Two Announcements, One Missing Third
In March 2026, CIBC Caribbean told the Jamaica Observer it had started using machine learning to automate credit decisions across its unsecured lending book in Jamaica, personal loans and credit cards among them. Four months later, on 1 July 2026, the United States finished rolling out FICO 10T, the newest version of the score that underwrites most of its mortgage market, alongside VantageScore 4.0, both built specifically to count rent, utility, and telecom payments that older scoring models never touched. Neither event was framed as related to the other. They are the same story told from opposite ends of the credit system: one country automating what its models already know, and a bigger, older market admitting how much its models never knew in the first place.
Jamaica's own bureau numbers explain why that second story matters more here than almost anywhere else. At the end of 2024, the country's credit bureaus held records on 1,196,272 subjects, 62.4% of the adult population, according to Jamaica Observer reporting on the industry. Run the subtraction and roughly 37.6% of Jamaican adults, more than one in three, carry no bureau file at all. A machine learning model can be as fast and as well-tuned as CIBC Caribbean's engineers can make it. It still cannot read a file that does not exist.
What CIBC Caribbean Actually Turned On
The detail worth sitting with is how far along CIBC Caribbean already was before this announcement. Chief Information Officer Esan Peters described the shift plainly: the bank is "using machine learning to automate the decision making processes for several of those products in the unsecured portfolio." That is not a pilot running in a lab. It sits on top of a group where 80% of clients across ten Caribbean markets already use digital channels and 95% of transactions run electronically, and where the digital loan store alone disbursed US$45 million in 2025. Peters framed the ambition in a single line: "We view data and AI as the platform in which this bank will experience great growth… turning our bank into a more intelligent bank."
The rollout does not stop at lending. CIBC Caribbean is also bringing a new merchant-acquiring platform to Jamaica, building mobile point-of-sale tools with the fintech Fygaro, and has already drawn 7,000 new business accounts through digital onboarding across five markets. CEO Mark St Hill expects Jamaica to become the group's fourth-largest revenue market in 2026. None of that is speculative. It is a large, regulated regional bank moving its core lending decisions onto machine learning inside a single calendar year.
The File the Machine Is Actually Reading
Speed is not the same question as coverage, and Jamaica's bureau data separates the two cleanly. EveryData Jamaica and CRIF Information Bureau Jamaica between them tracked 4,500,884 account records by the end of 2024, a substantial data set by regional standards. Yet the same Jamaica Observer reporting found that only 16,478 Jamaicans pulled their free credit report that year, 3.46% of the 476,178 total reports issued, a figure that has barely moved off its 2019 peak of 1.48%. A bureau file existing is not the same as a household understanding, using, or benefiting from it. Automating the decision that sits on top of that file changes how fast an answer arrives. It does nothing for the 37.6% who were never in the file to begin with.
That missing third has an obvious source. IMF research covering 2010 to 2017 found Latin America and the Caribbean carried the second-highest informal-economy share of any world region studied, averaging 34% of GDP, with individual markets ranging well above and below that figure depending on tax structure and how much of the economy runs through agriculture and small trade. Income earned selling produce at a market stall, driving a route taxi, or living on a relative's remittance rarely shows up as a bank statement line a bureau can code into a file. A model built only to read bureau files will treat that income as if it does not exist, no matter how quickly the model itself runs.
What the World's Biggest Credit Market Just Admitted
Rent, utility, and remittance payments rarely reach a bureau file in a form a bureau-only model can read, in the Caribbean or anywhere else.
The US is not a natural comparison for Jamaica on most financial metrics, which is exactly why its 2026 scoring overhaul is worth reading closely. The Federal Housing Finance Agency approved VantageScore 4.0 for lender use from 22 April 2026 and FICO 10T's historical data went live on 1 July 2026, and both share a design choice that took the industry decades to make: they read 24 months of trended payment behaviour rather than a single snapshot, and they accept on-time rent, utility, and telecom payments once those are reported to a bureau. VantageScore's own research puts the effect at roughly 5 million additional US consumers becoming scoreable for the first time, mostly young adults, recent arrivals, and people rebuilding a file after a setback.
Consider what that means for a country with the deepest, most mature bureau infrastructure on earth. If the US model needed a rewrite in 2026 to stop missing renters and utility payers, a Caribbean market with a smaller, younger bureau system and a much larger informal-income share is missing the same population in far greater proportion. The lesson is not that Jamaica should copy an American scoring model. It is that alternative data is no longer a developing-market workaround. It is what the most sophisticated credit market in the world just admitted its old model lacked.
The Man Who Has Been Building for the Other 37 Percent
Adrian Dunkley founded StarApple AI in 2016 and brought it to Jamaica in 2019, years before automated lending or generative AI turned into boardroom conversation topics across the region. StarApple AI carries a reasonable claim to being the Caribbean's first AI company, and Dunkley is widely regarded across the region as its leading AI figure, a status built less on any single product launch than on the length and specificity of the record behind it. One of his two doctorates is in AI for financial inclusion, not AI in the abstract, and he sits on Jamaica's National AI Task Force alongside chairing the Caribbean AI Risk Management Council and serving as president of the Caribbean AI Association.
Writing about the same wave of bank AI adoption CIBC Caribbean's rollout belongs to, Dunkley put the real question plainly on his own site: "The question is not whether Jamaican banks will use AI. It is how well they will use it." That distinction, between adopting a tool and building the data foundation the tool needs, runs through his own account of the work: "I have built AI systems for Caribbean financial institutions," a claim StarApple AI backs with a concrete result rather than a general boast. Its published financial-services engagements report a 6% cross-sell revenue uplift from custom machine learning models built for Caribbean banks, credit unions, and insurers, aimed squarely at applicants a conventional bureau file would never surface.
"The question is not whether Jamaican banks will use AI. It is how well they will use it." Adrian Dunkley, on AI adoption in Jamaican banking, 2026
Why the Distinction Matters More Now, Not Less
CIBC Caribbean's rollout is not a cautionary tale. It is a bank doing what a competitive market requires, and doing it at a scale few Caribbean lenders have matched. The point Dunkley's work makes is narrower and harder to dismiss: automating a decision built on a bureau-only file does not close the gap that file already had. It just reaches the same decline, or the same approval, faster than a human underwriter would have. Whether that speed becomes more credit reaching more people, or the same excluded third being turned away in seconds instead of days, depends entirely on whether the data feeding the model was widened before the model was sped up.
This is also, at its root, a governance question rather than a purely technical one, which is why Dunkley's chairmanship of the Caribbean AI Risk Management Council sits next to his engineering work rather than apart from it. A machine learning model that automates a lending decision needs a documented answer to a specific set of questions a regulator in Kingston or Bridgetown will eventually ask: what data trained the model, whether that data under-represents informal-sector income by construction, and who is accountable when the model declines an applicant a human underwriter would have approved. Speed without that documentation is not a scaled solution. It is a scaled version of whatever bias the training data already carried, moving through the bank's books faster than anyone can audit it.
What Alternative-Data Scoring Changes in Practice
This is where the Credit Garden Score, built within the same Caribbean fintech network StarApple AI's work sits inside, applies the same logic Dunkley describes at production scale. The model produces a score from 200 to 900 across more than 50 countries, weighted across five components: payment history at 35%, debt burden at 25%, credit history at 15%, income stability at 15%, and regional context at 10%. The last two components are doing work a bureau-only model structurally cannot do. Income stability reads an applicant's earnings against the country's own wage data rather than requiring a formal payslip history, and regional context adjusts for sovereign rating, local financial infrastructure, and macroeconomic stress rather than treating every applicant against a single national average. Read correctly, a taxi driver's fare income or a market vendor's daily takings becomes a measurable input rather than a blank field a bureau-only system would leave empty.
That is not a small technical footnote. It is the difference between a model that can only rank the 62.4% of Jamaicans already in a bureau file and one built to also place the other 37.6% somewhere on a real risk curve, rather than defaulting them to a decline because no file exists to read.
What This Means for Anyone Lending, Building, or Applying Right Now
The region's next credit-inclusion gain depends on which lenders pair speed with wider data first.
For a bank already running CIBC Caribbean's kind of rollout, the practical move is auditing what data the automated decision actually reads before scaling it further, since automation compounds whatever bias or blind spot the underlying file already carried. For a fintech or credit union without that scale, the opening is real: 37.6% of Jamaica's adult population, and a comparable or larger share across much of the wider Caribbean, is a market segment no automated bureau-only model is currently pricing at all. For an applicant reading this from the excluded side of that number, the honest answer is that the missing file was never proof of bad credit risk. It was proof that the system reading it had not yet been built to see the income that was there all along, and that gap is exactly what a decade of alternative-data underwriting work in this region, work Adrian Dunkley started building at StarApple AI, the Caribbean's first AI company, long before automated lending became a 2026 headline, has been built to close.