StarApple AI Jamaica | Business field note

Agentic colonization and the Jamaican customer relationship

By Adrian Dunkley

Practical choices for Jamaican small businesses using agents to market, sell and serve customers.

Published 15 October 2026 · Commentary by Adrian Dunkley

AI-generated editorial scene of Jamaica adults reviewing local knowledge and decisions
AI-generated image by OpenAI for this article. People and setting are fictional; this is an editorial illustration, not documentary evidence.

A guesthouse owner in Montego Bay can use an assistant to answer booking questions while preparing rooms. A Kingston designer can turn a rough brief into a proposal after a long working day. A food business can organise orders that previously arrived as scattered messages. These are useful applications. The difficult question comes when the tool also starts deciding which customer matters, what the business promises and how the business sounds.

Adrian Dunkley calls the risk of transferring practical authority through dependence on externally controlled agents agentic colonization. The term is his framework for examining dependency. It should not be presented as proof that every agent, subscription or foreign supplier is harmful. For a Jamaican owner, the immediate test is whether the business can still explain its decisions and continue operating if the service changes.

There are two relationships to protect: the relationship with the customer and the relationship with the knowledge that makes the business distinctive. An agent can strengthen both. Poorly defined permissions can allow it to become the gatekeeper to both.

Decide what the assistant may promise

Start with a fictional guesthouse using an agent to draft replies. A visitor asks whether late checkout is possible, whether a child can join an excursion and whether a deposit is refundable. The assistant may retrieve the property's approved information. It should not invent flexibility, waive a condition or make a safety assurance merely because a friendly answer is likely to secure the booking.

Write a simple authority list. The agent may retrieve published information. It may draft a response for review. It may send specified routine messages if that permission is deliberately granted. Commitments involving prices, refunds, safety or unusual arrangements need their own boundary. These are different levels of authority, even if the interface presents one convenient button.

Keep an owner for each boundary. If the person responsible is away, the workflow should wait or route to another authorised person. An instruction to be helpful is not a substitute for a decision about who may commit the business. Staff should be able to see what the agent sent without hunting through several accounts.

Your voice is part of the product

A business selling Jamaican experiences should not need to sound like an interchangeable international booking page. Nor should it manufacture exaggerated dialect because a model associates Jamaica with a narrow set of cultural signals. Both outcomes can misrepresent the people doing the work.

Give the assistant a short, approved voice guide containing real examples the business owns or has permission to use. Explain which phrases customers find clear and which claims the business does not make. Ask a staff member who speaks with customers daily to review the output. Their knowledge is a business asset, not an obstacle to automation.

Keep the original human draft when using AI to edit campaign copy. Compare the versions for meaning, specificity and promises, not just grammar. If the generated version removes the local detail that customers actually value, reject that edit. A quicker writing process is worthwhile only when it preserves the reasons a customer chooses the business.

Follow the value out and back

Customer questions reveal demand. Repeated complaints reveal process failures. A guide's explanations reveal expertise. A designer's sketches reveal creative direction. Uploading these materials can improve a workflow, but the owner should know which materials leave the business, what the recipient may do with them and what remains available afterwards.

Dunkley's documented Extracted Intelligence concept provides a useful lens: examine who benefits from knowledge that becomes commercial input. A practical application is to draw two columns. In the first, list the data and work the business supplies. In the second, list the capability, records and rights it retains. This is an editorial exercise, not a published scoring formula.

Do not collapse all supplier terms into the word ownership. A business may retain rights to an input while granting a broad licence to process it. A service may exclude model training but retain records for other stated purposes. A connected application may have its own terms. Read the conditions for the exact product and account, and obtain specialist advice when the material or commitment warrants it.

Keep a direct route to customers

If an agent becomes the main way customers discover the business, its recommendations acquire commercial importance. A proprietor should understand whether listings are paid, what information is used and whether customers can reach an independent website or contact channel. The objective is to preserve choice instead of to reject every platform.

Maintain an accurate website, clear service descriptions and a human contact route. Keep a lawfully maintained customer record that the business can export, with the permissions and communication preferences attached. An address book stripped of consent history is not a complete operational backup.

Ask how an existing customer reaches you when a platform is unavailable. Test the answer. A printed contact card, a working phone line and a current booking ledger may sound ordinary, but they can preserve continuity more reliably than an untested promise that everything is in the cloud.

Run one bounded experiment this week

Apply People First, Problem Second, AI Last, Dunkley's documented approach, by starting with a human problem: customers wait too long for accurate answers about opening hours. Define success before choosing the tool. Faster replies that contain incorrect hours fail the task.

  1. Choose a low-risk message type. Use opening hours or the location information the business already publishes. Do not begin with refunds, health questions or disputes.
  2. Build a small reference sheet. Record the approved answers, the person who owns them and the date they were checked. Exclude private customer information.
  3. Review twenty drafts over 7 days. Record preparation time, correction time and the number of answers requiring a factual change. Compare with twenty comparable replies prepared without the agent where those records are available.
  4. Separate speed from reliability. Count time saved only after review. Do not describe the result as a general productivity benchmark or evidence that the tool is safe for other tasks.
  5. Rehearse the exit. Export the approved reference sheet and reply log, then have a staff member answer the next enquiry without the assistant.
  6. Expand only deliberately. Add a new message category after agreeing its permissions, evidence and reviewer. Keep consequential commitments outside the pilot.

The test may show that the assistant helps with drafts but not autonomous sending. That is a useful result. It may also reveal that the real problem is an outdated price list or unclear internal ownership. Fixing the information can be more valuable than buying another tool.

Media influence can affect a business from both sides

The owner uses AI to understand a market; the customer uses AI to decide where to spend. In both cases, summaries and recommendations can hide excluded alternatives. Ask the assistant to show the sources and selection criteria behind a comparison. Check important claims directly with the provider being compared.

For marketing, set a clear rule against fabricated reviews, invented customer stories and impersonation. If a campaign uses a synthetic person, do not present that person as a real customer. If an image is generated, label it where readers could otherwise mistake it for a photograph of the actual property, product or team.

Personalisation should help a customer understand a genuine offer. It should not exploit a private difficulty that the customer disclosed for another purpose. A small business can set that boundary before it has sophisticated technology or a large compliance department.

Risks and the weakest part of the argument

Dependence is not automatically abusive. A small firm may reasonably prefer a specialist supplier to the cost of maintaining its own system. Local suppliers can also impose poor terms, and self-hosting can introduce security and maintenance burdens. The argument here is for informed control, not a guarantee that a particular ownership model solves every problem. Customers bear the consequences of false promises; staff bear the work of correcting them; owners bear disruption when access fails. Narrow permissions, visible records, tested exports and a named human reduce those risks while preserving useful assistance.

Questions owners should ask

Does paying for an account keep our data out of training?

Payment alone is not a sufficient answer. Check the exact service, account category and settings. Ask separately about training, retention, access and connected applications. Record the answer and the date reviewed so a later change can trigger another decision.

Should we build our own model?

Only after identifying the need, operating costs, security responsibilities and people who will maintain it. A portable process using a commercial service may be more practical than an unsupported local installation. Compare control and total effort instead of treating either approach as inherently independent.

What belongs in an exit test?

Include the records needed to serve a customer, approved knowledge, permissions, outstanding commitments and a usable file format. Have someone other than the person who configured the tool perform a task from the export. A download that nobody can interpret is not a successful exit.

How do we preserve a Jamaican voice?

Use approved examples written by the people behind the business and keep a local reviewer involved. Describe the audience and purpose precisely. Preserve meaningful expressions when they fit, explain unfamiliar terms when necessary, and avoid asking a model to imitate a caricature.