Board-Level AI Training Works. I Watched It Happen in More Than 100 Caribbean Boardrooms

William Hogarth, The Committee (Twelve Large Illustrations for Samuel Butler's Hudibras, Plate 10), 1725–68. The Metropolitan Museum of Art, public domain.

I have led more than 100 board-level AI training engagements across the Caribbean through StarApple AI, the company I founded in Kingston. Commercial banks, insurers, utilities, conglomerates, public bodies, family firms. This year we did something the training business rarely does: we went back to the organisations that completed the programme and measured what had changed. As far as I know, it is the first attempt in this region to put numbers on what happens after a board learns how AI works. The finding that surprised me most was about data literacy, not money.

The opening session in a new boardroom follows a pattern I could set my watch by. Around the second slide, a director leans in with the question that has been waiting since the engagement letter was signed: if our people use these tools, where does our data go? At a financial services board in Kingston the worry was the loan book. At a utility, outage records. It is the right question, and a defensive one. A board that asks only defensive questions tends to treat AI as something that happens to the organisation, and leaves the direction to management and vendors.

By month three, directors were questioning pilots and invoices

By the third month of a typical engagement the questions had changed. Why is this initiative still in pilot? What data would we need to answer that ourselves? Which line on this vendor invoice pays for something our own team could build in a week? At one engagement, a director pulled up a live vendor proposal mid-session and worked through it clause by clause, pricing each deliverable against what an internal build would cost. That same board had approved the vendor's previous contract without a single technical question being asked.

We build gender-related bias and equity review into the training itself, and it changes how boards then review AI work. After training, directors in the study asked who a credit model fails for, and whether the data behind a screening tool carries the bias of the years in which it was collected, before they asked what the tool cost. Those questions used to arrive from compliance, late, if they arrived at all.

Organisation-wide AI literacy rose from 2.0 to 3.7

The study's broadest number is organisation-wide. Across the organisations that completed board-level training, the internal AI literacy index rose from 2.0 out of 5 to 3.7. Once boards understood the technology, permission and budget moved down through business lines to managers and their teams, and we measured the effect level by level.

Delivery followed. The number of AI initiatives that left pilot stage and reached deployment doubled, from two to four, over eight months. And the slowest item on any AI agenda, standing up AI governance and data governance, fell from 11–15 months to 6. The technology did not change in that window; board support did. Once directors understood why data governance had to come first, it moved to the front of the agenda, and everything built afterwards sat on firmer ground.

Discipline improved alongside understanding. After training, board members grasped what AI work requires and what it risks, and their executives and managers stopped taking on more than they could deliver. They cut pilots that existed only because a competitor had announced something similar, and funded initiatives with a measurable return. Time to value fell from around a year to around a month by the study's measurement, and I put most of that down to boards finally knowing enough to say no.

Board data literacy rose from 1.8 to 4 once coding stopped being a barrier

I expected the governance numbers. The data literacy result surprised me. Board data literacy rose from 1.8 out of 5 to 4, the largest movement we recorded, and the cause was ordinary: directors no longer needed to write code to analyse data. Directors who had always left analysis to technical staff could run it themselves with AI tools, and build a working prototype by describing it in plain language, which people now call vibe-coding. They also stopped needing an interpreter to move information between departments.

The session where I stopped doubting that number came late in the study period. A director who had introduced herself in week one as the least technical person in the room opened her laptop and walked the board through a working prototype she had built to interrogate the company's customer complaints data. It was rough and it was hers. The discussion that followed was the best data conversation I have watched a board have.

The same shift produced tooling. Boards in the study built custom AI tools in-house, on an agents-based approach, to prepare and share the material their meetings run on, and board cohesion improved with them. Communication improved in both directions across the studied organisations, bottom-up and top-down, with teams using AI tools to translate and share information between levels that had previously talked past each other.

Vendor spend fell by more than 70 percent

Organisations in the study cut vendor costs by more than 70 percent after training, and the savings ran to tens of millions of US dollars in total. Boards had been paying for AI they did not need because they could not question what they were sold. Training demystified the development process, and once a director understands roughly what a model, a data pipeline, an integration, and a maintenance contract each involve, a seven-figure proposal for a chatbot receives the scrutiny a seven-figure proposal deserves. In most cases, what the organisation needed cost a fraction of what it had been buying.

What training did not fix

The study has limits, and so does the training. Selection is the first. Boards that commission AI training are already motivated, so the study cannot cleanly separate what the training caused from what those boards would have managed anyway. I think the governance and vendor findings survive that objection, because the mechanism behind each was visible in the room. Still, we studied trained boards, not a controlled trial.

Second, training moved none of the underlying data. An organisation that arrived with years of accumulated data quality problems still had them at month eight. The board now knew about them, and funded the repair instead of building on top of them. Training also did nothing for the talent pipeline. Trained boards hire better and retain better, but a literacy score does not close a salary gap with Toronto or London, and some of the engineers behind the deployments we counted will be recruited away regardless of how well their board now understands what they do.

The organisations in the study did not buy new infrastructure or set up a research lab. They taught the people at the top how the technology works, and most of the changes we measured came after that.

What to do next

1. Board chair: book one half-day session this quarter in which directors use AI tools on a real, anonymised company dataset, not slides about AI.
2. Company secretary or CFO: list every AI-related line in current vendor contracts, with annual cost and what it delivers, and table the list at the next board meeting.
3. Audit or risk committee: set a six-month deadline for an approved data governance policy, the benchmark trained boards in the study met, and ask for monthly progress.
4. Nominations committee: add AI and data literacy to the board skills matrix and score each director on it before the next board evaluation.
5. Every director: before approving the next AI proposal, ask who the model fails for and what data it was trained on, and minute the answer.