AI-Proof: What Machines Can Never Be — Volume II: Application, Part V: Conscious Strategy, Chapter 26: Finding AI-Proof Business Ideas
Part V looks at strategy. After vision, trust, checking, speed and the regulatory map, this sixth chapter of Part V asks what makes a business idea worth building, and what you would be selling when you are not there.
A Bangkok cook’s son wants his father’s chicken rice in five carts. The recipe is written down and the queue is real. Neither shows that strangers want a copy, or who will stand behind it when the cook is not there.
In Brief For readers who are weighing a business built on work that lives in their own hands. Chapter 26 asks what makes a business idea worth building when no idea is safe from machines. A chicken-rice cook in Bangkok and his son test whether the dish can exist without him; a designer in Mumbai asks whether his work can become a business without his hands, and a hotel client asks whose name will stand beside hers. The Mahābhārata’s Nala, known to his wife by his cooking, and the Zhuangzi’s Cook Ding help separate what others can taste of a maker from what the work is for the one who does it.
No business idea is safe from machines, and being hard to copy is not what makes work worth doing; before you build a business on work that lives in your hands, find out whether anyone wants it without you, and who will answer for it when you are not there.
Opening: Half Past Ten on the Soi
The khao man gai is ready at half past ten. It has been ready at half past ten, six days a week, for twenty-two years.
Somchai Rattanapong cooks it on a steel cart on a soi, a side lane, off Charoen Krung Road. The road runs through Bang Rak, a district of Bangkok beside the Chao Phraya river. The cart stands in the forecourt of a shophouse whose owner has rented him the space since 2004. It has no sign. There is no menu, because there is one dish: poached chicken over rice cooked with the chicken’s fat, and a sauce of fermented soybeans, ginger and chilli.1 The recipe was his mother’s, and before her his grandmother’s. The queue begins forming around eleven. By early afternoon the pot is empty. Somchai cleans the steel, covers the cart and goes home.
His son Tawan is twenty-three and in the last year of a business degree at Thammasat University. In March 2026 he showed his father a plan: twelve slides, with market research, projections and a franchise model, which he had built with an AI model in an afternoon. It would put the cart’s chicken rice in five places across Bangkok within two years. The arithmetic was right, and the demand was easy to see. Every afternoon people arrived at the soi after the pot was empty and went away. That disappointment, the slides said, was “wasted revenue”.
Somchai looked at the slides. He did not say no, exactly. What he said, twice, was that the chicken would not taste the same.
Tawan took this for sentiment, or for fear of growth. He was partly right. Somchai believes that the taste is not in the recipe. The recipe is written down; his mother made sure of that. The taste, he thinks, is in his hands and his timing, in what he does differently on wet days, in twenty-two years over the same pot. Under that belief is a fear more exact than fear of growth. He is afraid that if the recipe leaves his hands and the chicken tastes fine, then what he has been doing all these years was not special after all. Just cooking.
The title of this chapter promises “AI-proof” business ideas, and the promise needs an argument before anything else. “AI-proof” can mean safe: an idea that machines cannot reach. No idea is safe in that sense, and no book can honestly sell one. Or it can mean what this book’s title means: what machines will never be. In that sense the only thing that holds by meaning is that a machine is not the one a business belongs to, the one who answers for it. That grants no business anything. It does not make an idea good, wanted or lasting. So read the title as a question.
The question is Tawan’s, though he did not yet put it this way. If no business idea can be made safe from machines, what makes one worth building, and what exactly would you be selling when you are not there?
The print edition answered with a programme. A business idea was “AI-proof” if it needed “something that only consciousness can provide”; if it sat in three “layers” that machines could not take over; and if its plan was “moat-first”, starting from its defences. This edition withdraws the programme. Every founder is conscious, including those whose ideas fail, so consciousness cannot tell a sound idea from an unsound one. The layers were descriptions of where machines were slow in a given year. And a plan that begins with defence begins with what keeps other people out.
What this chapter holds is narrower. Whether a business deserves to exist is judged as a product is. The test is what it does to the people it serves, to those who pay for it, and to those who pay without being asked. To that, the chapter adds three things. An idea located by what machines cannot yet do is located by a date. A demonstration that a thing can be made is not evidence that anyone wants it, and a queue for the original is not demand for a copy. And work that lives in one person’s hands becomes a business only by becoming a different promise, whose parts travel differently: the tasks, the eye and the answering.
26.1 Ideas Built on a Gap
Nobody can count how many AI business ideas fail before they start, so nobody can say that most do. What can be said is narrower, and it concerns one kind of idea.
Chidi Okonkwo, a founder in Lagos, built a company in 2024 that wrapped another firm’s language model in a clean screen for small businesses. Users paid for convenience they could soon get free. He closed the company in December 2024. A month later he asked a machine to build it again, and by the end of the afternoon he had a rough copy that worked. Chapter 14 tells his story and the test his company failed. Can a competitor rebuild your product in a fortnight, using the same model? A yes is a warning, not a verdict. It shows that the product you can see is a thin layer over someone else’s machine. Whatever business there is lies in what the copy would lack.
A product that turns legal documents into plain summaries can be a good product in a temporary gap. That names the kind of idea this section is about. The gap is what users cannot yet do for themselves. When the machines close it, the product has no other reason to exist, unless it has earned one.
So the first thing this chapter can say is plain. An idea located by what machines cannot yet do is located by a date. A list of things machines cannot do is a list of dates, and a business built on one of them lives exactly as long as its date. That is not a reason never to build in a gap. Some good businesses began in one, and outlived it by earning something that did not depend on it: customers who came back for other reasons, a name, a way of working. The date is a reason to know which kind of idea you have.
A better question than what machines cannot yet do can be asked of any idea, and it needs no forecast. Would this still be worth doing if a machine, or anyone else, could do every task in it?
Try it on the cart. Every task in it can already be done by others. Khao man gai is sold across Bangkok, and every cook who sells it has a recipe. The cart’s competitor was never a machine. It is the cart on the next soi. Yet the cart is plainly worth running. Several hundred people eat a good lunch there each day, and the pot is empty by early afternoon. Its reason to exist never depended on a gap. The test says nothing about whether the cart is safe. It is not safe. Its landlord could sell the shophouse; the city could change its rules; its cook could fall ill. The test says only that its reason does not carry a date.
26.2 Arjun’s Question
This chapter has no filter for AI-proof ideas. It has a question, and one designer’s evening produced it.
Arjun Mehta is a brand and packaging designer in Mumbai. His studio is two rooms above a printing shop in Parel, and he lives in Bandra. By the start of 2026 he had more work than two rooms could hold, and he was turning clients away. He had also just learned what working alone costs. In the last months of 2025 a box he had designed turned out to resemble one registered in China. In December he revised his client’s packaging without charge and lost three weeks he could not bill. No toolmaker’s promise to defend its customers covered him. A studio, he thought, could carry what one man could not.
One evening in January 2026 he sat on the floor of his flat with the back of a printer’s proof and a pen. An autorickshaw idled at the junction outside. He could hear his neighbour’s television through the wall. He wanted to know whether the studio was an idea worth building, and he wrote questions.
The print edition gave him five, and four of them do not survive. Two asked whether the business needed “something that only consciousness can provide” and “the founder’s specific consciousness”. Every designer Arjun might hire is conscious, and every founder is someone in particular, so neither question can tell one idea from another. Two asked for promises: that the value would compound, and that the business would “get stronger when AI gets better”. Nobody can make either promise. The honest form of the second is smaller: would the studio’s reason to exist survive the machines getting better? The first question came with a worse claim: “If the idea works just as well with a sufficiently advanced AI at the centre, it will eventually have one there, and the human will be unnecessary.” That runs together the one who does a task and the one who must explain it and put it right. A machine may do every task in a project, and someone must still stand behind the result.
One question, the fourth, stayed on the page. Arjun was asking whether the work could become a business that existed independently of his hands.
This book already has a set of questions for whether a product will last. It does not need a second. Arjun’s evening adds one that those do not ask, because they assume a product that already exists apart from its maker. His work did not. Clients came to him for his eye. It had found the dusty rose of a label for a rose sharbat, taken from a sharbat maker on Mohammed Ali Road. It made him spend a day in a client’s world before he drew anything. Could that become something that worked when he was not in the room?
Put that way, the question is too large to answer, so split it. A design practice is three things that travel differently. There are the tasks: layouts, versions, production files, research. They can be written down, and a machine or another designer can do them. There is the eye: what he notices, what he leaves out, when he argues with a client. It is not a skill that can be handed over in a document. If a machine comes to do some of what the eye does, that part joins the tasks. The split does not depend on where machines stop. And there is the answering: whose name is on the work, who explains it when it fails, and who puts it right. A studio would have to do something different with each.
He looked at the fourth question for a long time.
He wrote: I don’t know.
26.3 Three Layers, Each With a Date
The print edition named three layers of opportunity: checking whether things are true, being physically present, and trust. It said that machines could not take these layers over, and that the best businesses sit “where two or all three layers overlap”. Each layer describes where machines were slow in a given year. None of them describes where a business deserves to exist, or where anyone will pay for what it sells.
Take checking first. As of 2026, machines check a text against other texts quickly and well. Checking a claim against the world, by a visit, a test or a telephone call, stays slow and costly, and Chapter 23 asks who pays for it. A business built on that difference holds while the difference holds.
Presence next. The same section admitted that the timeline for general-purpose robots “may shorten”, and told founders to “watch it honestly rather than assume permanence”. That was the right warning, and it dates the whole layer. Presence also protects less than it seems. Nobody can copy Somchai’s cart by copying a file. Anyone can copy its recipe, and a franchise is a way of putting one recipe into five other people’s hands. Presence protects the cart from software. It does not protect it from another cook.
Trust last. The section closed this layer with a sentence this edition withdraws: “The layer beneath the system is always a self.” Trust is not a layer that anyone owns. It is kept and spent, by how a business behaves when it is wrong, as Chapter 22 shows.
By that map, Somchai’s cart is the safest business in the chapter. He checks every pot himself, he stands there every day, and his regulars trust him. It sits in all three layers at once. Its competitor is still the cart on the next soi, and its future is still decided partly by a landlord and a city. A layer is a description of where machines are slow. A business placed by that description inherits its date.
26.4 The Inventory Before the Plan
The print edition advised a moat-first business plan, which reverses the usual order of planning. Do not begin with a problem and then look for a defence. Begin with the defence, the moat, and then ask which problems it lets you solve better than anyone. It added: “Eight moats survive the AI era. Every business that lasts will be built on at least one.” This book has already tested those eight moats one by one. None turned out to be a wall. Some were duties, done again each time, and some were reasons to stay that do not depend on a business being good.
Tawan’s deck is a plan of the usual kind. Its problem is the people who arrive after the pot is empty. Its solution is five more pots. The inversion would have him start elsewhere, with what the cart has that a rival cannot quickly get. That sounds wiser, and part of it is. But a plan that starts from defence starts from what keeps other people out, and then looks for customers it can hold. Holding customers who would rather leave is lock-in. It is a reason a business survives, not a reason it should exist.
The useful part of the inversion is the inventory. Before any plan, list what you actually have. Tawan’s list is short. A recipe, written down, for a dish sold all over the city. A queue, for this cart, on this soi. His father’s mornings, twenty-two years of them. A forecourt rented from a family that has known his father for as long. And the regulars, some of whom have eaten at the cart since it opened. What the list does not show is what the cart has never had to pay. Its rent was agreed between old acquaintances and never written down, and its cook is its owner, who takes what is left. A new cart would pay a lease, and a wage or a franchise fee, before anyone took what was left. Whether a plate still pays after that is a question the old cart has never had to answer.
Now ask which of these would travel to five new places. Only the recipe would. It is also the one thing every rival already has in some form.
You have an idea. And ideas, in 2026, are the cheapest commodity in the economy. That claim carries a date, and it is true for its year. An idea cannot promise a business that lasts. Nor can years spent on expertise, a network and a reputation. Depth can fill a place that someone is willing to give you. It does not make anyone give it.
So the inventory asks one more thing before the plan. What would the five carts be for, and for whom? For the people who arrive late, the deck says. But those people already have other carts to go to, and most of them go. The five carts would be for something else. They would be for Tawan’s future, which is a good reason. And they would be for a name that does not yet exist, the brand the franchise model needs. A brand is a promise to strangers. Whoever paints it on five carts must say who will answer for it. There is also a cost that no slide counts. Suppose the name on the carts is Somchai’s, and one of the five serves a bad lunch. The cart on the soi pays, in the trust of people who never ate at that one. That is one way a business makes people pay without asking them. Lend your name to work you do not see, and people you will never meet can spend it.
26.5 Validating Ideas in the Age of Instant Prototypes
Tawan’s twelve slides took an afternoon. As of 2026 a working prototype of many kinds of software takes days, and a plan with projections takes hours. The trap is that cheap building creates the illusion of cheap validation. A prototype tests whether something can be built. Building is cheap. Attention is not.
The print edition added a sentence that needs a correction: “Validation tests whether it should be.” That runs two questions together. Whether people want a thing is one. Whether it should exist is another, and buyers settle only the first. Kept apart, the first can be said plainly. A working demonstration shows that a thing can be made. It is not evidence that anyone wants it, and still less that it should exist. And a thing people want may still not pay its way. They can want it at a price that does not cover what it costs to make, deliver and put right.
The Pañcatantra, an old Indian book of fables, tells of four friends who found the bones of a lion in a forest. Three of them were learned. They put the skeleton together and gave it flesh. As they were about to wake it, the fourth, who had little learning and much sense, warned them, and they laughed at him. He climbed a tree. They woke the lion, and it killed the three.2 The fable is about learning without sense, and I borrow only its picture. Their prototype worked. The one person who asked what would follow was not heard.
Evidence of wanting is slow because it needs people who owe you nothing. Machines can make it slower in one way. Asked to judge a plan, a model built to please will tend to find it promising. A study published in Science in 2026 found eleven widely used models, asked for advice about people’s own conduct, affirming their actions about 49 per cent more often than people did.3 Asked instead to attack the plan, the same model will attack it. Both, as I read them, are obedience to the role it was given, and neither is a test of the market. A founder who asks a flattering machine to validate an idea is using a sophisticated mirror to admire his own reflection.
A test of an idea works only for someone who is willing to hear no. The cheapest test is the one that could show your favourite assumption false, run on people with no reason to please you.
Arjun did the cheap thing first. Between February and March 2026 he talked to the clients of his last two years. He asked each the same question: what did you get from working with me that you could not have got from a machine or a larger agency? The replies were kind. Several mentioned the day he spent in their world before he drew anything. One said, “You argued with me.”
Then he did what he had done with his own designs since his second design for a hotel chain in Goa, after it rejected his first website in October 2024. He wrote the case against his plan in the voices of the people who would live with it, and this time he asked a model to argue it with him. Its first objection was not about design. Every person he had asked, it said, had already chosen him. He had asked the people who stayed whether they wanted him. He had not asked anyone whether they would buy a studio.
The machine was right. He called two clients who had left. One had moved to a larger agency, for its size. The other now used design tools and a junior designer of her own. “I did not need you,” she said, without unkindness. “I needed it done.” She had not found a better designer. She had stopped needing one from outside. Who had renewed, and who had left, said more.
The most useful reply came from a client who had stayed. Vandana is the operations director of a hotel chain in Goa. In October 2024 she rejected the first website Arjun designed for the chain. In April 2025 the chain renewed his contract for a year, with her name beside his on every approval. “I don’t know if what you do is worth what you charge,” she told him. “But I know the AI version isn’t worth anything to me, because when it’s wrong I can’t tell.” She was not praising his eye. She was saying what she could not check. People believe what they cannot check as far as someone stands behind it, and she knew who stood behind his work. Whether she would believe a studio, none of his conversations had asked.
Tawan ran the other kind of test, the one his father’s claim invited. In June 2026, after his last examinations, he asked whether he could film the mornings at the cart. Somchai agreed, on the condition that the film stayed in the family. Tawan gave it to a model and asked it to write out every step. The model ran on a company’s computers, so the film went there too. Neither of them had thought of that as leaving the family. Some of the steps were in the recipe. Some were not. The chicken was lowered into the pot and lifted out three times before it was left. The stock was tasted twice. The flame was turned down when the rice began to tick against the lid. Somchai read the list and said that he did not lift the chicken three times. The film showed that he did.
For two weeks Tawan cooked a second pot beside his father’s, from his grandmother’s recipe and the machine’s list. Regulars who agreed were given a small plate from each pot, unmarked, and asked which was Somchai’s. Most could not tell. On most days their answers were no better than guessing. One woman, who had eaten at the cart since it opened, was certain every time. She tried on six days and chose wrongly on four. On the wettest day of the fortnight Tawan’s rice came out soft and heavy, and everyone could tell. The list said nothing about rain. Asked, Somchai said that he used a little less water when it rained. He had not known that he did it until he was asked.
The test had a price, and it was paid by the people who ran it. Tawan gave it two weeks of mornings. Somchai paid for the second pot of chickens, and some of it was thrown away. The regulars gave their patience, and one of them gave up being certain. Tawan ran it knowing that it might prove his father right.
So, in this test, Somchai was mostly wrong. To most of the people who knew it best, on most days, the chicken tasted the same. On the one hard day he was right, and even there what he knew was partly a record nobody had kept. A note of the weather beside each day’s pot would, in time, have taught it too.
Two old stories about cooks look at this test from different sides.
In the Mahābhārata, the great Sanskrit epic, King Nala loses his kingdom at dice and leaves his wife Damayantī in a forest. He lives in disguise as Bāhuka, a charioteer and cook in another king’s household. Later Damayantī suspects who he is, and sends her maid to bring her some meat he has cooked. In Kisari Mohan Ganguli’s translation4, “Damayanti who had formerly often partaken of meat dressed by Nala, tasted the meat that was brought by her hand-maid.” Then “she thereupon decided Vahuka to be Nala and wept aloud in grief of heart.” The Sanskrit calls her ucitā, accustomed: she knew the cook because she had eaten his food many times before.5 The epic gives his skill an unexpected source. When Damayantī chose Nala as her husband, the gods gave him gifts, and “Yama granted him subtle taste in food as well as pre-eminence in virtue.”6 The taste was a god’s boon, not years at a pot. The story is about a king’s disguise and a wife’s grief, not about shops. I borrow only its answer to one question: who can recognise a cook in his food? Someone accustomed to it, and perhaps only such a person.
The Zhuangzi is a Chinese book whose early chapters are traditionally dated to the late fourth century BCE.7 It tells of Cook Ding, who carves an ox so smoothly that his lord praises his skill. In Burton Watson’s translation8, which spells the name Ting, the cook replies: “What I care about is the Way, which goes beyond skill.” His knife is nineteen years old and still sharp, because he cuts along the spaces in the ox. But “whenever I come to a complicated place, I size up the difficulties, tell myself to watch out and be careful, keep my eyes on what I’m doing, work very slowly”. The lord’s reply ends the story: “I have heard the words of Cook Ting and learned how to care for life!” The passage is about how to live, not about meat. I borrow its order of attention: the master is slowest where the work is hardest. Somchai’s wet days were such a place, though, unlike Cook Ding, he had not known that he changed anything there.
The two stories seem to disagree, and they answer different questions. Nala’s story puts the cook in the food, but only for someone accustomed to it. Recognition there is a relation between a cook and the people who have eaten his food, not a property of the dish. And Tawan’s test adds a warning the epic does not give: the most accustomed person at the cart was wrong more often than right. The Zhuangzi puts the good of the skill first in the life of the one who has it. The lord learns nothing about carving. He learns how to live.
Somchai’s fear belongs to the first question. If nobody can taste the difference, he thinks, it was just cooking. The second question says that his twenty-two years are not settled by a taste test at all. A person’s worth never depended on being hard to replace, and neither did the worth of his work. Several hundred good lunches a day are not nothing because someone else could cook them. Neither story sells a franchise. Neither says that a machine could not cook the chicken.
What did Tawan’s test show? It showed that the chicken can taste the same, at his father’s cart, to people who came for his father. It showed nothing about strangers in five districts, at a cart with no history and a name they had never heard. It also showed only that Tawan could cook it, on dry days, beside his father. He had grown up beside the pot. Five new carts would need other cooks, at least four of them, who had not. Arjun’s conversations had the same limit: the people he asked wanted him, which is not the same as wanting a studio. The queue at the soi is evidence that people want this cart. A queue for the original is not demand for a copy. The test the plan needs is a copy among strangers, and it costs a season.
Plot twist
A franchise looks like the safe way to start a business: a tested recipe, a known name and help from the firm that owns it. An older American study found the opposite for the people who buy in. Timothy Bates, an economist, followed small firms started between 1984 and 1987 until late 1991, in research done at the US Census Bureau. The franchised firms began larger and better funded. Even so, the independent firms earned more, and “their survival prospects are better than those of franchises”.8
The firms that sell the name can fare differently. A study of new franchisors, published in 1996, found that a young chain’s emphasis on franchising had “a significant, positive effect on its growth and survival”.8
Both studies are decades old and American, and neither looked at street food in Bangkok. What they change here is a question the slides left out. On my reading, a franchise can be good for the owner of the name and bad for the people who run the carts under it. If five carts failed to draw strangers, the people who ran them would carry most of that season.
26.6 A Map With Dates
The print edition drew a global opportunity map. It sorted the world by region. Where machines were most used, checking would pay. Where “personal relationships have historically been central to commercial life”, in “significant parts of Africa, Latin America, South and Southeast Asia, and the Middle East”, trust would pay. Where crafts and food were, presence would pay. Then it listed businesses, from a weavers’ cooperative to a taproom. This edition withdraws the map and the list. A region of hundreds of millions of people is not a trust market, and a list of safe trades is the promise this chapter cannot make. Building on networks that fill the gaps left by weak institutions is practical, and the builder should want the institutions to strengthen, not stay weak.
An honest map is drawn smaller, and it carries dates. Part of it is drawn at City Hall.
Between 2014 and 2016, Bangkok’s registered street vendors fell from over 21,000 to 10,700, under a policy whose motto was “Return the footpath to pedestrians”.9 By 2023 the city’s rules required a vendor in a permitted area to trade in person.10 In 2024 the governor signed new rules for selling on public pavements, dated 29 August and made public by the city on 19 September. A vendor must be Thai, and must also show a low income in one of four ways. One is the state welfare card; another is earning no more than 300,000 baht a year after business costs.11 A vendor may not run or help at a second stall,11 and may have one assistant, who must also be Thai.12 In October 2024 the city said that a permit is not for life and cannot be transferred, leased or sold, and that it planned no new vending zones.13 Early in 2025, 618 vendors were permitted to trade in 32 areas of Bang Rak.14 By 2026 the city was cancelling some pitches, adding a few others, and moving vendors off main roads into hawker centres.15
Read that against Tawan’s slides. Somchai’s cart has never stood on a public pavement; it stands on a forecourt he rents. His own institution is a family, the shophouse owners, who have taken his rent for twenty-two years on terms nobody has written down. If they sell, the cart moves or stops. Under these rules Tawan could not own or run five carts on public pavements. Each pitch is held by one vendor on a low income, who sells in person and cannot pass it on. Any cart he ran would, in practice, have to stand on private land, paid for in rent. The city has reasons of its own: walkways, hygiene, and support for vendors on low incomes.16 They are not this book’s reasons. Even so, the rules say plainly something the slides left out. In Bangkok, a pavement stall is treated as one person’s work, not as a unit of something larger. That keeps a poor vendor’s pitch from becoming a chain’s. It also keeps that vendor from growing on the pavement.
Whether a cart can become a business is decided partly by people who never taste the chicken. That is true of most business ideas, and it is the part of the map that changes most often. An app that brings a cart its customers writes rules too: it can change its fees, or the order in which carts appear. A map of opportunities that leaves out who writes the rules is a map of the founder’s hopes.
26.7 What a Business Can Promise Without You
In April 2026 the hotel chain’s year was up. Arjun had a reason to be careful that month. He had just paid to reprint a drinks company’s labels, because his guide had given their colour only for screens (sRGB), with no value for ink.17 In a studio the next such error might be someone else’s, and the bill and the explanation would still be his.
He told Vandana about the studio before he had decided anything. If he built it, other designers would do part of her work, and he would review it.
“Whose name goes beside mine?” she asked.
He had no answer ready, so they went through the work together. The brand and the website would stay his, and his name would stay beside hers on those approvals. The menus, the signs by the pool and the offers for the monsoon season could be done by anyone he would stand behind. The condition was that he looked at the work before it went, and put it right when it was wrong. She renewed the contract for six months, not a year. “Show me who does what,” she said. He had told her before he knew, and the telling cost him half a year of certainty. He signed.
Her question splits Arjun’s into the three parts it already had, and each travels differently.
The tasks travel easily. Layouts, versions, production files and research can go to machines and to other people. That is what the tools are for. Vandana’s menus were a list of tasks.
The eye travels only by forming people, and forming people is slow and paid for by someone. What Arjun notices is what this book calls a disposition: worn into a person by repeated choices among other people, and kept only while there are occasions to use it. A designer does not acquire it from a document. She acquires it from real work with real stakes, room to fail, and correction from someone who looks. Written examples and reasons help, but the forming still happens in the work. Room to fail means that someone agrees to carry the cost of her mistakes, and in a small studio that someone is Arjun, or his clients. The first projects of a junior designer are her first rung, the place where the next person learns, and they cost hours that a client does not see. That place is hers as well as the studio’s. A studio that hires a junior designer for her hours and gives her only the tasks has used her, not formed her. Forming her costs Arjun too. The hours he would spend looking at her work are hours he would not spend on his own. For a while the studio’s work might be worse than his alone. And once she is formed, she may leave and take the eye with her. The studio will have paid for it all the same. There is a danger the other way too. A studio that hires nobody to do the making, and keeps only those who learned their eye by doing it, uses up the eye it inherited.
The answering must be placed again. Whoever’s name is on the work must answer for it: explain it when it fails, and put it right. A studio can do that as a studio: explain, redo, pay. Vandana’s question was about something else, and the renewal shows it without needing a verdict. For the brand and the website she wanted the person whose first website she had rejected, and whose second design had been aimed at her actual guests. For the menus she wanted someone who would stand behind them. The arrangement has a ceiling. If every piece passes under Arjun’s eye before it goes, the studio can grow only as far as his hours of looking. His name stays honest only while he has them.
The print edition said that for the studio to be real, “the consciousness had to become transmissible”. That asks for the wrong thing. Nothing about Arjun’s awareness needs to pass to anyone. What must pass is a set of habits, formed in other people at someone’s cost. And there must be a name that those people can keep honest when he is not in the room.
So Arjun’s question can be met, though not settled. Can work that lives in one person’s hands become a business that exists without them? It can become a different business. Work that lives in one person’s hands becomes a business only by becoming a different promise. A studio that sells only what Arjun does himself is a practice with overhead. A studio that sells something else must say what it is selling. A sign makes a promise to strangers who have never met the person behind it, and someone has to keep that promise every time.
He did not decide that spring. The page with the question on it went into a drawer, under the proofs.
Ending: The Sign in Huai Khwang
In September 2026 Tawan opened a cart of his own, on a rented forecourt in Huai Khwang, across the city. His father had given him the recipe in the one way he could. Tawan photographed his grandmother’s page in his father’s kitchen, and the page stayed there. His father did not give him his name. “That one stays at my cart,” Somchai said, and said nothing more about it.
Tawan’s cart has a sign, the first the family’s chicken rice has ever had. It says khao man gai, and his own name.
The first morning it rained. He used a little less water, and watched the rice. No queue formed. People came one at a time: office workers, two drivers, a woman with a child. None of them had ever eaten at the cart in Bang Rak. At half past one there was still chicken in the pot.
The woman finished her plate, paid, and said that it was good. Tawan thanked her. He wanted to ask whether it tasted the same, and there was nobody to ask. Nobody there had ever eaten his father’s.
Practice
The morning without you. Write down one ordinary working morning of the work you would build a business on, step by step. Beside each step, mark any of four things that apply. Mark it written down if another person or a machine could do it from the page. Mark it learned if it can only be learned by doing it under correction. Then note whose hours would form it in someone else, and who would carry the cost of their early mistakes. Mark it answered for if a customer would call your name when it goes wrong. Mark it depends on if it needs something you do not control: a landlord, a city’s rules, a supplier, an app, one person’s health. Then write the one sentence a sign would promise a stranger who has never met you, and check it against your marks. Ask what it costs, in money and hours, to keep that promise to one customer, including putting it right when it fails. Finally, choose the assumption in your plan that you most want to be true. Find the cheapest test that could show it false, and run it on people who have no reason to please you; your best clients are not those people. For the self-employed this costs unpaid evenings; an employee proposing a venture inside a company should ask for the time. The exercise does not tell you whether a business will pay. It shows what you would be selling, what keeping that promise costs, and who would answer for it.
Words to keep
- disposition: a way of seeing or acting, worn into a person by repeated choices among other people and kept only while there are occasions to use it; it cannot be handed over in a document.
- franchise: an arrangement in which other people sell a product under one name and to one recipe or method.
- khao man gai: Thai chicken rice: poached chicken over rice cooked with the chicken’s fat, served with a sauce of fermented soybeans, ginger and chilli.
- prototype: a working early version of a thing; it shows that the thing can be made, not that anyone wants it.
- rung: the place in a line of work where the next person learns, usually the simpler work given to beginners.
- soi: in Thai cities, a side lane off a main road.
- ucitā: Sanskrit, “accustomed”; the word the Mahābhārata uses for Damayantī, who knew Nala by his cooking because she had often eaten it.
Sources
- MICHELIN Guide Thailand, “Iconic Dishes: Khao Man Kai, Thai Chicken Rice and Where to Find It,” 14 November 2024 (guide.michelin.com): a “Hainese-inspired [sic] Thai dish” of poached chicken and rice; the dipping sauce of “fermented soybeans (taucheo), vinegar, ginger, sugar, dark soy sauce, and green chillies”. The rice: the dish’s Thai name means “chicken oily rice”, and in the dish generally “some of the fat and liquid ... are used in the cooking of the rice” (“Hainanese chicken rice,” Wikipedia, accessed 23 October 2026). The cart, Somchai and Tawan are composites.
- Pañcatantra V, “The Lion-Makers”; Arthur W. Ryder, trans., The Panchatantra (Chicago: University of Chicago Press, 1925), Book V, “Ill-considered Action”, pp. 442–444. Ryder leaves the fourth man unnamed; in the Sanskrit text that Chapter 11 follows he is Subuddhi, “Good Sense”. Chapter 11 tells the story in full. Its use here as a picture of a prototype is mine.
- Myra Cheng, Cinoo Lee, Pranav Khadpe, Sunny Yu, Dyllan Han and Dan Jurafsky, “Sycophantic AI decreases prosocial intentions and promotes dependence,” Science 391, no. 6792 (March 2026): eleven models; users’ actions affirmed about 49 per cent more often than by humans (a relative figure; the prompts asked for advice about the users’ own conduct). DOI 10.1126/science.aec8352.
- The Mahabharata of Krishna-Dwaipayana Vyasa, Book 3, Vana Parva, Nalopakhyana Parva, Section LXXV, trans. Kisari Mohan Ganguli (Calcutta, 1883–1896; SriPedia mirror of the Internet Sacred Text Archive, ibiblio.org).
- Mahābhārata 3.73.22 (Critical Edition numbering), socitā nalasiddhasya māṃsasya bahuśaḥ purā / prāśya matvā nalaṃ sūdaṃ prākrośad bhṛśaduḥkhitā: “she, accustomed to meat prepared by Nala many times before, ate it, took the cook to be Nala, and cried out in great grief” (my rendering; text as in the electronic Critical Edition at sacred-texts.com, The Mahabharata in Sanskrit, Book 3, and at enjoylearningsanskrit.com, Book 3, chapter 73, verse 22). Socitā is sā ucitā.
- The Mahabharata, Book 3, Vana Parva, Nalopakhyana Parva, Section LVII, trans. Kisari Mohan Ganguli (mahabharataonline.com; also sacred-texts.com, m03057). The gods give the boons “after the daughter of Bhima had chosen Naishadha as her husband”; the wedding follows after they leave. In the Critical Edition the boon is 3.54.31: yamas tv annarasaṃ prādād dharme ca paramāṃ sthitim.
- Chad Hansen, “Zhuangzi,” Stanford Encyclopedia of Philosophy (substantive revision 24 March 2024): Zhuangzi “flourished through the latter half of the fourth century BC”; the Inner Chapters (1 to 7) are attributed to his period, and his authorship of any of them “remains a hypothesis”.
- Zhuangzi, chapter 3, “The Secret of Caring for Life” (養生主), the story of Cook Ding (庖丁), trans. Burton Watson, Chuang Tzu: Basic Writings (New York: Columbia University Press, 1964), as reproduced by the Bureau of Public Secrets (bopsecrets.org). Watson’s sentence about the complicated place is quoted only as far as “work very slowly”; it goes on to the cut itself. Chapter 3 of this book quotes Watson’s Complete Works (1968) for chapter 2. For the Plot twist in 26.5 (added in this edition): Timothy Bates, “A Comparison of Franchise and Independent Small Business Survival Rates,” Small Business Economics 7, no. 5 (1995): 377–388, abstract (firms started 1984 to 1987, followed to late 1991; franchises “larger scale, better capitalized”; independents more profitable, and “their survival prospects are better than those of franchises”; the research was done at the US Census Bureau’s Center for Economic Studies, Working Paper 94-3, 1994); Scott A. Shane, “Hybrid Organizational Arrangements and Their Implications for Firm Growth and Survival: A Study of New Franchisors,” Academy of Management Journal 39, no. 1 (1996): 216–234, abstract (a new franchisor’s emphasis on franchising “has a significant, positive effect on its growth and survival”).
- Gisele Yasmeen and Narumol Nirathron, Vending in Public Space: The Case of Bangkok, WIEGO Policy Brief (Urban Policies) No. 16 (1 May 2014), with an update by Sarah Reed, 26 April 2017 (wiego.org): registered (licensed) vendors fell from over 21,000 in 2014 to 10,700 in 2016, citing the Bangkok Metropolitan Statistics Book 2016; the BMA’s motto “Return the footpath to pedestrians”.
- “กทม.คืนฟุตปาทให้คนเมือง คุมกำเนิด 21,368 หาบเร่แผงลอย” (BMA returns the footpath to city people, caps street vendors at 21,368), Prachachat Turakij, 20 February 2023: among the conditions for vendors in designated areas, “ทำการค้าด้วยตนเอง” (to trade in person).
- “ประกาศเกณฑ์ใหม่ ‘หาบเร่ฯ-แผงลอย’ ผู้ค้าเป็นคนไทย มีบัตรคนจน-รายได้ไม่เกิน3แสนต่อปี,” Daily News (Bangkok), 19 September 2024, and “เปิดหลักเกณฑ์ เงื่อนไข หาบเร่-แผงลอย หลัง กทม. ประกาศกฎใหม่,” Krungthep Turakij, 22 September 2024. The governor had signed the BMA announcement on trading areas on roads and public places, made public on 19 September 2024. Vendors and their assistants must be Thai and meet any one (ข้อใดข้อหนึ่ง) of four conditions: the State Welfare Card; a home-purchase contract with the National Housing Authority under the Baan Mankong programme; welfare payments from the Ministry of Social Development and Human Security; or income of no more than 300,000 baht a year after business costs, as shown in tax filings. They may not hold or assist at another stall. Trading areas are reviewed every two years on roads of three or more lanes and every year on smaller roads. Pattaya Mail (TNA), 4 October 2024, dates the announcement 29 August 2024; Daily News, 5 October 2024 and 7 April 2026, gives the criteria as applying from 29 August 2024.
- “Only ‘poor Thais’ can be street vendors, says City Hall,” Bangkok Post (undated on the page; it reports the announcement of September 2024): one sales assistant, who must be Thai; no migrant workers; stalls of three square metres; the paper said the rules would take effect on publication in the Royal Gazette (no Gazette notice was found; Daily News gives 29 August 2024 as the date they applied). An earlier draft, open for public comment from 17 to 31 January 2024, had a ceiling of 180,000 baht. It made the permit personal and barred its sale, lease or transfer, and anyone else trading in the vendor’s place (Thansettakij, 22 January 2024; Thai PBS World reported the same draft: “The licenses cannot be transferred or sold”).
- “Bangkok reduces over 10,000 street vendors in two years, pushes for new trading space regulations,” Pattaya Mail (from TNA), 4 October 2024: Governor Chadchart Sittipunt, “In the past two years, we’ve reduced the number of street vendors by over 10,000”; permits “will not be granted for life”, and “vendor permits are non-transferable and cannot be leased or sold to others”; “The BMA does not plan to expand street vending areas”; review of existing zones from January to March 2025, results in June 2025.
- “BMA ends vending in 2 key locations,” Bangkok Post, January 2025 (dated from its text: vending ended “since Jan 1”, and a Silom stretch cancelled “in October last year”): “In Bang Rak district, there are currently 32 areas where 618 street vendors are permitted to operate.” The deputy governor, Jakkapan Phiewngam, asked that vendors in permitted areas meet standards “in line with the law on public orderliness and hygiene”.
- “เดินหน้าจัดระเบียบพื้นที่ทำการค้า ชูมาตรฐานความปลอดภัย-สุขอนามัย,” Daily News, 7 April 2026 (the criteria in force from 29 August 2024; 36 vending points outside the permitted areas cancelled; a target of 30 replacement points; 47 areas for local identity and tourism approved by the committee and awaiting the governor); “ราชกิจจาฯ ประกาศ กทม. เพิ่ม 4 จุดผ่อนผันค้าขาย ‘ยานนาวา-ภาษีเจริญ’,” Krungthep Turakij, 3 July 2026 (four new permitted points, three in Yan Nawa and one in Phasi Charoen, published in the Royal Gazette on 3 July 2026); AFP, “Bangkok food vendor curbs push city staple from the streets,” The Jakarta Post, 3 May 2026 (five city hawker centres, the latest opened in April 2026 beside Lumphini Park, for vendors moved off main roads). The Daily News headline’s en dash is given as a hyphen.
- The Thaiger, report on the BMA’s new vending rules (headline as in its address, “Bangkok street vendors must be poor enough to operate, BMA”; undated on the page; it reports the same announcement): in the paper’s own words, the rules “aim to support low-income Thais while improving public spaces”; income counted after deducting business costs. On hygiene, see the Bangkok Post report of January 2025 cited above (the deputy governor’s instruction that vendors in permitted areas meet standards “in line with the law on public orderliness and hygiene”); on walkways, the WIEGO brief.
- Michael Stokes, Matthew Anderson, Srinivasan Chandrasekar and Ricardo Motta (Hewlett-Packard and Microsoft), “A Standard Default Color Space for the Internet - sRGB,” version 1.10, 5 November 1996 (w3.org, marked obsolete and kept for historical purposes): an RGB colour space “well suited to” monitors, television, scanners, digital cameras and printing systems, proposed as a default for the Internet; it has since been standardised as IEC 61966-2-1. The printing systems meant are those that accept RGB files; an offset press prints in CMYK inks, so an sRGB colour has to be converted. The drinks company’s reprint, and its date (April 2026), are told in Chapter 24.
Named people in this chapter are composites, and their conversations and circumstances are constructed; the studies, laws and figures cited are real and sourced.
Revised from the print edition (2026)
- What the print edition claimed: that an “AI-proof” idea requires “something that only consciousness can provide”, such as “taste that requires a self behind the choosing”, and that “If the idea works just as well with a sufficiently advanced AI at the centre, it will eventually have one there, and the human will be unnecessary”; that Arjun’s practice passed because his qualities “required someone home”; that an AI-proof business accumulates “a growing body of knowledge that no competitor can shortcut”; that a studio using machines “would become more valuable as AI tools improved, not less. This one passed.”; that the studio’s problem was “the ability to transmit his consciousness to other people”; that “trust cannot be automated and verification cannot be outsourced”; that “The layer beneath the system is always a self”; that the overlaps of the three layers are “where competition is thinnest and where AI’s reach is shortest”; that “Eight moats survive the AI era. Every business that lasts will be built on at least one”; that “No general-purpose model will maintain current, accurate, jurisdiction-specific knowledge”; that a moat built over years gives a plan “a longer life expectancy”; that “Validation tests whether it should be”, measured by “real willingness to pay”; that an AI asked to evaluate a plan “will, in most cases, tell you it is promising”, because of “the probabilistic nature of language models”; that wanting to be proven wrong is “the rarest quality in entrepreneurship” and “the most reliable signal that a founder’s consciousness is actually switched on”; that the client needed “Not a system. A person. ... Someone conscious.”; that regions with strong personal ties are a “trust layer”, and that their businesses “run on consciousness, not computation”; that the three layers “have proven durable across every wave of technological change”; that “transmitting consciousness is not the same as transmitting skill”, and that “The machine does not tolerate uncertainty. It generates the most probable next token.” It also told the commodity trap again through a founder in São Paulo, Lucas, whose summaries of legal documents launched in early 2025 and were overtaken “by August” (free assistants had read Portuguese well before 2025); built 26.4 on a Johannesburg consultant, Naledi Dlamini, whose liquor licensing it called municipal (it is provincial) and whose regulations changed after a Braamfontein nightclub fire and a Sandton incident of which I found no record; listed safe trades (physiotherapists, coaches, a weavers’ cooperative in Oaxaca, a coffee farmer in “the Sidama zone”, a regional state since 2020, a practitioner in Kerala, a taproom in Guadalajara, a food inspector in Lagos); gave Arjun a forty-minute call to Rajan Subramanian, the Coimbatore textile manufacturer of other chapters, who called his trust “not mine anymore” and said “It belongs to the business”, with “twenty-seven years”, “a mill floor in Erode” and a knee operation that the rest of the book does not hold; called Chapter 24’s colour-profile error “the Pantone error”; gave Arjun “twelve years in Mumbai’s visual culture” and a named brand of notebook; said that Kavitha Sharma, the science teacher in Sagar, had “shifted from delivering content to verifying”; ended with the son watching his father’s hands and changing a slide to “recipe scalability: unknown”; and closed with a bridge to the next chapter. It also asked “What problems will AI not solve next year, or in five years, or ... ever?”, and promised that “a year accumulating domain expertise, two years building a network, three years establishing a reputation” would lay the foundation for a business that lasts.
- What I now hold: “AI-proof” read as safe is a promise no idea can keep; read as this book’s title reads it, it grants no business anything. Whether a business deserves to exist is judged by what it does to the people it serves, to those who pay for it and to those who pay without being asked. An idea located by what machines cannot yet do is located by a date, though a business begun in a gap can outlive it by earning something that does not depend on it; the better question is whether an idea would still be worth doing if a machine, or anyone, could do every task in it. The three layers are dated descriptions of where machines were slow, not of where a business deserves to exist or where anyone will pay for what it sells; presence protects a cart from software, not from another cook. A plan that starts from defence starts from what keeps people out; the useful part is an honest inventory of what you have, which of it travels, and what the business has never had to pay (rent agreed between old acquaintances, a cook who is also the owner). A demonstration shows that a thing can be made; it is not evidence that anyone wants it, still less that it should exist, and a thing people want may still not pay its way; flattery and attack from a machine are, as I read them, the same obedience to the role it was given; a test works only for someone willing to hear no, and what clients do says more than what they say. A queue for the original is not demand for a copy, and a dish cooked on dry days by the son who grew up beside the pot says nothing yet about other cooks. A film given to a model that a company runs goes to that company. Recognition of a maker in his work is a relation, held by those accustomed to it, and even they can be wrong; the good of a skill to the one who has it is not settled by a taste test, and work is not worthless because someone else could do it. Who writes the rules, here a city’s rules for pavements, decides much of what an idea can become; Bangkok’s rule keeps a poor vendor’s pitch from becoming a chain’s and keeps that vendor from growing on the pavement, and an app that brings customers writes rules too. Work that lives in one person’s hands becomes a business only by becoming a different promise: the tasks travel easily; the eye travels only by forming people, at someone’s cost, helped but not replaced by written examples; the answering must be placed again in a name that others can keep honest, and a founder who reviews every piece caps the studio at his hours of looking. Whatever part of the eye a machine comes to do joins the tasks, so the split does not depend on where machines stop; and a designer who has been formed may leave with what the studio paid to form.
- What changed my mind: Chapter 14’s finding that surviving is not deserving to exist; the earlier test of the eight moats; the book’s rule that claims about what machines can do carry a date; the audit’s corrections (the colour profile, the provincial licensing, the Sidama region, the invented incidents, Rajan’s years and mills); the choice to give Rajan less room in this Part, after he had carried two chapters of the last; Bangkok’s own rules for pavement vending, which treat a stall as one person’s work; the plain fact that, as of 2026, a plan with projections takes a machine an afternoon; and an outside reading of an earlier version of this chapter, in October 2026, which pointed out that the chapter kept wanting apart from deserving but not from paying its way, that a reproduced dish is not a reproduced business, that review by one person caps a studio, that the Bangkok rule cuts both ways, that apps write rules too, that a business begun in a gap can outlive it, and that the colour note rested on Wikipedia.
- What survives: the cart, “ready at half past ten”, with no sign; Somchai’s “twice”; the fear, word for word: “that if the recipe leaves his hands and the chicken tastes fine, then what he has been doing all these years was not special after all. Just cooking.”; Tawan’s slides and “wasted revenue”; “a good product in a temporary gap”; Arjun’s question, “whether the work could become a business that existed independently of his hands”, and “He wrote: I don’t know.”; “cheap building creates the illusion of cheap validation”; “A prototype tests whether something can be built”; “Building is cheap. Attention is not.”; the sophisticated mirror; the client’s “when it’s wrong I can’t tell”, now Vandana’s; “You argued with me.”; “ideas, in 2026, are the cheapest commodity in the economy”; the builder who “should want the institutions to strengthen”; “a practice with overhead”; and Tawan’s question, now the chapter’s: what exactly would you be selling?
- Section titles in the print edition: 26.1 was “Why Most AI Business Ideas Fail Before They Start”, 26.2 was “The AI-Proof Idea Filter”, 26.3 was “Three Layers of Opportunity”, 26.4 was “The Moat-First Business Plan”, 26.6 was “The Global Opportunity Map”, and 26.7 was “Consciousness as Competitive Advantage in Entrepreneurship”. 26.5 keeps its print title.