A Willemstad business considering an AI assistant has several decisions to make. Can staff and customers use it in the languages they need? Can they reach it on their own devices? What will it cost after checking and support are included? A local server can help with part of that work, but it cannot answer every question.
The Inter-American Development Bank's September announcement provides a reason to take those practical details seriously. Its regional scenarios make worker mobility important to the distribution of AI's benefits. Curaçao's computing services, Papiamentu language project and national strategy process offer concrete places to examine that issue, without pretending a regional model predicts an individual island's wages.
TL;DR
- The IDB's regional scenarios are conditional estimates, not a wage forecast for Curaçao.
- Current provider and project pages describe GPU services and a Papiamentu model under development. Their claims need to be checked against a buyer's particular task.
- Census figures concern facilities in living accommodations, not the percentage of people who can use AI.
- A small business can compare costs and run a supervised language exercise with public or invented material. A short test cannot certify a system for client work.
What the IDB announcement says
In its 21 September 2026 announcement, the IDB described findings from its forthcoming report, From Digitalization to Artificial Intelligence: Turning Promises into Productivity. Under broad adoption and large productivity effects, regional GDP could be 5.1% higher after a decade. With limited adoption and small effects, the estimate is 0.3%.
The wage scenarios range from increases of 2.3% to 5.3% where workers move into expanding job areas, to falls of 13.5% to 20.9% where they cannot. These are conditional regional results. They do not establish an island-specific outcome or show that retraining alone determines which scenario occurs.
The announcement says the full report is due in November. This article uses the primary release; the full report's methods are outside this review. The practical question for a business is narrower: what changes in the work, who needs support, and how will the result be assessed?
Separate the measures
The Central Bureau of Statistics Census 2023 table lists internet in 71.7% of living accommodations, a computer in 45.2% and a smartphone in 67.7%. These are accommodation-level facilities, not counts of individual users, firms or AI skills. By subtraction, 28.3% were not recorded as having internet. That does not reveal whether a particular person could connect elsewhere.
A founder should test the intended service with the actual users and devices it is meant to serve. The census supports asking about access; it does not prove that every product must be phone-only or that a laptop-based service excludes most individuals. Device availability, accessibility needs, connection reliability and preferred language need separate checks.
ECLAC and CENIA’s October 2025 index release places the region's share of global AI investment at 1.12%, compared with 6.6% of global GDP. It also estimates 14% of visits to AI solutions against 11% of internet users. Those measures describe different activities and denominators. They are not a single adoption score, and none of them gives Curaçao's position by itself.
For a business, visits to a tool are a weak substitute for an outcome. A useful measure might be the time needed to prepare and check a routine response, the number of corrections, or whether a customer could complete the task. Decide which result matters before choosing the software.
Start with the work
Adrian Dunkley's People First, Problem Second, AI Last framework starts with the human need and the problem before selecting a technology. Applied here, it means asking what a Willemstad business needs its assistant to do, rather than treating access to computing as a reason to automate.
Consider a hypothetical accounting office that receives routine enquiries in Dutch, English and Papiamentu. Staff might want help finding an approved appointment policy or drafting a reply. That is a different task from interpreting a client's tax position. The latter carries greater consequences and needs qualified professional judgement. An assistant should not move from one task to the other because its answers sound confident.
Map the current process first: who receives the question, where the approved answer lives, who checks it and how an uncertain case is escalated. Keep that route available during a trial. If staff spend more time correcting drafts than they save, record the result rather than counting generation speed as productivity.
What Curaçao's projects currently describe
The BlueNAP.AI provider page, checked on 13 October 2026, advertises GPU and container services. Its pricing panel lists an on-demand starting rate of $2.99 per GPU-hour, alongside commitment periods, and lower reserved starting rates. The panel uses a dollar symbol without an explicit currency code. Ask for a dated written quote with the currency, hardware, term and full charges confirmed.
That page is a supplier's description, not an independent performance or compliance audit. Neither the word sovereign nor the location of a data centre settles every question about access, subcontractors, retention or applicable law. A buyer needs the service terms and technical design for the actual workload. There is no like-for-like price comparison here because the hardware and contractual terms may differ.
LengaGPT’s own project page describes a Papiamentu language model developed and managed by Stichting SPLIKA. It says the system is under active development, can make mistakes and is not an official language authority. It also warns users against entering confidential or sensitive information. The page presents future open-source availability as a plan; it does not establish that a production-ready release is available for this office.
The project offers a way to investigate local-language tools. It is not evidence that a particular task is already solved. A business should check release status, permitted uses, supported language varieties and how to report errors. A bilingual reviewer should assess meaning and tone, including whether the answer fits the local context, rather than judging fluency alone.
The Regulatory Authority of Curaçao’s 14 July 2026 release announced a forthcoming consultation for the national AI strategy, with RAC coordinating participation. It named education, science, business, civil society and government as contributors, and identified privacy, safety and transparency among the concerns. This dated announcement is not confirmation that submissions remain open today. Check the current process before sending material.
Publisher disclosure
14West was founded by Adrian Dunkley, who also founded StarApple AI. StarApple describes itself as the Caribbean's first AI company. The framework application here is this publisher's analysis. Affiliated sites such as Caribbean AI and the Caribbean AI Risk Management Council are related reading, not independent verification of the external claims.
A small exercise for this month
The dates below are suggested planning dates, not official deadlines. Agree the exercise with the person responsible for the service and use an approved tool. Keep client documents, account details and private messages out of it.
- Choose one low-risk task by 20 October. Write down the user need, the approved information source and who checks the answer. Prepare a baseline by timing how staff handle the same type of task without the assistant. Avoid legal advice, tax decisions or other consequential judgements in this exercise.
- Prepare ten paired questions by 23 October. A competent bilingual reviewer writes ten public or fictional enquiries in English and Papiamentu with equivalent meaning and checked answers. That produces twenty inputs. If Dutch is part of the intended service, assess it separately too. Record the tool version and conditions so another reviewer can understand the result.
- Review every answer by 26 October. Count factual errors, unsupported claims, wrong-language replies and cases needing escalation. Record review time as well as response time. Keep the findings separate by language and task. There is no pass mark that makes this small set a reliable estimate of real-world safety.
- Compare the whole workflow. Include correction time, subscription or computing charges and support needs. Ask the supplier for a written cost example. Try fresh questions after any change, and retain human review of client-facing material. Decide whether to continue, narrow the task or stop based on the evidence and risk.
- Check the consultation route. If the business wants to contribute to the strategy, ask RAC about the current invitation and format. Share a short account of the task and findings using invented examples. A 30 October follow-up would be the business's own planning choice, not a government submission deadline.
Who carries the risk?
Workers may have to learn a changed process while still meeting their ordinary responsibilities. Give them time, support and a way to raise problems. Do not turn a small tool trial into a ranking of staff or assume that an error reflects an employee's ability. Decisions about jobs and pay require much more evidence than this exercise supplies.
Customers can be harmed by a wrong answer that sounds authoritative, especially when it concerns money or obligations. Keep the assistant within the agreed task, make escalation easy and ensure a responsible person checks material before use. A model built for a local language still needs testing for the particular purpose. Local origin is not a safety certificate.
Data donation needs its own decision. The SPLIKA donation page now links a portal and tells prospective donors to read an agreement and appendix, then request an account. It says they should donate only material they are entitled to share. Do not upload customer files merely to help a model learn Papiamentu. Rights, privacy and the stated training purpose need review before any transfer.
Access can also shape who benefits. Provide an alternative route for a customer who cannot use the tool or does not want to. Test accessibility and low-bandwidth use where relevant. Keep the ordinary service available while assessing whether the new route improves it.
What this analysis cannot establish
We have not benchmarked BlueNAP's service, evaluated a LengaGPT release or measured Curaçao's wage response to AI. The evidence combines regional scenarios, dated census facilities and project descriptions. Those sources help frame questions, but cannot establish the commercial return or reliability of a particular product.
The strongest next evidence would come from a clearly described local task with independently checked results and full costs. Even then, a successful small trial would justify further evaluation within its scope. It would not establish that every office, language use or customer situation will work the same way.
Questions readers ask
What should a supplier quote include?
Ask for the GPU model and memory, the billing unit, minimum commitment, taxes, storage, data-transfer fees and support costs. Request a worked example for your proposed workload. A headline hourly rate alone cannot establish the total bill or a like-for-like saving against another service.
Can a small language test certify a chatbot?
No. Ten paired questions can expose problems worth investigating, but they do not represent every speaker, dialect, topic or task. Keep the answer key, tool version and test conditions. Use fresh examples after changes and retain qualified review for consequential or client-facing material.
What should a data donor check first?
Confirm that you have the rights and authority to share each document. Read the project agreement and appendix, including permitted uses and any limits on withdrawal. Remove private information and obtain suitable advice where others have rights in the material. Possessing a file does not automatically authorise its donation.
How should a business report an unresolved problem?
Describe the task, tool version, language, expected answer and observed error using public or invented examples. Keep client records out of the report. Send it through the supplier or project’s official channel, and distinguish an isolated example from a measured error rate across a defined sample.
Sources
About the author: Lancelot Williams
Lancelot Williams is Startup Correspondent at 14West. He covers founders, funding and national AI policy. See more of the publisher’s analysis in the 14West blog.
Supported by StarApple AI.