Cheap to Build, Hard to Win: Four Checks Before Your AI Business

Making money with AI has never looked easier: a few commands spin up a website, a logo, an assistant. Sandeep Swadia argues the danger hides inside that very ease. A speaker who once ran technology companies worth billions, he insists that cheaper production and easier winning are not the same thing, calling the confusion the artificial mirage . His analogy comes from television: forty years ago in America, an aspiring show knocked on three doors, and when all three closed the show barely existed; today anyone with a phone and a connection publishes without permission, yet twenty million new clips land by morning. Production belongs to everyone, attention to the few. Forbes issues a twin warning for the startup economy: the generative AI boom unleashed a flood of newcomers, yet the overwhelming majority are thin shells resting on the ready-made interfaces of large model providers. When everyone holds the same tools, the tools protect no one; a bubble keeps inflating in the middle of it all.

Against that picture Swadia lays out a four-gate route: the DEAL frame . Four letters, four doors: demand, edge, access and loops. Demand proves somebody pays; edge explains why the business is yours; access maps the real road to the payer; loops make each lap teach the next one. The order is deliberate: without proof of payment, edge means little, and without a road, even the finest product stays invisible. That sequence holds in a sluggish economy as in a hot one, never the reverse. The rest of the talk tests each gate with its own case, from electric cars to razor subscriptions, from neighbourhood laundries to a violin workshop.

Demand: the wallet rules

The first gate fits in one sentence: a good idea reveals itself the moment somebody starts paying. The evidence reaches back to 2006, to Tesla's most fragile days. The company wants millions of electric cars, yet no buyer is certain; against Toyota and BMW, almost nobody gives it a chance. The answer is to unveil the Roadster and offer the first hundred cars as a signature series: buyers place a hundred-thousand deposit and receive no car. The series sells out in three weeks; by August 2007, 570 Roadsters are spoken for, before manufacturing even begins. The lesson is blunt: customers vote and veto with their wallets. Without payment, be ready to change the offer, the crowd, even the idea itself.

Testing demand without fooling yourself takes three moves. First, pick the niche inside the niche: small businesses in general are no customer at all, while dental practices with fewer than three offices across three nearby towns are. Independent hiring advisers and online sellers under a million in yearly turnover do the same job. The narrower the buyer, the clearer the gap between assumed needs and real ones. Then talk to payers, not friends: friends comfort, while true buyers name the pain, the obstacle and the sum they would spend to make it vanish. Finally, build the smallest test: hand a rough AI-built version to future buyers and keep adjusting until value shows. Most people build backwards: a site assembled in ten minutes proves the tool can build sites, never that anyone will pay you for one. AI supplies answers, never demand; demand remains your job.

Edge: what the tool cannot give

Automated bookkeeping shows what the discipline pays. In America, QuickBooks holds most of the market while AI-native newcomers march on it — so the opening is not AI accounting in general but a narrow buyer with troubles current tools cannot fix. The speaker offers his own story: during his consulting years no ready-made program fit his routine, so he brought in an independent bookkeeper who built a custom process around his habits, with AI in service of that process. The advice is to narrow further — independent consultants, corner laundries — ask ten of them what neither QuickBooks, nor the newcomers, nor their current keeper fixes, and build the smallest version on that pain. PitchGrade sizes the fight: Intuit draws 16 billion in yearly revenue, 8.5 of it from the small-business arm, guarded by habit lock-in and data complexity. LedgerBrief completes the table: the online edition of QuickBooks passes six million subscribers, yet AI tools come in two breeds — helpers bolted onto legacy programs and platforms built on learning from day one. Between the incumbent's de facto monopoly and the buyer's hunger, the gap belongs to whoever narrows hardest.

The second gate asks the core question: with rivals holding the same tools, what should you build that resists copying? Since AI itself is never the edge, the sources split three ways: imagination, focus and context. Imagination belongs to Guy, the Canadian street performer who breathed fire and played instruments yet dreamed of a circus — just as show business was dying, children fleeing to theme parks, traditional troupes closing at a loss. Guy married the circus to the Broadway show: elite acrobats, striking light, original music, storytelling, no animals. Out came Cirque du Soleil, credited in the talk with roughly a billion in yearly turnover. Focus comes from the speaker's years as chief executive: three growth paths ahead, two in larger and seductive markets, then the reverse decision to narrow further — one market, one problem, one product, one buyer type. That stubbornness paid off in a handsome sale within a few years.

The third source of edge is context, and here the floor goes to Andrew Ng. MIT renders his thesis: algorithms are widespread, some open to all; the hour belongs to engineering the needed data in a systematic way. Your knowledge, relationships, conversations, mistakes, workflows and subtleties exist in no tool. IEEE fills the portrait: pioneering GPU training with his Stanford students, co-founding Google Brain, years at the scientific helm of Baidu, Ng now runs visual inspection on factory floors through his company Landing AI. And at twenty-six, with no decade of field practice? The talk reassures the young founder: context builds fast by narrowing and joining those who live the problem. Passion for photography gives the example: serving online brands, workflows for estate agents, packages for wedding photographers — three separate troubles. Whoever picks wedding shooters and keeps asking where the workflow breaks, where money and time leak, what clients value, what the off-season brings, builds an insider view nobody else owns. AI researches the niche, maps the flow, ships a rough version in a day; what counts as good stays your call.

Access and loops: first the road, then the mind

The third gate is access: every product must marry the road where it meets its buyer. Hence the Dollar Shave Club story. Selling blades to men with the Gillette giant across the shelf made supermarket war pointless, so the company found the direct road — subscription boxes landing on schedule. Five years after launch, Unilever bought the firm for a billion. Sometimes the buyer picks the channel: software for laundries may live on online maps, cold calls and door-knocking. The interview simulator drives the point home: studying the resume, the posting and the company culture, then bending questions to answers, is the easy half; hunting job seekers one by one is impossible. Distribution runs through campus career offices, coding boot camps and hiring advisers, since users and payers differ. Foundra seals it with Thiel's line: markets punish missing distribution before faulty product; a mid product that owns its channel beats a fine one without. The templates of Notion, the tireless support engineer of Linear, the plugin crowd of Figma show one rule with three faces — an early classic of the subscription economy.

The fourth gate turns usage into assets: each customer touch leaves data, and each lap should teach the next. The image comes from Formula 1: drivers run endless practice laps and pit; engineers compare readings from hundreds of sensors against the driver's feel, tune small things, send the car out. One aim only: every lap makes the next smarter. A young firm repeats the ritual at its scale: sales calls, support logs, reviews and interviews hold insights that go into ready models, surface patterns no eye would catch, and flow back into the work. Nvidia gives the mechanism its technical name, the data flywheel: a cycle starting from company data through customisation, evaluation, guardrails and release, with models refreshed each turn, results improving, costs held down. The rule is plain: the loop must make the business leaner with every pass. When one lab's lead melts into the next, learning speed is the only capital in a speed economy.

The craftsperson at the bench

The close turns from the business to its builder. In Cremona, Italy, violin-making counts as a sacred craft; masters release three to six violins a year, choosing each spruce and maple piece by its voice, carving for months on a single instrument. Nobody else hears what they hear in the wood; years of struggle, listening and refining built the ear. The verdict follows: struggle teaches what good looks like. Hand the grunt work to AI, never the learning and the judgment; run at machine speed without surrendering the wheel. Whoever cheers a low entry step alone will learn the difference when the bubble bursts and only the workbench remains. Forbes repeats the warning: a business with no entry barrier copies over a weekend. The opening question returns to close: where to build the hard-to-copy thing, so the work lives for years?