and may need approved software. Twenty examples can expose problems; they cannot prove a product is reliable, establish a national readiness score or justify deployment. Choose success criteria for the task and its risks rather than an arbitrary 25% reduction in rework.
Where this can go wrong
Protecting personal data is a prerequisite, not something to postpone until a pilot looks promising. Guyana's Parliament publishes the Data Protection Act 18 of 2023. This article has not established its current commencement or institutional arrangements. A project owner should confirm the applicable law and oversight with qualified local advisers and current official records. Do not infer that uncertain commencement means no privacy duties or remedies exist. Use synthetic examples first; restrict access and agree retention and deletion rules before handling real data.
Where this analysis is weakest
We checked the World Bank report's growth table and workforce discussion as well as its release. Its labour categories describe regional exposure, not predicted job losses or a Guyanese payroll. A small test of one process cannot measure an organisation's overall capacity, and low rework in that sample does not prove there are no other barriers. National internet averages also hide variation. The proposed exercise should guide further questions rather than settle them.