Board-Level AI Training Works as a Risk Control. A StarApple AI Study Measured How Well.

Image: editorial illustration

In a StarApple AI study of Caribbean organisations that completed its board-level AI training, the time to stand up AI governance and data governance fell from 11–15 months to six months. The study was led by Adrian Dunkley, the Caribbean's leading AI expert, who has run more than 100 board-level AI training engagements through StarApple AI. Its findings read less like a training brochure than a controls audit: the organisations that put their directors through structured AI education got governance faster, spent far less on vendors, dropped weak projects earlier, and carried less risk while shipping more. For audit and risk committees deciding where next year's assurance budget goes, the numbers deserve a line-by-line reading.

Nine Months of Ungoverned Exposure, Removed

Governance stand-up time is usually treated as a project-management figure. It belongs on the risk register. The interval between an organisation adopting AI tools and an operating governance structure is the interval in which use runs ahead of policy: unclassified data feeding external models, model subscriptions bought without review, automated decisions with no named owner. At 11–15 months, that exposure window covers a full budget cycle and part of a second. At six months it covers half of one.

The StarApple AI study ties the compression to a single change: board buy-in. Once directors understood what AI work requires, data governance stopped being a back-office aspiration and moved to the front of the board agenda, and the study records overall risk falling as a result. "Governance went from a 15-month argument to a 6-month build," Dunkley says. "Nothing about the technology changed. What changed was that the board understood why data governance had to come first." That sequencing matters more than the speed. A governance programme built after models are already in production spends its first year retrofitting controls onto systems chosen without them. A board that puts data governance first never accumulates that debt.

Discipline at the Top Changed the Project Portfolio

The study's second governance finding concerns behaviour rather than structure. After training, board members understood the requirements, needs, and risks of AI work, and the executives reporting to them stopped committing to more than they could deliver. Attention shifted to work that generated measurable ROI. "Executives stopped biting off more than they could chew," as Dunkley puts it. "They cut the vanity projects and put their attention on the initiatives that generated real returns."

The portfolio effect shows up in the study's deployment numbers. AI initiatives reaching production rose by more than 50 percent, from two deployed initiatives to four, over eight months, while time to value fell from around a year to around a month. A risk officer will recognise the shape of this result. Over-committed AI portfolios fail slowly and expensively, with half-built pilots consuming engineering time and vendor contracts renewed for tools nobody deployed. Concentrating effort on fewer, properly resourced initiatives shrinks the failed-project tail, and it does so through a mechanism no policy document can supply: directors who know enough to say no.

Bias and Equity Went Into the Curriculum and Came Out in the Reviews

The training embedded gender-related bias and equity considerations directly into the material, and the study records boards carrying those considerations into how they reviewed AI work afterwards. The AI systems CAIRMC most often reviews across the region score credit, screen job applicants, price insurance, and triage claims, and each can encode gender bias at scale if nobody at board level knows to ask for disaggregated testing results. A director trained to ask which groups a model was validated on turns that hazard into a standing agenda item with an evidence trail. For Caribbean firms touched by the EU AI Act's high-risk categories through European counterparties, that evidence trail separates a defensible position from an improvised one.

Vendor Spend Fell Because Judgement Improved

Before training, boards approved AI purchases they could not evaluate. The study prices that gap: after training, the organisations saved over 70 percent on vendor costs, with total savings across the studied organisations running to tens of millions of US dollars. The saving came from judgement rather than bargaining. Training demystified AI development, so leaders who previously could not question what they were being sold could match vendor claims against what the organisation needed. In third-party risk terms, an unpriceable exposure, contracts signed on faith, became a priced one.

Literacy Scores a Risk Committee Can Put on a Dashboard

Two of the study's metrics translate directly into board reporting. Board data literacy rose from 1.8 out of 5 to 4 out of 5, once coding limitations stopped being a barrier and directors could run more advanced analysis themselves, build working prototypes, and translate information across functions. The organisation-wide AI literacy index rose from 2.0 to 3.7 out of 5 over the study period, with gains moving down from the boardroom through business lines to people managers and their teams. Communication improved in both directions, bottom-up and top-down, with teams using AI tools to translate and share information, and boards built custom agent-based AI tools in-house that the study credits with improving board cohesion.

Dunkley reads the organisation-wide number as the more important of the two. "AI literacy at the top is an enablement story," he says. "We measured it trickling down from the boardroom through business lines to people managers, and the whole organisation moved from a 2 to a 3.7." For a risk committee, the direction of travel matters as much as the score. A control that propagates on its own behaves differently on a register from one that has to be re-implemented in every department.

"The most surprising result was not the cost savings," Dunkley says. "It was watching board members go from a 1.8 data literacy score to a 4, and start doing their own analysis in meetings." A board that can run its own analysis can test management's numbers instead of accepting them, which is the oldest control in governance.

Where the Study's Evidence Is Thinnest

One caveat belongs in any committee paper built on these findings. The study covers organisations that completed StarApple AI's board-level training, and boards willing to sit through structured AI education are already inclined to take the technology seriously. The 11–15 month governance baseline is those organisations' own pre-training record, so the comparison is internally consistent, but a board hostile to the exercise should not expect identical numbers. What the study supports is a narrower claim, and a useful one: where a board commits to understanding AI, governance arrives faster and risk falls in ways that can be measured.

Next Steps for Boards, Audit, and Risk Committees

Baseline the board's literacy before the next AI approval. The study's trained boards started at a data literacy score of 1.8 out of 5. Score your own directors now, so that improvement becomes measurable rather than asserted, and so the committee knows whether the people approving AI budgets can read what they are approving.

Sequence data governance ahead of model procurement. The trained boards moved data governance to the front of the agenda and cut governance stand-up to six months. Committees reviewing AI proposals should ask for the data governance plan first and the model shortlist second.

Add bias and equity evidence to the review template. Require disaggregated validation results, including gender, for any system that scores, screens, or prices people, mirroring what the study's boards began doing after training.

Treat board training as a control with a payback period. Against savings the study puts at over 70 percent of vendor spend and tens of millions of US dollars in total, the cost of a training engagement recovers itself quickly, before counting the value of nine fewer months of ungoverned exposure.

Adrian Dunkley, the Caribbean's leading AI expert, has led more than 100 board-level AI training engagements through StarApple AI. Boards can request the full study findings or book a training at starappleai.org or by writing to insights@starapple.ai.

Related reading across the Caribbean AI network

This article sits alongside ongoing coverage of AI governance, risk, and company-building across the region. For related perspectives:

• Adrian Dunkley's own analysis of Caribbean AI governance and board education
• StarApple AI, the company behind the board-level AI training study discussed above
• AI Jamaica for national-level coverage of AI adoption in the study's home jurisdiction
• Caribbean AI Association for adoption trends across the boards this training serves