The useful preparation
You can finish a report faster with AI and still send the wrong figure to a client. You can learn a new tool and accidentally expose the information your employer trusted you to protect. The useful skill is knowing what to check before the output leaves your desk.
In this guide, “AI Apocalypse” is a metaphor for disruption to work, skills and trust. The preparation is practical: understand the task, protect the information, check the answer, keep control of actions and practise recovery. These five habits give Caribbean professionals a way to use AI without handing over their judgement.
Why these checks matter now
On 8 October 2026, the OWASP Gen AI Security Project published its Q3 2026 exploit roundup. It brings reported incidents together with the project’s security analysis. Those reports are a prompt to test your own controls, not evidence that your organisation has suffered the same failures.
The NIST AI Risk Management Framework offers a voluntary structure for governing, mapping, measuring and managing risk. The five tips below translate that work into everyday checks. They do not replace the specialist review a regulated or high-impact use case needs.
1. Check the job, the vendor and the model
Start with the person who needs your work. A client waiting for an accurate quotation needs the right price and a clear scope. A polished paragraph does little for them if it changes either. Write down what a successful answer must contain before you compare tools.
My published approach, People First, Problem Second, AI Last, puts the human need before the technology. Applied here, it means defining the job and the possible harm before deciding how much work to delegate.
Check the company supplying the application and the provider supplying its model. Ask which version will handle your work, where your information travels, what external sources the answer draws on and who handles a failure. A local sales office does not establish local hosting. A model benchmark does not establish performance on your files.
Your checklist: Name the task owner. Record the tool and model version. Read the data terms. Ask how updates are communicated. Test a difficult local example with an answer you can check. Record unresolved questions before your manager approves wider use.
2. Keep private information out of casual prompts
A useful test rarely needs a real passport, salary sheet or customer complaint. Replace real people with invented examples first. Removing a name may leave an address, unusual job title or account history that still identifies someone.
Check the settings and contract for the specific account you use. Personal, business and API arrangements can differ. Find out whether prompts are used for training, how long records remain, who can read them and how deletion works. Ask your data owner before uploading restricted information, even when a colleague already uses the tool.
OWASP’s sensitive-information guidance explains why data access and sanitisation matter. An instruction such as “keep this confidential” is not an access control. A supplier’s promise about model training also does not answer every question about storage or support access.
Your checklist: Use the approved account. Minimise the input. Replace identifiers. Check attachments, screenshots and spreadsheet tabs. Keep passwords and access keys out. Pause when you cannot tell whether the information is allowed.
3. Verify the parts that could hurt someone
AI can invent a citation and format it beautifully. Open the source yourself. Check that it supports the sentence, applies to the right country and is current enough for the decision. A second chatbot repeating the claim is not independent evidence.
For a quotation, compare it word for word. For a calculation, use the underlying figures and a spreadsheet or calculator. For an unfamiliar local rule, ask the responsible specialist. Keep a short record of what you checked and what remains unknown. If your output needs certainty that the evidence cannot supply, return it for review.
This is an application of Cognitive Debt: skipping the checking step also removes a chance to develop judgement. Try completing a small part of the task yourself before reading the AI answer. That gives you something to compare, particularly when you are learning a new subject.
Your checklist: Check names, dates, amounts and sources. Mark assumptions. Recalculate totals. Ask who bears the cost of a wrong answer. Have a qualified person review consequential advice before anyone relies on it.
4. Separate drafting from permission to act
An assistant that drafts a response does not need permission to send it. A tool that summarises a folder does not need to delete files. Start with the smallest access needed for the task and add permissions only after you understand the consequence.
For actions affecting customers or staff, set an approval point outside the generated answer. A person should inspect the actual recipient, attachment, amount and proposed change before approving it. Do not rely on the model to decide whether its own action is allowed. OWASP describes this risk as excessive agency.
Treat instructions found inside emails, websites and documents as untrusted material. They may ask an assistant to ignore its task or send information elsewhere. A clear prompt helps, but permissions and approval controls must still hold when the prompt fails.
Your checklist: Prefer read-only access. Limit connected folders. Require review before sending, paying, publishing or deleting. Know how to revoke access. Keep a usable manual route for the same job.
5. Keep a record and practise stopping
Write down the tool, task, date, reviewer and any important correction. Keep only the evidence you need in an approved location. A risk log that copies confidential prompts into a shared spreadsheet can create another problem.
Agree who to contact when a tool exposes information or gives a harmful answer. Rehearse pausing the workflow and switching to the manual method. Preserve relevant evidence safely; avoid spreading the material through an informal group chat. The responsible team can decide whether customers, management or regulators need notification.
Repeat a known test when the model, prompt, data source or permissions change. NIST’s Generative AI Profile treats risk management as work across the system’s lifecycle. A successful demonstration last month does not settle the question after an update.
Your checklist: Record corrections and near misses. Name an incident contact. Test the stop mechanism. Check the fallback. Set a review date and retest after meaningful changes.
A Dominica example: keeping local facts intact
Consider a worked example in Dominica: a Roseau-based administrator helps a small tour operator prepare replies about port pickup and trips to the island’s interior. This is an illustration, not a report about a named company. The assistant can turn approved facts into a clear message; a staff member still checks the port, pickup point, date and currency before it goes to a visitor.
Dominica and the Dominican Republic must not be interchangeable in the source material. Nor should a model turn Eastern Caribbean dollar amounts into US dollars without saying so. Route advice drawn from an old page needs a current check with the operator. An attractive itinerary may still be impossible on the day a customer arrives.
This is our practical application of the research, not a measured result for Dominica. The studies below do not establish how well a particular model handles the operator’s local knowledge. A small, supervised test is the place to find out, with the person who knows the work checking every proposed reply.
What the evidence can tell you
The November 2023 version of Generative AI at Work studied 5,179 support agents. It reported 14% more issues resolved per hour on average with an assistant. That is a result for the studied setting, not a forecast for your team.
A field experiment with 758 consultants found 12.2% more completed tasks and 25.1% faster completion on tasks within the tested AI’s capability. The same paper reports worse performance on a task outside that capability. Its GPT-4 experiment does not validate today’s product on your workflow.
The ILO’s 2025 occupational-exposure update estimates that one in four workers worldwide (25%) is in an occupation with some exposure to generative AI. Exposure describes potential task change. It is not a count of people who have lost their jobs.
Stanford’s 2025 AI Index recorded 233 reported AI incidents in 2024, up 56.4% on 2023 in its cited database. Reporting patterns and coverage affect that count; it cannot tell you the probability of an incident in your business.
A practical test for this week
This is a suggested exercise, not a claim about likely savings. Choose a low-risk task with public or synthetic information. Do not use the trial to make decisions about real customers or employees.
- Write the task, owner, permitted data and stop conditions on one page.
- Measure a manual baseline on 10 representative examples. Record time, correctness and missing information.
- Test the same examples with AI and human review within seven days. Include checking and correction time, then compare both quality and total time with the baseline.
- Review every error with the owner before expanding use. Pause if the trial exposes information or acts beyond its permission. Keep a record of what must change.
Ten examples can reveal a problem; they cannot establish a reliable error rate or certify safety. Higher-impact work needs a larger, more varied evaluation and specialist review before deployment.
Where the checklist can still fail
A careful reviewer can miss an error, especially when tired or unfamiliar with the subject. A supplier can change a model or a setting. A small test can overlook a rare case that causes serious harm. The people bearing that harm may be customers whose information is exposed, staff judged on an incorrect record or a business that makes a promise it cannot fulfil.
Match the control to the consequence. Keep sensitive decisions with qualified people, test access boundaries and give affected people a way to challenge an outcome. The weakest part of this advice is the local evidence: the cited studies do not measure these example Caribbean workflows. Use them to justify careful evaluation, then collect your own evidence. Training and written rules need enforcement and time for review.
Questions to settle before you start
What if my employer has no approved AI policy?
Ask your manager or data owner for a written boundary before using work information. You can practise with public material or invented examples while that decision is pending. Record which tool you want to use, the task, the information needed and who will check the output.
Should I disclose that AI helped write my work?
Check your employer’s rules, your client’s requirements and any professional obligations before delivery. Keep a record of the tool’s contribution and your own checks. If the audience asks, describe the assistance accurately. Never claim that a person independently produced or verified work when that did not happen.
Can I use a personal AI account if the approved work tool is unavailable?
Do not move work information to a personal account simply to meet a deadline. Follow your organisation’s approved fallback, such as completing the task manually or using an authorised alternative. Tell the task owner about the outage and agree a revised delivery time if needed. Ask for written approval before changing tools.
Does buying the book enrol me in the course?
The book and the course enquiry are separate. The book page gives the current ordering options. To discuss training, contact StarApple AI and confirm the delivery format, duration, availability, fees and any assessment requirements before making a booking. This article promises none of those arrangements.
Put the checks into practice
Survival Guide to the AI Apocalypse: A Practical Guide to AI Risk Management
Discuss the course with StarApple AI and work through the controls for your context. Confirm format, availability, fees and any assessment requirements before booking.
View the courseRead Survival Guide to the AI Apocalypse by Adrian Dunkley for the book’s current ordering options.