For seven years, ministers, permanent secretaries and CEOs across the Caribbean asked me some version of one question: where do we actually start with AI?
My answer used to be "it depends": on your sector, your data, your regulator and your staff. That was accurate and no help at all to a Minister of Finance in Barbados or a Permanent Secretary in Trinidad who has to put a concrete business case for an AI budget line before a Finance Committee. I built the LATAMC AI Playbook to give a better answer.
What the Playbook contains
The LATAMC (Latin America and the Caribbean) AI Playbook is a searchable library of more than 15,000 AI use cases, mapped across the sectors that matter in the region: tourism, agriculture, financial services, public health, education, criminal justice, port logistics, energy and disaster response.
Each use case is drawn from a real deployment, academic literature with measured outcomes, or government implementation records. Each records what was built, what it achieved, under what conditions, what it cost to run, and where it failed. The failures are included because a team that does not know how a use case breaks is not ready to deploy it.
Fifteen thousand use cases means that within any sector you can find specific approaches to specific problems, with the evidence of what worked and what did not.
The regional focus is deliberate. Most AI playbooks assume deep capital markets, mature data infrastructure, a large domestic tech sector and well-staffed regulators. Caribbean economies have smaller, more interconnected financial systems, newer data infrastructure, capable but stretched regulators, and economies weighted towards tourism and agriculture. A playbook written for Germany or Singapore is sound for Germany or Singapore and hard to apply in Jamaica or Barbados.
Why the number of use cases matters
The standard route to an AI strategy goes like this: hire a consultancy, receive a 100-page document naming a handful of priority areas, then spend eighteen months working out what those areas mean in practice, until the effort stalls.
The priority areas are usually sensible. The gap is in what each one means in practice. "AI in financial supervision" could mean transaction monitoring, compliance automation, systemic risk early warning, credit scoring for micro-lenders or fraud detection in remittance flows. Each needs different data, infrastructure, approvals, timelines and risk tolerance.
With thousands of documented cases, a team can look up what a central bank in a comparable economy did with a particular transaction monitoring system: what it cost, what data it needed, what went wrong in implementation, and what it achieved after two years. That is the level of detail a budget submission needs.
Three use cases a Caribbean government can start on now
These three have solid evidence, fit Caribbean conditions, and rely mostly on data governments and large firms already hold.
Tourism demand forecasting
Hurricanes, health scares, economic cycles in source markets and changing travel habits make tourism demand hard to forecast with traditional statistics, and many hotels plan on assumptions weeks out of date. Machine learning models that combine booking data, flight searches, social media sentiment, weather forecasts and source-market economic indicators have improved forecast accuracy in several island tourism economies.
Better forecasts feed staffing, room pricing, marketing spend and food ordering. The data is already held by tourism boards, national airlines and the larger hotel groups. The remaining barrier is technical.
Agricultural climate risk scoring
Farmers need credit to adapt to a more variable climate, and lenders cannot price that credit because they cannot assess risk field by field. Climate risk scoring models combine satellite images, past crop yields, local climate projections, topography and soil data into a risk score for each field. Similar approaches have helped lenders extend credit to smallholders in East Africa who had been judged too risky.
The method transfers; the model does not. A model built for maize in Kenya will not score sugarcane in Barbados or bananas in Saint Lucia, so the local data work has to be done here. The Playbook includes guidance for that localisation.
Public safety resource planning
Crime follows patterns by time, day, place, weather and economic stress, and models can learn them well enough to plan patrols better than intuition alone.
This must not repeat US-style predictive policing, where models trained on biased arrest records sent police back to the same neighbourhoods and reinforced the bias. The responsible version works only on aggregated time and place data for planning. It never scores individuals. A typical output reads: "This district on a Friday evening before a public holiday has historically needed 40 percent more officers than the baseline." Several Caribbean territories have the data to do this today.
Key point
Every use case in the Playbook carries its governance requirements with it. Accountability is designed alongside the technology from the first draft, because it is much harder to add once a system is running.
Governance is built into every use case
Each sector section sets out the data rights involved, the consent approach, the regulatory approvals typically needed in CARICOM jurisdictions, the transparency standard that should apply, and the route of appeal for people affected by automated decisions.
This matters more in small states. A flawed credit scoring model in Jamaica reaches a far larger share of adults than the same model would in the United States, and there are fewer alternative lenders to serve the people it wrongly excludes.
How to use the Playbook
The Playbook is a search and filter tool, not a document to read front to back.
- Start with a problem: reduce claims fraud, improve student retention, forecast clinic demand, cut port clearance times.
- Filter by sector and problem type to find matching use cases.
- Read the evidence for the most relevant ones, including the failures.
- Shortlist two or three that match your data and staff capacity.
- Run the Playbook's readiness assessment, which scores data maturity, technical capacity, governance and change management against the use case.
- Build a pilot around the strongest candidate.
The readiness step guards against the most common failure: technically sound AI dropped into an organisation that was not ready to run it.
The limits of a use case library
A use case from another country tells you what is possible and roughly what it takes. It does not tell you whether your ministry's data is clean enough, whether your procurement rules allow the vendor you need, or whether your staff will use the result. Evidence from comparable economies is also uneven: some sectors have hundreds of well-documented cases, and others, such as small-island port logistics, have few. The Playbook shortens the search, and a pilot is still needed.
What to do next
- Permanent secretaries and heads of agencies: pick one operational problem your agency must improve this financial year and write it as a single sentence with a measurable target, before looking at any AI tool.
- Agency data or IT leads: list the datasets you hold for that problem, who owns each one, how far back it goes and how complete it is. This takes about two weeks and decides which use cases are realistic.
- Policy teams: shortlist two or three matching use cases from the Playbook and summarise each on one page: cost, data needed, results elsewhere and known failure points. Use that as the basis of the budget submission.
- Ministries of national security: if you pursue crime pattern analysis, publish in advance that the system uses only aggregated time and place data and will not score individuals, and name the oversight body.
- Agricultural lenders and development banks: start collecting field-level yield and loss data from borrowers now, since localised climate risk scoring cannot be built without it.
Frequently Asked Questions
How do I judge whether a use case from another country will work here?
Check four things: whether the source economy is comparable in size and data maturity, whether the outcome was measured against a baseline, whether anyone independent evaluated it, and whether the cost of running it is reported. A use case that fails two or more of these is a lead to investigate, not evidence to budget on. Ask the implementing agency directly where you can; public bodies often share lessons with peers.
How is a use case library different from a national AI strategy?
A national AI strategy sets goals, priorities, funding and institutions at government level. A use case library sits underneath it and shows how a specific problem has been solved elsewhere, with data needs, cost and results. A strategy without that layer tends to stall at implementation, and a library without a strategy produces scattered pilots that never scale.
Are there other tools for assessing a country's AI readiness in the region?
Yes. UNESCO's Readiness Assessment Methodology evaluates a country's legal, social, economic and technical preparedness for ethical AI, and several Latin American and Caribbean countries have completed it. The Latin American Artificial Intelligence Index (ILIA), led by Chile's National Centre for Artificial Intelligence (CENIA), compares AI capacity across the region each year. Both are useful benchmarks alongside an organisation-level readiness assessment.
What is the most common mistake governments make when choosing a first AI project?
Choosing the most visible problem, not the one with the best data. A high-profile project on messy or incomplete data usually fails in public and sets the programme back years. A first project should have clean historical data, a clear owner, a measurable baseline and results within six to twelve months.
How can a government agency get access to the LATAMC AI Playbook?
Contact Adrian Dunkley's team through the contact page on this site and describe your organisation and the problem you want to address. It helps to have the one-sentence problem statement and a rough inventory of your data ready, since those decide which parts of the Playbook are most useful.