The Island Factory
Enterprise AI-factory discipline, scaled to a crew of one.
The world's largest companies stopped treating AI as a pile of projects and started running it like a factory: standardized parts, a production line, honest gauges. We run the same discipline at the other end of the scale — live, every day, in our founder's own island tourism business — and install it for operators who will never have an IT department. An AI factory small enough to run on an island, strong enough to run without you.
Moat & Meter decides where your AI dollars go. The Token Ledger controls what they cost. The Island Factory is how the thing gets built — from standard parts, shipped with guardrails, adopted for real, and steered on autopilot.
Why factories beat projects
Most small-business AI fails the same way: a bespoke build for every idea, no dollar figure attached before the work starts, nothing reused, nobody trained, and no gauge that says whether money followed. HBR research puts AI project failure rates as high as 80% — roughly double ordinary IT projects.* A factory attacks every one of those failure modes structurally: every candidate workflow earns its slot, every build is made of parts the next build reuses, everything ships with guardrails, adoption is somebody's job, and the whole line reports to the owner.
The five stations
Station 1
The Intake — what earns a slot on the line?
Tell Projects get picked because they demo well, with no dollar figure attached before anyone builds — that's AI Theater with a budget.
Move Every candidate workflow gets a Factory Ticket: its Moat & Meter quadrant, a Meter number (revenue touched or hours saved per month), a data check, and the standard parts it would reuse. No ticket, no build.
We run this ourselves: our AI concierge was built for the ferry window — the 90–150 minutes before visitors land, when booking decisions actually happen. The intake was a revenue thesis, not a tech demo.
Station 2
The Parts Bin — why does every project start from a blank file?
Tell Each new automation gets a bespoke build and a fresh invoice; nothing from the last project survives into the next.
Move The Second-Use Rule: nothing custom gets built unless it's designed to be reused by the next workflow. The small-business parts bin is about six stock blocks — booking & checkout core, payment rail, AI brain with an owned context store, message rail, event tracking, scheduler. A "new product" is those blocks plus a thin business layer.
We run this ourselves: e-bikes, ferry tickets, glamping, and grocery delivery all run on the same six blocks — one operator, roughly 230 production functions — and every new vertical costs less to launch than the one before it.
Station 3
The Loading Dock — is it production-grade, or does it just work on a laptop?
Tell An AI feature touching money or customers with no duplicate-action guard, no kill switch, and no plan for the day the model is down.
Move The Ship Checklist: duplicate guards on anything that charges, a kill switch on every risky behavior, fail-open design — if the AI layer dies, the business still runs — and a verify-live step after every deploy.
We run this ourselves: after a real multi-tap incident produced three charges, we installed a four-layer duplicate guard. Our premium tier sits behind a kill switch. Our cost-cache is flag-gated and fails open. We learned each of these the expensive way so clients don't have to.
Station 4
The Floor Shift — launched is not adopted. Who's actually using this?
Tell The gauge lies. Dashboards report activity, not usage; "sent" isn't delivered; funnels inflate; nobody on staff was ever trained. This is the station everyone skips — HBS's research on enterprise AI operations shows the giants dedicate entire budget cycles to adoption alone, while small businesses budget zero for it and then blame the tool.
Move Meter real usage tied to money. Fix the leak the data names, not the one you assumed. Name one adoption owner. Train staff on the tool in week one.
We run this ourselves: our own "68% checkout abandonment" turned out to be a broken gauge — the real leak was sticker shock at one step, so we fixed the pricing display, not the bot. And every night we mine unanswered guest searches to learn what people want that we don't sell yet.
Station 5
The Control Room — does the factory report to you, or are you the monitoring system?
Tell The owner reads raw dashboards at 11pm; costs drift silently; nobody would notice a dead scheduled job for a week.
Move An automated ops digest that tells the owner what needs a human
today; cost governance via the
Token Ledger as a standing discipline; model-swap readiness so the factory never marries a vendor.
We run this ourselves: an AI Autopilot digest reads our whole business twice a day — 6am and 2pm — and surfaces exactly what needs the owner. Model tiering and caching keep the AI's own bill on the right side of the Meter.
The Front Desk — the GenAI layer
The rule
Rent the brain. Own the memory.
The customer-facing AI voice bolts onto the factory — it isn't the factory. The model is rented and swappable behind an abstraction layer; the assets that compound are yours: the context store (your facts, prices, policies), the prompts, the event history, the customer data. A business that owns its memory can change brains in an afternoon. A business that rents both is a feature of someone else's product.
We run this ourselves: our concierge runs a frontier model as the primary brain with a cheaper model tiered behind a flag — and every fact it knows lives in our database, not in the prompt. When something changes on the island, the brain updates without a deploy.
The three factory rules
I
Second use pays
Standard parts beat bespoke builds. If it can't be reused, it doesn't get built.
II
No meter, no machine
Nothing ships uninstrumented. Every automation must report its own dollar impact.
III
Adoption is a job, not a hope
Someone owns usage — or the machine rusts while the invoices keep coming.
How we install it
$397The AI Teardown — Station 1 as a product. Sixty minutes on your operation; you leave holding three Factory Tickets: workflow, quadrant, monthly dollar impact, parts required, install price. Three automatable workflows or it's free — and the fee credits toward the install.
$1,500–2,500The Island Factory Install — Stations 1–4, done for you. Your highest-Meter workflow built from battle-tested blocks (that's why it isn't $20,000), shipped with duplicate guards and kill switches, your staff trained, real usage metered.
$150–300/moThe Control Room — Station 5 as a subscription. We watch the gauges, keep the Token Ledger, re-tier models as vendors reprice, and send you the digest. You get fewer phone calls; the factory reports to you.
Deciding which workflows deserve a ticket at all is Moat & Meter — for larger organizations, that's the AI Value Diagnostic.
The reading spine
Iansiti & Lakhani, Competing in the Age of AI (HBS Press, 2020) — the AI-factory concept.
Bojinov, "Keep Your AI Projects on Track," Harvard Business Review (Nov–Dec 2023) — *the failure-rate research and the five-step AI lifecycle.
Davenport & Mittal, All-in on AI (HBR Press, 2023) — why scattered pilots don't compound and systems do.
The Island Factory is original Isla Tech methodology, informed by the AI-factory concept developed at Harvard Business School by Karim Lakhani, Marco Iansiti, and Iavor Bojinov, and by HBS's published research on enterprise AI operations. Isla Tech is not affiliated with or endorsed by Harvard Business School or Harvard Business Publishing.