Home / Blog / Agentic colonization

Founder field note

Your AI Startup Might Be Building Someone Else’s Moat

Map the dependencies that can reprice your business before your customers depend on it.

By Adrian Dunkley15 October 20267 min read

AI-generated scene of Caribbean founders reviewing a laptop and notes in a Kingston-inspired office
AI-generated illustration created with OpenAI on 15 October 2026. Fictional people and setting, not documentary photography.

You can launch an impressive AI product while owning very little of the business underneath it. The model comes from a platform. Discovery happens inside another platform. Customer history sits in a managed integration. Your team supplies the positioning, local knowledge and support, then pays recurring fees for the machinery that makes those contributions usable.

In his essay Agentic colonization and the decisions we give away, Adrian Dunkley calls the broader risk agentic colonization. Access becomes integration; integration becomes influence over decisions. For founders, the uncomfortable question is whether the product is building durable customer value that the company controls, or strengthening dependencies that someone else can reprice.

Using external infrastructure is ordinary and often sensible. The danger grows when a startup mistakes rented capability for a defensible position and discovers the difference only after customers arrive.

Draw the dependency map before the pitch deck

Map the full route from discovery to renewal. Who introduces the customer? Who controls the interface? Where does the customer’s working history live? Which service interprets the request, chooses the next tool and completes the action? Who can change the terms of each step?

Mark every point where a single provider can interrupt delivery or alter the economics. Include identity, payments, messaging and retrieval, not just the model API. An architecture with two model providers can still have one irreplaceable customer-data store or distribution channel.

Now ask what the startup genuinely owns or can reliably carry elsewhere: customer relationships, properly licensed datasets, workflow expertise, product design, evaluation cases and operational knowledge. A list of integrations is useful engineering. It is not, by itself, a strong answer to the ownership question.

Run the price-change test

Consider an illustrative startup selling an agent to small tourism businesses. The product drafts guest messages, updates schedules and answers local questions. Usage looks healthy. But successful customers require more tool calls, longer context and additional support, increasing the cost of serving the accounts the company most wants to retain.

Build contribution-margin scenarios using actual provider pricing and observed usage. Include retries, failed calls, retrieval, storage, monitoring and human escalation. Separate the cost of a demonstration from the cost of completing a real customer task reliably.

Stress the assumptions: higher usage, a provider price increase, a service interruption or a move to a replacement model. These are planning scenarios, not forecasts. Identify the point at which the current subscription stops making sense. If nobody can answer that question, growth may be concealing a dependency problem.

Customers should recognize your company’s value

Ask a customer why they would stay if your model provider released a similar feature tomorrow. Useful answers might include trusted local support, a deeply understood workflow, reliable integration or expertise that helps them avoid costly mistakes. “We have access to the model” is fragile.

For a Caribbean founder, local knowledge can be valuable when it is specific and operational. Knowing how an independent operator handles an exception may matter more than producing another generic travel paragraph. Validate that value with customers rather than assuming that a local label creates a defensible business.

Maintain a direct, consent-based customer relationship. Provide useful exports of customer-owned records. Document the workflow so service does not depend entirely on one opaque agent session. Customers should be able to understand what the company contributes and how their information moves through the product.

Creative extraction can become a bad business model

A founder may feed a platform a library of local campaigns, stories, images or language examples to make a product feel distinctive. That material has creators, rights and context. Calling it “data” does not answer who should benefit from its use.

The commercial risk is circular: creators supply cultural value, a platform packages useful capabilities around a wider body of material, and the same creative economy pays subscriptions to access the resulting tools. How that applies to a particular provider depends on its practices, contracts and the material involved. It should not be presented as a proven theft claim without evidence.

Know the licences for the material you use. Pay for commissioned work. Review whether uploaded content may be retained or used beyond the requested service. Give contributors clear terms where their work is central to the product. If the startup’s distinction depends on a community’s expression, build a credible way for that community to share in the value.

Do not outsource your editorial judgment

An automated growth system can research trends, generate messages, select audiences and distribute variations. Combining those steps creates a route by which a platform’s framing can become the startup’s public voice. It also creates opportunities for persuasive messaging to scale faster than careful review.

Mass manipulation is a risk to design against. NIST’s Generative AI Profile identifies the potential for large-scale disinformation and misinformation as an information-integrity concern. Watch for hidden sponsorship, unsupported claims, synthetic testimonials and personalization that exploits vulnerability. Keep responsibility for campaign objectives and sensitive audience choices with accountable people.

Cultural dilution can be less dramatic and more persistent. A tool repeatedly replaces distinctive language with safe global phrasing. Soon, every local brand sounds like the same subscription service. Use local editors who can challenge both the wording and the assumptions. Measure whether customers understand and trust the communication, not simply whether more content was produced.

Earn the right to automate consequential actions

Start with a narrow workflow whose errors are observable and reversible. Give the agent only the permissions that workflow needs, consistent with OWASP’s guidance on excessive agency. Require explicit approval for actions with material financial, legal or reputational consequences, according to the customer’s policy.

Keep action records, define a suspension procedure and test recovery. Tell customers which parts of the service depend on external providers. Confidence should come from demonstrated reliability and clear responsibility, rather than from hiding the supply chain behind a polished interface.

Before expanding, run a practical portability test. Move a representative task to another implementation or a manual fallback. Record the quality loss, migration effort and missing information. You do not need to rebuild every component in-house; you do need to understand the price of losing access.

The strongest answer to agentic colonization is a business that can choose its dependencies deliberately. Build something customers value, compensate the people whose work gives it distinction, and retain enough control to keep serving them when a platform changes its mind.

The dependency worksheet
1Map discovery to renewal
2Price one completed task
3Test a fallback

A suggested editorial exercise for founders, not a measured research result.

Do this now

Within seven days, map one paid workflow, measure ten completed tasks and run one synthetic case through a fallback. Record the missing records, extra cost and permissions you could not revoke.

Founder questions

How do I test dependency without moving customer data?

Use a synthetic case that reproduces the workflow without personal information. Run it through your current provider and a manual fallback. Compare output quality, time and permissions. Test a real export only after checking the customer’s agreement and the destination’s safeguards.

What should the worksheet calculate?

Start with revenue minus provider calls, retries, storage, monitoring and human support for one completed task. In a hypothetical example, US$20 revenue minus US$8 delivery cost leaves US$12 before fixed costs. Replace every input with your measured figures; this is arithmetic, not a forecast.

How do I protect a creator’s contribution?

List each commissioned asset, its creator, permitted uses and payment terms. Keep the licence alongside the original file. Check separately whether a vendor may retain uploads or use them for training. A licence to publish a campaign does not automatically settle those additional uses.

Put your business assumptions on paper.

Use the dependency map alongside the operating questions in Kill My Startup.

Buy on Amazon