Running a real business with AI as the operator.
Every framework we sell has a lab where it either survives or dies: an island tourism company we built and operate — run essentially by one person, because the AI carries the load — off the east coast of Puerto Rico. This is what the AI actually does there, and what it proves for your business.
One person. A whole tourism company. An island.
Culebra is a small island east of Puerto Rico's main island — reached by ferry, home to some of the best beaches in the Caribbean, and busy with visitors who step off the boat with a few hours to spend and a lot of questions. Todo Culebra is the company we built to serve them: e-bike rentals, a ferry-ticket concierge, food delivery, and glamping, all running from a live interactive map of the island.
Here's the part that matters for you: the company is built and operated essentially by one person. Not because it's small in scope — it runs quote-to-cash, customer support, logistics, and analytics every single day — but because AI systems carry the operating load.
The operating reality is unforgiving in a useful way. Guests plan their day on the ferry ride over, which means the business has to be able to answer, quote, and close a sale in the window before they step off the boat — with no staff standing by. Payments have to reconcile themselves. A cancellation at dawn has to release inventory without anyone touching a keyboard. Island tourism compresses every operations problem a larger company has into a single day, over and over.
Todo Culebra is our lab. Every AI strategy we put in front of a client has already survived contact with real customers, real refund requests, real early-morning ferry questions, and a real P&L — ours. When we tell you what AI can run and where it falls over, we're not quoting a survey. We're reading our own dashboard.
What the AI actually does.
None of this is a demo or a pilot. These systems are in production, facing paying customers, right now.
Coqui, the AI concierge
A conversational AI on the map and site answers guest questions around the clock — which beach for snorkeling, where to eat tonight, how the ferry works — quotes prices, and closes bookings. It runs inside guardrails and escalates to a human the moment a conversation needs one.
Booking changes by chat
Guests change their time, their date, their bike count — or cancel — right in the chat. No phone tag, no inbox archaeology. The system updates the booking, the payment, and the confirmation on its own.
Automated payments
Stripe checkout with a small deposit that locks the card; the balance auto-charges before the ride. Duplicate-charge guards protect guests from multi-tap mistakes, and refunds are handled from an owner dashboard.
Ferry-ticket concierge
Tiered volume pricing on the island's scarcest resource. A human makes the final regulated purchase — by design — while the AI handles everything around it: quote, payment, delivery emails, change requests.
Nightly demand signals
A nightly job mines what guests actually search for and surfaces trending and unmet demand on the live map. The market tells us what to build next — in writing, every morning.
The business briefs its owner
First-party analytics feed a morning operations digest: money owed, chats needing attention, the day's schedule. The owner starts each day with the business reporting to him — not the other way around.
AI that watches its own bill
Prompt caching, a semantic answer cache, and cheaper-model routing — engineered to cut a monthly Anthropic bill measured in hundreds of dollars. The Meter applies to the AI itself.
Price experiments from config
Per-date price experiments run from configuration — no deploys, no engineering cycle standing between a pricing question and its answer. Change the number, watch the market respond.
A one-person AI factory.
Harvard Business School researchers — Marco Iansiti, Karim Lakhani, and Iavor Bojinov — have documented how the world's largest companies industrialize AI: standardized, reusable components instead of bespoke builds; models deployed into real business processes; adoption measured honestly. They call it an AI factory. Todo Culebra is that same architecture at the other end of the scale — a miniature factory, run by one person.
Five verticals, one parts bin
E-bikes, ferry tickets, camping & glamping, food delivery, and grocery prep all run on the same reusable blocks — booking core, payment rail, AI brain, message rail, event tracking, scheduler. Every new vertical costs less to launch than the one before it. That's a factory floor, not a pile of projects.
~230 cloud functions, one operator
The platform runs roughly 230 production functions — quoting, charging, confirming, reconciling, reporting — maintained by a single person, because every component was built to be reused, monitored, and swapped.
Shipped the expensive lessons
A real multi-tap incident once produced three charges on one booking — today a four-layer duplicate guard makes that impossible. Risky features sit behind kill switches. The AI layer fails open: if a model goes down, the business keeps selling.
The dashboard that lied
Our own analytics once reported a 68% checkout abandonment rate — instrumenting the real funnel showed a broken gauge and a very different leak: sticker shock at one specific step. We fixed the pricing display, not the bot. Measuring adoption honestly is a discipline, not a dashboard.
We've turned that operating discipline into our third framework: The Island Factory — enterprise AI-factory practice, scaled to a crew of one. Moat & Meter decides where your AI dollars go. The Token Ledger controls what they cost. The Island Factory is how the thing gets built.
The AI-factory concept was developed at Harvard Business School; the Island Factory is original Isla Tech methodology informed by that published research. Isla Tech is not affiliated with or endorsed by Harvard Business School or Harvard Business Publishing.
Moat & Meter, eaten as our own dogfood.
We don't just sell the matrix — we run our own company on it. Here's where the systems above actually landed, and what each placement forced us to do.
The concierge, enclosed
Coqui started as pure Commons — anyone can rent a chatbot. We enclosed it with what only we hold: proprietary island knowledge, live pricing and availability, and the accumulated booking and conversation history of the business. A rival can rent the same model tomorrow and still not have the thing that makes ours sell. That's the Enclose-the-Commons play, executed on our own P&L.
Killed on sight
Plenty of experiments demoed beautifully and defended nothing — impressive output, no fence around the value, no movement on the Meter. They got shut down, which is exactly the advice we give clients. The discipline to kill your own toys is most of the framework.
Demand signals compound
The demand-signals loop is the feedback flywheel: guests search, the map learns, the concierge answers better, more guests use it, the signal sharpens. Every night the moat gets a little deeper — on first-party data no competitor can rent, buy, or scrape.
Cost engineering as strategy
An AI system that moves the P&L but costs more than it moves is theater with better lighting. Caching, semantic reuse, and model routing keep the concierge's own bill on the right side of the Meter — the same audit we run on every client system.
The point isn't tourism. The point is where you aim.
Nothing about e-bikes or ferry tickets makes this story special. What Todo Culebra demonstrates is a claim most leadership teams still treat as hypothetical: one person plus well-aimed AI can run quote-to-cash, customer support, logistics, and analytics — simultaneously, in production, with real money and real customers on the line.
The operative word is aimed. The same models, pointed at the wrong problems, would have produced an expensive chatbot and a graveyard of pilots. The difference between AI that compounds and AI that burns budget was never the model — it's where you point it, what proprietary asset you enclose around it, and what you're willing to kill.
Notice, too, what stayed human. The regulated ferry purchase. The escalation path out of the concierge. The final tap on a refund. The system is deliberate about which decisions a person keeps — that boundary is a design choice, and drawing it well is half the work.
That judgment is what you're hiring. When we plot your AI portfolio on Moat & Meter, we're applying tests we've already run against our own money — including the uncomfortable ones, like admitting a system we enjoyed building was theater. The Diagnostic is where we run those same tests against yours.
Every AI strategy we sell has already survived contact with real customers and a real P&L — ours.
See it applied to your business.
Bring your current and planned AI bets. Every one comes back plotted on Moat & Meter — the Fortress bets to fund hard, the AI Theater to shut down — on a single slide your board reads in ninety seconds.