A woman in Kingston has sold produce from the same market stall, six days a week, for twelve years. She pays her stall fees on time. She keeps a mobile money account and pays her child's school fees in full every term. She has never defaulted on a loan, because no one has ever lent her money.
To a conventional credit score, no history looks the same as bad history. So she gets no loan, has no capital to grow, and her business is the size it was in year one. Her twelve years of discipline are invisible to the system that decides who gets credit.
Her case is common. Estimates vary by country and method, but across CARICOM many adults have either no formal account or an account with no access to credit. AI credit assessment is starting to change that for some of them. This piece covers what works, what still blocks progress, and what lenders, regulators and borrowers can do about it.
Why Caribbean credit scoring misses informal earners
Scorecards were built for economies with mostly formal jobs, full credit bureau coverage and steady payslips. The Caribbean has large informal economies: estimates for Jamaica put the informal sector at 30 to 45 percent of GDP.
Market vendors, domestic workers, contractors, farm labourers and cook-shop owners manage cash flow, stock and risk every day. That work produces almost none of the documents a scorecard asks for. Without a formal history they cannot get credit, and without credit they cannot build a formal history. In many families that loop has held for generations.
Conventional credit scoring does not find the Caribbean untrustworthy. It finds the Caribbean invisible.
Cost is the second problem. Manual assessment takes a loan officer's time, and that time costs the same whether the loan is JMD 50,000 or JMD 5 million. The small loans that would help micro-entrepreneurs most are the ones lenders can least afford to process by hand.
What alternative-data models look at
An AI credit model can be trained on the records people already create through ordinary economic activity. Useful inputs include:
- mobile wallet inflows and outflows, and how regular they are
- utility and phone bill payment history
- supplier payments and stock purchases
- savings transfers, even small ones
- how long the same pattern has held
Lenders in sub-Saharan Africa, Southeast Asia and Latin America have run models of this kind at scale. Published results report repayment prediction close to conventional scorecards, in populations where a scorecard has almost nothing to work with.
Mobile wallets matter most here. A vendor who takes payment by wallet every day, moves a steady share into savings and pays suppliers on schedule is showing financial discipline in a form a model can read. Ten years ago that record did not exist in digital form.
Why a model trained on Caribbean data does better
StarApple AI has worked on credit risk models for Caribbean lenders since the company's early years. The models are trained on Caribbean repayment data and calibrated against Caribbean default patterns.
An imported model learned its idea of default from another economy, with different loan products, pay cycles and regulations. Applied in Jamaica, it looks for patterns that may not exist here and misses local ones, such as seasonal income around tourism and harvest, or remittance-driven cash flow.
In our deployments with Caribbean financial institutions, locally trained models performed better on Caribbean applicants than offshore models adapted for the region. The gain came from local repayment data, local variables and a team that knows the economy the model is scoring. It showed up as approvals that a conventional scorecard would have declined, including for applicants much like the market vendor.
Fairness and appeal rules a regulator should ask for
Regulators' concerns about AI credit scoring are legitimate, and two stand out.
The first is discrimination. A model trained on historically biased lending data can reproduce that bias without ever using a protected characteristic, through proxies such as address or phone type. This has happened in credit scoring elsewhere. A lender should be able to show approval and default rates broken down by group, and how often it reruns that check.
The second is accountability. When a model declines someone wrongly, the applicant needs a named person to contact, a reason they can understand and a way to have the decision reviewed.
Manual assessment has the same risks. It is slower, costs more, depends on each officer's judgement and already excludes informal workers. The useful regulatory question is what documentation, fairness testing and appeal process a lender must have before it deploys a credit model.
What still limits reach: data, explainability and price
Data coverage. A model can only score people who leave a digital trail. People who deal only in cash are still outside it, so wider wallet adoption and cheaper digital payments come before AI inclusion can reach the most excluded.
Explainability. A loan officer who cannot explain a model's decision cannot defend it to a customer or a supervisor, and will not trust it. Providers need to supply plain reason codes with every decision, such as "irregular inflows in the last three months".
Pricing. A model that lets a lender serve new borrowers profitably does not guarantee a fair rate. If the new loans carry rates that wipe out the benefit of access, inclusion has failed on the borrower's side even while the lender's numbers look good.
How to measure whether it is working
Model accuracy and loan volume are lender metrics. The borrower-side test is what happens after the loan. Track, for each cohort of first-time borrowers:
- repayment rate at 6 and 12 months
- share who take a second, larger loan
- change in business revenue or stock where it can be recorded
- share who now have a credit bureau file they did not have before
The vendor who borrows to double her stock and repays on time, and the contractor who finances equipment, takes a bigger job and hires two people, are the outcomes these numbers should capture. Individual cases like these exist today. The numbers are still small next to the excluded population, and I do not yet have regional data showing how many first-time borrowers move on to larger formal credit.
What to do next
- Lenders: pick one product, such as a working capital loan under JMD 250,000, and run an alternative-data model alongside your scorecard for one quarter. Compare approvals and 90-day arrears before switching anything over.
- Lenders: write the appeal process before launch. Name the officer who reviews declined AI decisions and set a response time, for example ten working days.
- Regulators: ask every lender using a credit model for three documents: the data sources, a fairness test by group, and the appeal procedure. Ask for the fairness test again each year.
- Fintech and wallet providers: give customers a way to download or share their own transaction history with a lender. That single feature turns wallet data into a credit file the customer controls.
- Borrowers without a credit file: run your sales and supplier payments through one wallet or bank account for at least six months before you apply. A steady record is what these models score.
Frequently asked questions
Who regulates AI credit scoring in Jamaica?
No single law covers AI credit models yet, so several existing rules apply. The Bank of Jamaica supervises deposit-taking institutions, licensed credit bureaus and, under the Microcredit Act 2021, licensed microcredit lenders. The Data Protection Act 2020, in force since December 2023, governs how lenders collect and use personal data, and the Office of the Information Commissioner oversees it.
Can a lender use my mobile wallet or phone data to decide on a loan?
Only with a lawful basis, which in consumer lending usually means your informed consent. Under Jamaica's Data Protection Act you can ask any organisation what personal data it holds on you and how it uses it. If a lender will not tell you which data fed its decision, treat that as a warning sign.
How do I check my credit file in Jamaica?
Contact the licensed credit bureaus directly; the Bank of Jamaica publishes the current list on its website. The Credit Reporting Act gives you the right to see your report and to dispute errors in it. Many first-time borrowers find they have no file at all, which confirms the thin-file problem this article describes.
How is an alternative-data model different from a microlender's quick approval?
A quick approval can rest on little more than an ID and a phone number, and it is priced high to cover the unknown risk. An alternative-data model builds a risk estimate from months of real payment behaviour, which should support a lower rate for reliable borrowers. Ask any lender what data it used and whether a good repayment record will lower your rate on the next loan.
What does it cost a small lender or credit union to start?
The main cost is data preparation, not software. A lender needs at least a year or two of clean loan and repayment records to test a model against, and staff time to label defaults consistently. Running the model in parallel with the existing scorecard for one quarter, as a pilot, keeps the risk low while the numbers come in.