The Fastest Way to Take Over the Caribbean with AI

My concern is that Caribbean countries could spend years developing national AI strategies while surrendering practical control through the tools they adopt. Each country may have its own policy, yet depend on systems whose design and commercial priorities are settled elsewhere.

The easiest opening for outside control would be a region that negotiates separately, tests separately and keeps its failures to itself. We could find ourselves with less influence over our own development while decisions affecting our citizens increasingly reflect interests outside the region.

For a business owner, the question becomes immediate: who can change the system you rely on, what can it do with your information, and can you keep serving customers if access disappears?

A national policy needs enforceable supplier terms

A strategy can describe our ambitions without changing the terms on which we buy technology. If we cannot inspect a consequential decision, move our records or continue an essential service without the supplier, our freedom to choose has narrowed.

Consider a hypothetical Jamaican insurer connecting an AI assistant to claims documents. Preparing summaries may help staff give customers faster answers. But if the firm gradually loses the ability to review the underlying evidence without that assistant, it has created a dependency alongside the efficiency gain. A contract renewal then affects how it can operate, not simply what software it purchases.

I describe the wider concern through Agent Colonialism: AI introduced into communities without their interests adequately represented in its design and decisions. The danger is that our data and experience create value elsewhere while our people carry the consequences of poor decisions. We should examine that possibility in procurement, before a system becomes difficult to replace.

Foreign technology can be useful. Local ownership also needs scrutiny. The relevant tests are whether people affected by the system have protection and whether the organisation retains enough knowledge and authority to challenge it.

An agent needs a clear limit on whose instructions it follows

An AI agent can use connected tools to carry out work, such as handling files or preparing messages. The authority we give it matters because a wrong instruction can reach beyond the answer on a screen.

There are different routes to harm. A supplier's commercial objectives may conflict with a customer's interests. An agent may pursue a badly specified goal, such as closing cases quickly at the expense of resolving them properly. A hostile actor may deliberately redirect an agent. State influence is another risk to assess where evidence supports it; a supplier's country of origin alone does not establish that its product is acting for a government.

NIST's agent-hijacking experiments demonstrate a concrete security problem. Malicious instructions placed in material an agent reads can redirect its behaviour. Its tests included scenarios involving data theft and automated phishing. These were controlled evaluations, not evidence that Caribbean institutions have experienced those incidents.

For the insurer in our example, a document supplied for review must never acquire the authority to approve a claim or release customer files. Enforce that limit in the connected systems. A prompt asking the assistant to be careful cannot carry the whole responsibility.

Regional coordination must reach the contracts and the tests

The Caribbean has already begun this discussion. In July 2026, UNESCO reported COTED-ICT endorsement of its Caribbean AI Policy Roadmap. The CTU's December 2025 interim report called for harmonised governance, shared oversight and procurement standards. These initiatives give the region a basis for practical collaboration.

My recommendation is to connect that work to the decisions organisations make when they buy and deploy AI. A shared procurement checklist could help a hospital in Barbados and a public agency in Jamaica ask comparable questions about access, subcontractors and recovery. Each institution would still decide whether a system suits its task.

Governments and companies should also share what fails. A confidential regional reporting channel could circulate a verified warning without exposing customer records or helping an attacker. Universities and technical teams could maintain common tests that participating organisations adapt to their own services.

Coordination carries risks too. A regional platform could concentrate sensitive information or create another dependency. Shared testing does not require a central store of citizens' records. We should share methods and relevant findings with clear access controls, and test the regional arrangements themselves.

What businesses and governments should do next

A board or permanent secretary can start with an inventory of the AI systems that influence consequential work. Record the supplier, the information each system accesses, its permitted actions and the person authorised to stop it. Include tools staff have connected to work accounts.

Procurement teams should request clear terms for data use, changes to the service, access to audit records and the return or deletion of information. Ask the supplier to demonstrate how an organisation can leave. The relevant legal and technical teams should assess whether those promises are enforceable and workable in their jurisdiction.

For an initial test, select one workflow within the next 30 days. Use synthetic records to attempt a prohibited action, then disable the agent and complete the task through the fallback process. Measure the time required and record every missing permission, document or skill. Keep consequential actions restricted until the process owner resolves the failures. This is a proposed exercise, not a published measure of safety.

At regional level, I recommend a joint working session between government procurement leads, regulators, industry and technical researchers. Give it specific outputs: common supplier questions, a test-sharing process and a protected incident-reporting route. Assign an owner and a delivery date to each output.

Invest in the people who will challenge these systems. Caribbean institutions need staff who understand their own processes well enough to recognise when an impressive answer serves the wrong objective. That expertise should remain inside the organisation even when it buys external technology.

Keep the ability to refuse

We cannot yet quantify how much regional coordination would reduce these risks, or assume that agreement between governments will translate into effective enforcement. That is where this argument needs practical testing. Begin with shared procurement questions and controlled exercises, then publish what participating organisations learn without exposing sensitive material.

My position is that countries should retain their national priorities while building the protections they can enforce together. Before approving the next AI deployment, ask the team to demonstrate who controls its actions and how the service will continue without it. Put that demonstration on the board agenda.