Sovereign AI Has Become an Unverified Claim. CARICOM Should Make Vendors Disclose

Sovereign is currently the most valuable word in Caribbean AI sales. It opens doors in ministries, it shortens procurement conversations, and it costs a vendor nothing to say. In the procurements CAIA members have brought to us over the past year, and in the ones I have been asked to review personally, the word has usually meant a locally built interface over a foreign foundation model, with every prompt travelling to a data centre in North America or Europe.

Renting a frontier model through a capable regional integrator is frequently the right purchase for a Caribbean government, and the integrator earns its fee in workflow, integration and support. What goes wrong sits above the architecture, in the fifteen governments approving these systems against assumptions the word creates, recorded in Cabinet notes and procurement files that will not survive being asked for evidence.

Three Categories That Get Sold as One

Sovereignty over a capability means you can determine, on your own terms, whether it keeps running and who observes what passes through it. Applied to AI procurement, that resolves into three conditions that must hold together. The contracting party owns the model or licenses weights on terms it controls. Inference executes on infrastructure the vendor or the buyer operates. And the location of that infrastructure is written into the agreement, because a regulator or an auditor general asking where citizen data is processed is asking for a document.

Against that definition, three distinct products are currently sold under one heading. Genuine sovereign AI, where a vendor runs weights it controls on infrastructure it or the client operates. Self-hosted open-weight deployment, where the buying institution downloads a model from a family such as Llama, Mistral, Qwen, DeepSeek or Gemma and runs it on its own hardware. And a foreign foundation model behind a local interface, which is the most common and the least often named. Fine-tuning a hosted foreign model on Caribbean data belongs in the third category rather than the first, because the training corpus has now crossed the same border and the resulting weights sit with the provider.

The middle category is the one this region systematically under-buys, and the reason is commercial rather than technical. Open-weight models are downloadable, run on modest hardware once quantised, and settle the residency question by inspection, because the machine is somewhere a person can be walked to. Nobody has a sales team for that, because there is no recurring licence inside it.

Why Fifteen Governments Asking Separately Does Not Work

A procurement officer in a ministry of under a million people who asks a vendor to name the foundation model behind its product is doing something unusual, and both sides know it. The request reads as suspicion rather than process. The officer has no comparable answer from another government to benchmark against, no standard format for the reply, and limited leverage if the vendor declines. Most officers, entirely reasonably, do not ask twice.

A common schedule changes the social dynamics of that conversation more than it changes the technical content. When every CARICOM tender carries the same six fields, asking stops being an accusation and becomes paperwork. When the answers come back in the same format across member states, a procurement office in Castries can compare what a vendor told Kingston against what the same vendor is telling them. That comparison is the actual leverage, and no single government can generate it alone.

This is the same argument the Caribbean AI Task Force made about infrastructure when it proposed a Caribbean Data Commons and shared regional computing facilities under its data sovereignty pillar, and the same argument CAIA has been making since 2024 about the pattern of small island states each negotiating separately with the same handful of foreign providers. The infrastructure version of that argument now has institutional backing. The procurement version does not yet, and it is considerably cheaper to implement, because it requires no capital at all.

The Six Fields

CAIA proposes these as a common schedule for AI tenders across CARICOM member states, scored before any demonstration is seen. A demonstration is engineered to persuade. A written answer is engineered to be checked, and to be produced again eighteen months later when somebody asks what was agreed.

• Model identification. The model or model family behind the service, and the licence under which the weights are used.
• Inference location. The operator and the physical region where inference executes, with confirmation that both will appear in the contract rather than in an appendix the vendor may revise.
• Sub-processors. The complete current list of third parties processing buyer data, and the notice period before it changes.
• Retention and reuse. Whether prompts, outputs and uploaded documents are retained by any third party, for how long, and whether they may be used for model training or human review.
• Availability record. Twelve months of incident and outage history, with causes, which a buyer can lay against the public incident histories the major providers publish at status.openai.com, status.anthropic.com and status.cloud.google.com.
• Verification undertaking. Confirmation that the bidder will run a demonstration with outbound access to foreign model providers blocked, during evaluation, on the buyer's infrastructure or under the buyer's observation.

The sixth field carries most of the weight. A model executing locally keeps answering with the outside world cut off. A rebranded product returns an error or falls back to a canned response, and no architecture diagram alters that outcome. It costs a firewall rule and an afternoon, it needs nobody in the procurement office to read code, and in practice it is the step most frequently agreed in principle and then quietly dropped from the evaluation timetable under schedule pressure.

What a Wrong Answer Costs a Government

Four exposures follow, and they arrive on different timescales. A residency position that cannot be evidenced when an auditor general or a data protection commissioner asks for a data flow diagram. Continuity that sits outside the contract entirely, because a rebranded product inherits every upstream decision, from a deprecated model version to a regional access restriction, and the vendor's service agreement will accurately describe the resulting outage as beyond its control. Costs that cannot be negotiated, because a vendor with no cushion upstream has nothing to argue with when the price moves. And a public record that says something the ministry cannot support.

The regional data protection statutes make the first of those concrete rather than theoretical. Jamaica's Data Protection Act 2020 and Barbados' Data Protection Act 2019 both treat transfer of personal data outside the jurisdiction as a regulated act requiring a lawful basis and documentation. Trinidad and Tobago's Data Protection Act 2011 remains largely unproclaimed, which produces the opposite problem: thinner statutory recourse, and therefore heavier reliance on contract terms that most tenders never ask for. Either way, a prompt containing a citizen's personal data becomes a cross-border transfer the instant it is sent, and an institution that cannot name the receiving jurisdiction cannot write the record.

The Part of This That Is Not Commercial

There is a second reason to give the term a definition procurement can enforce, and it has nothing to do with vendor competition. Investment fraud across this region has learned the same vocabulary. Securities regulators, including the Trinidad and Tobago Securities and Exchange Commission, have issued repeated public alerts about schemes using AI-generated video of recognisable public figures to sell fraudulent products, and proprietary or sovereign AI is a standard element of those pitches.

When a phrase carries no testable meaning in legitimate commerce, it becomes free credibility in fraudulent commerce. A CARICOM disclosure standard that establishes what the word requires does more than protect procurement budgets. It gives regulators and the press a definition to hold a claim against, which is the thing currently missing when a scheme describes itself in exactly the language a ministry used in a press release last month.

Where This Proposal Is Weak

CAIA is not a standards body and has no authority to require anything of any government. This is a proposal from an association whose membership includes both buyers and vendors of these systems, and members on the vendor side of that list will have views about it. I would rather publish it with that tension visible than pretend the association speaks with one commercial voice on the question.

The proposal also has a real technical weakness. Publishing a verification standard degrades it, because a vendor reading this page can prepare for five of the six fields and rehearse an answer that satisfies a non-technical reviewer. What survives preparation is what costs money to fake: the egress-blocked demonstration, and a sub-processor schedule the vendor's own lawyers have signed, where misrepresentation is a contractual breach rather than a marketing exaggeration. If a member state adopts only two of the six fields, those are the two.

The deeper limitation is that a disclosure schedule governs the systems a government procures deliberately, and the material exposure in most public bodies is accumulation. A ministry with a properly assessed core system, an unlogged transcription tool that joins meetings, a browser extension installed by staff, and a chatbot from a fourth supplier has one good contract and an unmapped surface. In the institutions I have worked with across the region this year, a complete AI inventory has generally not existed before we sat down and built one, and its eventual length has been the first finding rather than the last. A disclosure standard does not reach that problem. An inventory requirement does, and it should probably come first.

What CAIA Is Asking For

Three things, in order of how easily they could actually happen. That the six fields be adopted as a standard schedule in AI tenders by any member state willing to move first, since a single government can do this without waiting for regional agreement and would immediately give the others a template. That the Caribbean Telecommunications Union consider the schedule as an implementation artefact under the Caribbean AI Task Force's data sovereignty pillar, where it fits alongside the Data Commons proposal as the low-cost half of the same argument. And that member states publish the completed schedules for awarded contracts, because the comparison across governments is where the leverage lives, and it produces nothing until the answers are visible to one another.

The Caribbean AI Risk Management Council has published the risk-register version of this diagnostic, mapping each exposure to a quantification method and to controls under ISO/IEC 42001:2023 and the NIST AI Risk Management Framework, for institutions that need it in the format an audit committee already uses. StarApple AI runs the vendor diligence module for boards and procurement teams that want a trained person in the room before signature rather than a consultant after the audit. Both are available to CAIA members, and both exist because the alternative is fifteen procurement offices independently discovering the same thing at their own expense.

Disclosure: I founded the Caribbean AI Association in 2024 and lead StarApple AI, which builds custom and sovereign AI models for Caribbean institutions and therefore competes with some of the vendors this proposal would require to disclose. Every check set out above can be run against StarApple AI, and clients do run them.