AI-Proof: What Machines Can Never Be — Volume II: Application, Part IV: Conscious Creation, Chapter 15: Building Real Moats in an AI World

Part IV is conscious creation. After the commodity trap of Chapter 14, this chapter asks what can still give anyone a reason to choose you when a machine helps others do your work for less, and what cannot.

← Previous: Chapter 14: The Commodity Trap · Series hub · Next: Chapter 16: Embed AI Deeply, Don’t Bolt It On →

A platform offered one of Rajan’s oldest buyers the same cloth for twelve per cent less, with an insurer behind the promise. What Rajan had left was not a moat. It was a man who phones first, and nobody owes him the order.

In Brief: For readers who wonder what will still make anyone choose them when machines make their work cheap. Chapter 15 asks what can still give anyone a reason to choose you when a machine helps others do your work for less. A cloth maker in Coimbatore hears that a sourcing platform will sell one of his oldest buyers the same cloth, cheaper, with an insurer behind it; a founder in Lagos builds a record of what traders know about one another, and finds that the people in it want to answer back. Warren Buffett on moats, the Tirukkuṟaḷ and Adam Smith on why traders keep faith, George Stigler on whom rules serve, and Kauṭilya and Guido Calabresi on who should pay when work fails help answer what a machine cannot copy, and why it gives no one a right to be chosen.

What a machine cannot copy is the person who answers; that is not a moat but a duty done again each time, and it gives no one a right to be chosen.

Opening: The Call from Hamburg

“I am not calling to negotiate,” Jürgen Weiss said.

It was February 2026, late in the afternoon in Coimbatore. Weiss had bought cloth from Rajan Subramanian since 2013, for a firm in Hamburg. Rajan knew Weiss’s short sentences the way he knew the sound of the looms below his office. He also knew when Weiss was leaving something out. Today Weiss was leaving nothing out, and that was what made the call unusual.

“I am calling because I owe you an honest conversation,” Weiss said. “I have been offered a better deal. I want you to hear it from me.”

Then he gave the facts. Since late 2024 Weiss had bought his standard weaves through a sourcing platform. Software on the platform matched his specifications to mills across South India and handled the shipping and the papers. Now the same platform had made a new offer. It would supply the yarn-dyed checks that Weiss still bought from Rajan: cloth for table linen, where the yarn is dyed before weaving and the pattern must come out square. Its price was twelve per cent below Rajan’s. Cameras would inspect every lot at the mill, and the records would be online. And the promise was written down. If a lot failed, the platform would refund it and replace it within a stated time, and an insurer stood behind that promise. There was even an account manager in Hamburg, a person Weiss could telephone.

“On paper, I should take it,” Weiss said. “I have not decided. But the offer exists, and the next one will be harder for me to refuse.”

Rajan thanked him, and meant it. After the call he sat for a long time, looking down at the floor through his office window. One hundred and forty people work there. They weave and finish cloth for buyers in fourteen countries. The yarn is dyed outside, by dye houses in Tiruppur; the weaving and the hand-finishing are done here. Since April 2025 there had been cameras on that floor and a scheduler deciding which looms ran which cloth; Chapter 18 tells how that change came, and what it cost. Nobody could say that Rajan was simply afraid of machines.

And the offer had not come from a machine. It came from a company that used machines, and that company had something Rajan did not: a written promise with an insurer’s money behind it. What Rajan had was harder to name. For nearly thirty years he had done what a buyer in Stuttgart once said of him: when something goes wrong, he calls you before you call him. Was that worth twelve per cent? Weiss would decide. Rajan could not decide it for him.

So this chapter asks: If a machine can do your work for less, what reason is left for anyone to choose you, and does anything give you a right to be chosen?

The print edition’s reply was eight “moats”. The word, in this business sense, was made famous by the American investor Warren Buffett. The print edition said that the strongest moats “are built from things that require consciousness”. It said that a machine “cannot build them” because “it has nothing to lose”. And it ended: “The exposure was the moat.” This edition keeps the eight subjects and tests each one. It does not keep the conclusion. Chapter 3 already found that “nothing to lose” is not quite true of machines. What a machine has to lose was put there by its makers, and it can be copied and restored. But the deeper mistake is simpler. The competitor in this story is a company, with a balance sheet, an insurer and an office in Hamburg. Machines do not compete with Rajan. Companies that use them do. Even if Weiss one day ran such software himself, with no platform in between, the cloth would still come off someone’s looms. He would still be choosing whom to trust.

Buffett explained the word plainly in his letter to Berkshire Hathaway’s shareholders for 2007. “A truly great business must have an enduring ‘moat’ that protects excellent returns on invested capital. The dynamics of capitalism guarantee that competitors will repeatedly assault any business ‘castle’ that is earning high returns.”¹ A moat, in his sense, is whatever stops a rival from taking your customers and your margin. His examples are a lower cost and a famous brand; others would add a licence nobody else can get. It protects the owner. Whether it also serves the customer is another question. The print edition asked why Weiss might stay, and never whether he should.

The title of this chapter needs an argument before the sections begin. In the same letter Buffett explains which businesses he will not buy. “A moat that must be continuously rebuilt will eventually be no moat at all.” And: “Additionally, this criterion eliminates the business whose success depends on having a great manager.” By the test of the man who made the word famous, Rajan has no moat. What he offers must be done again with every lot, and it depends on him. To say that a moat is a rate of deepening, not a wall, is to see half of this. Buffett would call that no moat at all, and on his own terms he would be right. This chapter draws the other conclusion. What is left is not an asset that protects returns. It is a practice, a person answering again and again for what leaves his hands, and a company can match much of it. Read “real moats” as a hope. This edition looks at what is left when the hope is tested.

The eight sections that follow are eight places where a business hopes to be safe. Each is tested with one question: who answers there? To answer, in this book, is to explain what happened, undo it where it can be undone, and put right what can be put right. Answering runs through all eight, from the trader who signs an entry to the man who calls first. The last, the liability shield, asks whether answering can be sold.

15.1 Proprietary Data: The Record Other People Write

It is easy to call proprietary data the moat you build by listening. The listening was real. But the knowledge it gathered belonged first to other people.

Chapter 14 told how Chidi Okonkwo, a founder in Lagos, lost his first company in 2024. In February 2025 he spent a week beside the chair of his uncle Emeka, who trades cloth in Balogun Market, on Lagos Island. The traders there know things no bank records. They know which suppliers deliver in the rains, whose cloth holds its colour, whose payments clear. They pass this on in conversation, and most of it is gone the moment it is said.

Chidi stayed in the market for two more months before he wrote any code. In August 2025 he launched a second company. At its simplest it was a ledger: a shared record in which traders wrote down the judgements they were already making. This supplier delivered on time. That buyer paid late. This lot held its colour after washing. Each entry was small. Gathered over months, the entries became something that existed nowhere else in writing.

The machine came second. Language models summarised a supplier’s history and noticed patterns across hundreds of entries. Chidi knew from his first company that this part was not his. Any rival could rent the same models. The record could not be rented, because it was made by people choosing to share what they knew.

So he built consent into it. Writing in the ledger was a choice. A trader could see, change and withdraw her own entries, and the record would not be sold to anyone else. Then he made one more decision, and it cost him entries. Every entry carried the name of the trader who wrote it, visible to the other members. Many traders would say a thing over tea that they would not sign. Fewer entries came in. Those that came had someone behind them.

That decision matters more than the software. Chapter 8 said that we should believe a source as far as we can check it, or as far as someone stands behind it. An entry in a ledger is a piece of testimony about another person. Unsigned, it is a rumour with good formatting. Signed, it is a claim someone can be asked about. The records were valuable to the traders because each one had a writer who could be found.

In the investor’s sense, this is partly a moat. A rival could copy the software in weeks. It could not quickly copy the habit of hundreds of traders writing down what they saw and putting their names to it. But the advantage lasts only while the knowledge stays outside other records. Banks, mobile-money firms and credit bureaus record more of commercial life each year. What the ledger holds that they do not is the judgement a credit score leaves out: who delivers in the rains, whose colours run. As other records grow, that space shrinks.

And the record was built from what others gave. The traders wrote it; the company holds it; the company’s value rests on it. Consent to write an entry is not consent to every use of it. If the company were sold, or turned its record to deciding who gets credit or insurance, the traders who wrote it would have no say in that choice. They would keep only the right to withdraw their own entries. What they are owed is Chapter 20’s question. Here it is enough to say that “proprietary” was the wrong first word. It was other people’s knowledge, kept in one place, and it would last only as long as they kept giving it.

15.2 Network Effects: Staying Because the Others Are There

Network effects look like the moat you build by connecting. The connecting happened. Whether it built a moat is what this section tests.

In February 2026, six months after the launch, a cloth trader in Accra looked at a dye house in Lomé, in Togo. She had joined the ledger because a cousin in Lagos told her to. She had never used the dye house and had never met its owner. The entries she read had been written by traders in Lagos, Cotonou and Abidjan, and she had never met them either. Each entry carried a name, a date and what the writer had seen. She placed a small order. The cloth arrived as described, and she wrote an entry of her own for the next stranger to read.

That is a network effect: a product becomes more useful to each person as more people use it. A product with a million users who never touch one another has a million separate customers, each free to leave when something cheaper comes. A product whose users help one another has a network, and the network is a reason to stay.

Here the record did something Emeka’s chair could not do. Emeka’s trust ran through people he knew, and through what he knew of their families. The ledger let a stranger trust a stranger, because other people had signed what they saw. The machine’s summaries made that fast. A summary, though, is the company’s statement, not a trader’s. If it misreports what the traders signed, the company must explain and correct it, and the signed entries beneath it are how anyone can check. Trust between strangers is a real gain. A trader in Accra could judge a supplier in Lomé without a cousin, an uncle or a father in common.

The network was not new. Balogun’s traders had been a network for generations, sharing news across West Africa long before any software. The company gave that network a written memory. It gave trust that already existed a way to travel.

But look at what a network effect is. It is a reason to stay that has nothing to do with whether the company is good. You stay because the others are there. That is why investors prize it, and why a user should look at it twice. A product with a strong network can grow worse and keep its people. Passengers on a bus switch when a better bus comes, and members of a community stay. That is true. Members also stay when the bus gets worse.

In March 2026 the ledger met someone who was not in it. A woman who runs a small workshop dyeing adire, the indigo-patterned cloth of Abeokuta, north of Lagos, heard from a customer that she was described in the ledger. Two traders had written that her deliveries were late in the rainy season. She did not dispute the dates. She wanted to say why, and what she had changed since. Nothing in the ledger let her. She was being judged, by name, in a place she could not enter.

She found Chidi’s number and called him. “Who wrote this?” she asked. “I want to say my side.”

In April 2026 Chidi added a right of reply. Anyone named in an entry could write one reply beneath it, and the reply would be shown with the entry. It had a cost. Some traders complained that the replies turned a record into an argument. A few stopped writing about suppliers who might reply. Others said the record was now more useful: a late delivery with a reason is a different fact from a late delivery without one. By the autumn of 2026 it was not clear which effect was larger.

A right of reply works only if the person knows that a claim about her exists, can read it, and can reply without risk to her business. The dyer heard by chance, from a customer. Whether the others named in the ledger will learn that they are there, the new rule does not settle. Nor does a reply correct anything. Had an entry been false, a reply would only have set her word beside the writer’s. The rule did not say who could check an entry, or take it down. A network that writes about people outside it holds judgements made in their absence. The traders stood behind what they wrote, because their names were on it. The reply lets the person judged explain. In a record about people, the right to explain has to run both ways.

The glitch

Chidi’s ledger rests on two choices: every entry carries a name, and the person named can reply. eBay’s record of strangers judging strangers shows what such choices can cost. There, buyers and sellers once rated each other after a sale, under their user names, and each could see what the other wrote. In 2013 the economists Gary Bolton, Ben Greiner and Axel Ockenfels reported on 742,829 eBay users with any rating. Four in five had more than 99 per cent positive ratings.² Two other researchers, Dellarocas and Wood, estimated that buyers were at least mildly dissatisfied in about 21 per cent of sales. Many buyers, they argued, left no rating at all for fear of a bad one in return.

eBay’s repair ran against Chidi’s design. In 2007 it added detailed ratings of sellers that were one-sided and anonymous: the seller could not see who gave them and could not reply. Bolton and his colleagues found that many traders used them to avoid retaliation, and that ratings became more informative. From the spring of 2008, sellers could leave buyers only positive feedback.²

So a signed record with a right of reply can be honest and still say too little. The traders who stopped writing about suppliers who might reply show the same effect on a small scale. A named entry can be questioned; its price is the bad news nobody signs. The ledger counts no entries that were never written, so it cannot show that price.

15.3 Distribution: How People Come to Choose You

Distribution control sounds like the moat you build by being chosen. Being chosen is real, but nobody controls it.

Weiss did not find Rajan through a search. In 2013 he asked a buyer in Stuttgart, who had bought from Rajan since 2008, what kind of supplier Rajan was. The reply was specific. “The quality is consistent, the communication is honest, and when something goes wrong he calls you before you call him.” Weiss placed a trial order. It arrived on time and to specification, with a note in Rajan’s own hand thanking him for the business and asking what he should know about Weiss’s standards. That was thirteen years before the call in February.

Distribution is how a business reaches the people who buy from it. A useful line runs between it and marketing. Marketing is what you do to be seen; distribution, in this sense, is what happens when people choose to find you. Rajan has buyers in fourteen countries, and most of them came the way Weiss came: one person telling another.

Being chosen this way has two faces. It is slow, and it is honest about the right thing. The Stuttgart buyer did not recommend a price or a product. He recommended a habit: what Rajan does when something goes wrong. But it also favours the people already known. A weaver with better cloth and no friend in Stuttgart waits outside. Chapter 7 asked whether part of what buyers pay Rajan for is not that habit at all. It may be the comfort of an old relationship: a premium for being known, like the picture of a reliable man that Rajan himself once held of Senthil’s father. Being recommended is partly being recognised, and recognition runs along old lines.

The platform reaches buyers another way. It does not wait to be recommended; it ranks mills by price, speed and record, and it found Weiss mills he would never have heard of through Stuttgart. Some of them were good. For his standard weaves, the platform’s way of choosing widened his choice. A recommendation carries the experience of the person who makes it, and also his field of view. The platform widens the field without making its own ranking neutral: it may favour the cheapest mills, the largest, or those best at filling in its forms.

Calls like Rajan’s are sometimes counted as investments in distribution, with returns their maker will never trace. That description takes back what it gives. A call made as an investment has a transaction in mind. The Stuttgart buyer’s sentence was true because Rajan did not call him to be recommended. A buyer may choose a business for a habit like that; it is a reason, not a right. A business cannot buy the habit by calling it marketing.

15.4 Brand and Trust: What Being Believed Is Made Of

Brand and trust can be called the moat you build by being believed. Being believed matters. This section asks what it is made of, and whether anyone owns it.

As of 2026 a machine can design a logo, a colour scheme and a slogan in minutes. Making a brand has become cheap. A brand in the older sense, a promise kept so often that people expect it, has not become cheaper. It is made of what happened the last time something went wrong, and no design tool can make that.

Chapter 3 described trust, with the philosopher Annette Baier, as accepting that someone could let you down, while not expecting them to. It also said that trust in a machine is really trust in the people behind it. A brand is trust in people under one name. The platform’s name stands for its staff, its rules, its records and its insurer. Rajan’s name stands for Rajan.

What trust between traders looks like was shown in Chapter 7. For years Rajan doubted Senthil, the mill owner in Tiruppur whose father first supplied him in 1998. Then his nephew Karthik scored ten years of Rajan’s own records, and Senthil’s mill came out the most reliable on his books. In February 2025 a lot from that mill failed a wash test. Senthil phoned Rajan before Rajan could phone him, explained what had gone wrong, and offered to re-run the whole order at his own cost. Rajan phoned his buyer and took the five days’ delay on himself. That is the shape of trust in a trade: who calls first, who explains, who pays.

On the Sunday after Weiss’s call, Karthik telephoned from Singapore, as he does every Sunday. He had read the platform’s terms. “Their promise is written down,” he said. “Yours is in your head, and in Weiss’s memory.” He was right, and his uncle had no reply to it. A written promise can be checked and enforced. Rajan’s habit of calling first cannot be enforced by anyone; it can only be remembered. Some of it could be written down. Rajan could put two promises into his terms of sale. He would tell a buyer of a known fault before shipping. He would replace a failed lot within a stated time, at his own cost. That much could then be checked and enforced. The money behind it would be his own, not an insurer’s, and a firm of his size can promise only so much. Part of what passes for depth may be a record nobody kept. A kept record can be a rung for someone else. An unkept one can pass for something deeper than it is.

Two old accounts disagree about why a trader keeps faith at all.

The Tirukkuṟaḷ, the classical Tamil book of couplets, has a chapter called naṭuvunilaimai: standing in the middle, or impartiality. Its last verse is about traders. In V. V. S. Aiyar’s translation of 1916: “Behold the business man that looketh after the interests of others as his own: his business will expand.”² G. U. Pope, translating in 1886, made the same turn in rhyme: “As thriving trader is the trader known, / Who guards another’s interests as his own.”³ Both English versions turn the verse into a promise of profit. The Tamil is shorter, and its modern commentators read it differently. The twentieth-century Tamil scholar M. Varadarasan glosses it this way: to trade guarding the goods of others as your own is the right way of trade.⁴ Read so, the verse does not promise that honesty pays. It says what trade is.

Adam Smith, lecturing in Glasgow in the 1760s, gave the opposite account. “Whenever commerce is introduced into any country, probity and punctuality always accompany it.” He did not think it came from national character. “It is far more reducible to self-interest, that general principle which regulates the actions of every man, and which leads men to act in a certain manner from views of advantage, and is as deeply implanted in an Englishman as a Dutchman.” His explanation was the number of dealings: “A dealer is afraid of losing his character, and is scrupulous in observing every engagement. When a person makes perhaps twenty contracts in a day, he cannot gain so much by endeavouring to impose on his neighbours, as the very appearance of a cheat would make him lose.”⁵ Where people deal with each other rarely, he went on, a clever trick can pay; where they deal often, it cannot.

The two lamps are not a choice between virtue and self-interest. Smith explains why repeated dealing makes honesty the rational course. The verse, on its modern commentators’ reading, says what good trading is; I read it as speaking to the cases the calculation does not decide. Most honest traders have both reasons, and few could say which one moved them on a given day. The English translators, by turning the verse into a promise of profit, slide towards Smith; the commentators keep the two apart.

On Smith’s account, the platform should be the most honest trader of all. It makes not twenty contracts a day but thousands, and every one is recorded. Its cameras inspect every lot, and every complaint is recorded against the mill. That is the strongest case against Rajan in this chapter, and it should be stated plainly. If honesty in trade comes from frequent dealing and the fear of being found out, a well-run platform has more of both than any single man.

Read this way, the verse bears on something Smith’s account leaves to habit: what a trader does where no inspection reaches. Smith’s dealer is honest because “the very appearance of a cheat” would cost him. Where inspection reaches almost everything, his reason reaches almost every case. For the rest, a habit decides, and a habit can be formed by the fear of being found out as well as by care for another’s goods. The two accounts do not settle which. A buyer cannot see which reason moves a trader. He can see only what the trader did after something went wrong. The verse speaks of merchants in an old Tamil market, not of export contracts, and I borrow only its question: why does a trader keep faith?

Trust, then, is a record of acts, kept partly in other people’s memories, and it must be renewed with every lot. Rajan does not own it. The platform’s trust is a record too, kept in its systems and backed by its insurer. Weiss may reasonably prefer either. Neither is owed to him, and neither owes him the other.

15.5 Regulatory Positioning: Rules That Name Who Answers

The print edition advised companies to make themselves partners to regulators, so that regulators would “consult rather than constrain” them. And it gave Rajan a line: “His competitors know the regulations. Rajan knows the people who enforce them.” Both should be withdrawn. A business that is safe because it is close to the people who enforce the rules is protected by its access, whatever its quality.

The American economist George Stigler described where that leads. “A central thesis of this paper,” he wrote in 1971, “is that, as a rule, regulation is acquired by the industry and is designed and operated primarily for its benefit.”⁶ Economists call this regulatory capture: rules written and run mainly for the firms they are meant to govern. A moat dug by friendship with regulators is capture seen from inside the castle. It may protect the castle. It does not protect the people the rules were for.

What can honestly be said about rules is narrower. They change, and they differ by place. In the European Union, Regulation (EU) 2026/1744 entered into force on 27 July 2026. It moved the Artificial Intelligence Act’s duties for the high-risk systems listed in Annex III to 2 December 2027, and for AI built into products to 2 August 2028.⁷ India’s Ministry of Electronics and Information Technology published AI governance guidelines on 5 November 2025. It chose, for now, to rely on existing laws, strengthened where needed, rather than a new AI law.⁸ Chapter 25 maps the rest. Here what matters is a pattern. Rules raise the cost of entering a trade, and the cost falls hardest on the small. A platform can hire people to fill in every questionnaire. Rajan fills in his himself.

That is happening already. Weiss’s own customers now send him questions about the mills behind his cloth: working hours, wages, the chemicals in the dye baths. The platform replies from records online, in a day. Rajan replies from what he knows of Senthil’s floor and of the dye houses he has visited for years. What he knows is true, and most of it is unwritten. The rule rewards the record, not the acquaintance. On this point the platform is better, and the reason is the one Karthik gave on Sunday. Documentation is not a trick of the large. It can make working hours, wages and the chemicals in a dye bath visible to people who will never visit the floor. An unwritten assurance, however sincere, is hard for them to check. A small firm needs simpler ways to show that it complies, not an exemption from showing it.

There is a legitimate kind of regulatory knowledge, though it protects no one from rivals. It means knowing what a rule is for, and being able to show a regulator, a buyer or a worker who answers for what. A rule that names who must answer is the law’s version of the Stuttgart buyer’s sentence. Positioning, in that sense, means being the kind of business that can be named.

15.6 Community: Belonging, and the Door It Closes

A community can look like the moat you build by belonging. Belonging is real, and so is what it shuts out.

Rajan’s world is larger than his factory. It holds the mill owners in Tiruppur, the dye houses and the finishing house in Karur. It holds a shipping agent in Chennai who once held a container for him because he knew Rajan would pay. They go to one another’s family weddings. They share meals when business brings them to the same town. They telephone with news that might matter, and not only when they want something.

The print edition called this “a world”, and said that members of a community do not leave because “the switching cost is social. It is the loss of relationships, shared history, and mutual obligation.” Some of that is true. A world like this is built over decades, and no platform can rent it. But a switching cost is what a customer loses by leaving. Making it social is still making leaving costly. Belonging that works as a moat is a kind of lock-in, described in warmer words. Lock-in means that customers stay because leaving would cost them too much.

It also keeps people out. The verse from the Kuṟaḷ in 15.4 sits in a chapter about standing in the middle. Two verses earlier, Aiyar’s translation reads: “Behold the weighing beam, for it is straight in itself and weigheth justly: the glory of the wise is to be like unto it and to incline neither to this side nor to that.”² The verse speaks of the wise, not of markets; applying its beam to a trading world is my reading. A trading world leans. It leans towards its own families, its own towns, the people whose fathers its members knew. That is a comfort to those inside and a cost to everyone outside. The cost is paid by people who never appear in the story: the newcomer whose cloth was good and whose name was unknown. The Kuṟaḷ put honest trade in a chapter about not leaning. A community moat is made of leaning.

The platform does not lean in that way. Its rankings have their own biases, but they do not care whose father anyone knew. Among the mills it found for Weiss were small units that nobody in Rajan’s world would have recommended, and some of them were good.

None of this makes Rajan’s world worthless. It holds knowledge, help in bad seasons and a habit of explaining to one another: the agent who held the container could be asked why, and so could Rajan. What it cannot honestly be called is a right. It is a reason some members stay, and a door some outsiders find closed. Keeping it also costs hours: evenings at weddings, calls on Sundays. Some of those hours are Rajan’s. Some belong to his family, and their names are not on the factory.

15.7 The Physical World: Atoms and Bits

The print edition claimed that machines move bits fast but atoms slowly. “You cannot compress the curing time of a dye or the drying time of a finish,” it said. “The atoms resist, and the resistance is a moat.”

The atoms resist everyone. The platform’s mills wait for the same dyes to fix and the same cloth to dry. Their containers spend the same weeks at sea between South India and Hamburg. The platform works in cloth too: it books looms in other people’s mills and moves real bales. Physics slows the whole trade. It does not protect Rajan from a rival who also works in cloth.

Nor is the physical layer free of machines. On Rajan’s own floor, the cameras mentioned at the start inspect faster and more consistently than tired eyes at the end of a shift. The platform’s mills use the same kinds of tools.

What the atoms give Rajan is something else. He owns looms and a floor, and he employs about 140 people. The platform rents capacity; Rajan holds it. That gives him control over how his cloth is made, and obligations the platform does not carry. His workers’ wages must be paid in a slow season. If he matched the platform’s twelve per cent, the margin would come first from the floor, before it came from him. The platform’s insurer does not carry that wage bill. This is not a reason to choose Rajan. It is why matching the price is a different act for him than for the platform. Who would pay for the cut is a choice, not a law of the trade.

Rajan’s hands are another matter. What his fingers tell him, he learned the year before, is a reason to ask for a test, not a verdict. Such knowledge is depth, and depth fills a position but does not grant one. The same holds for Kamala, who is sixty-eight and finishes the speciality cloth by hand, and for Muthu, who has run the floor for twenty years. Kamala’s name is on the tag of the seasonal collection. Calling her skill a moat does not settle whether the business will pay to teach it. A moat that rests on one woman’s hands is a debt as much as an asset. Who will learn her work, and who will pay for the learning? Nobody has decided.

15.8 The Liability Shield: The Moat You Build by Taking Responsibility

In ordinary business English, a liability shield protects you from paying for harm. A limited company is one; a contract clause or a safe-harbour rule, a law that excuses a firm from paying if it meets set conditions, can be another. The print edition used the phrase the other way round. It meant a moat built by taking liability on: being the professional whose name, licence and money stand behind the work. Both meanings are in this chapter’s story, and the difference between them is the chapter’s argument.

The shield that protects from paying was tried in 2024. Chapter 8 told how Air Canada argued before a tribunal in British Columbia that it was not responsible for what its website’s chatbot had told a grieving customer. Air Canada, the tribunal said, “in effect” suggested that the chatbot “is a separate legal entity that is responsible for its own actions”. The tribunal refused.⁹ The company answered for its machine. The Prologue’s second never says the same as a commitment: “A machine’s makers and users will never be free of what it does.”

The other shield, the moat, rests on a claim about the law. When a machine drafts a diagnosis, a brief or a design, the person who signs it is held responsible. So the value seems to move from making the work to guaranteeing it. As of 2026 that is partly true. Courts have sanctioned lawyers who filed invented citations.¹⁰ But whether a signature should carry the liability is a question two old voices settle differently.

The Arthaśāstra, the Indian treatise on statecraft attributed to Kauṭilya, has a book called kaṇṭakaśodhana, “the removal of thorns”. It begins with rules for artisans, and its rules are hard. In R. Shamasastry’s translation, weavers who return less than the thread they were given was meant to yield must “not only pay either a fine equal to twice the loss in threads or the value of the whole yarn, but also forfeit their wages.” Washermen who return the wrong cloth “shall not only be punished with a fine equal to twice the value of the clothes, but also be made to restore the true ones.” And: “Physicians undertaking medical treatment without intimating (to the government) the dangerous nature of the disease shall, if the patient dies, be punished with the first amercement,” the lowest of the treatise’s three grades of fine.¹¹ The state laid liability on the hand that did the work, and it priced the failure. It also asked for more than money: the washerman must restore the true cloth, and the physician must report a dangerous case before he treats it. These are rules for a king’s market, not for a modern court, and I borrow only their shape.

The American legal scholar Guido Calabresi asked a different question. His book The Costs of Accidents (1970) argued that accident law should aim, among other things, to reduce the total cost of accidents and of avoiding them. In an article of 1975 he put the question that follows: “Who is best suited to make the cost-benefit analysis between accident costs and accident avoidance costs?”¹² In plain words: who is best placed to weigh what failures cost against what it would cost to prevent them? Put the liability there, because that person can actually reduce the harm.

The two lamps disagree, and the disagreement matters for every professional who signs a machine’s work. Kauṭilya puts liability on the hand that did the work. Calabresi puts it on whoever is best placed to weigh what failure costs against what preventing it costs. When a machine’s error lies in how it was built or trained, that is often its maker, not the person who signs at the end. His question does not always point away from the signer. Checking a citation is cheap, and the lawyer who files it is best placed to check it. So, as I read Calabresi, the courts that sanctioned lawyers put the cost in the right place. The law is moving partly in the direction his question points. The European Union’s new directive on defective products treats software, including AI systems, as a product, for products placed on the market after 8 December 2026.¹³ It pays only for death or personal injury, damage to private property, and the loss or corruption of private data. Under it, a maker may have to pay when a machine’s wrong reading injures a patient. A failed lot of cloth, or a lawyer’s sanction for invented citations, is a business loss, and the directive pays nothing for it. A separate proposal for rules on AI liability claims was withdrawn; the notice appeared on 6 October 2025.¹⁴ As liability moves towards makers, the signature as a business narrows.

Where it does not move, the signature can become something worse than a moat. A person paid to sign what she has no time to check is not a shield for anyone. Chapter 1 called her a moral crumple zone: the one built to take the blame. A firm that sells “a human in the loop” while giving that human no time is selling a crumple zone.

What a signature can honestly carry is better shown by Chapter 6. In February 2025 a machine found a small nodule on a chest X-ray that the radiologist Sanjay Rao had missed at the end of an evening shift. He wrote an addendum under his own name and agreed that the patient should be told. The machine was the better reader that evening. Sanjay was the one who answered. But responsibility follows control as well as signatures. The hospital set the shift: he missed the nodule at hour nine, with a backlog of cases. And it will not fill three posts as their radiologists retire; the machine and the remaining doctors will absorb the work. It answers too: an institution that makes checking hard shares the duty to explain and put right.

Then there are the insurers. In 2025 several insurers at Lloyd’s, the London insurance market, began to back a policy sold by a company called Armilla. It pays damages and legal costs when an AI tool performs below what was expected of it. “A mistake by an AI tool would not on its own be enough to trigger a payout under Armilla’s policy,” the report said.¹⁵ The platform’s guarantee to Weiss is this kind of promise: money from a balance sheet, pooled across many customers and priced in advance. So a buyer can now ask plainly: do I want the person, or the balance sheet?

Chapter 5 drew part of the line. Liability, being bound to pay, can be bought, through contracts and insurance; explaining what broke and putting it right cannot be bought from an insurer. This chapter must add something: explaining and putting right can be bought from a company. The platform employs a person in Hamburg whose job is to explain and to arrange the replacement. A model cannot sell its own answering. As of 2026 no law treats a model as a party; the EU’s new directive will treat it as a product. The Prologue’s second never holds that a machine’s makers and users will never be free of what it does. A company can sell answering for a model, through its people and its insurer. What Rajan offers is narrower still: one particular person’s answering, from a man who knows the looms the cloth came off. The print edition said accountability was the one product a model cannot sell. Narrowed, that survives. It is not a product only a human being can sell. It is a product a company can sell for its machines, and one person can sell only for himself.

Weiss tested it. In March 2026 he split the yarn-dyed checks for the spring and summer: half from Rajan, half from the platform. He did not have to choose one supplier for everything. He already bought standard weaves from the platform and speciality pieces from Rajan; the question was which cloth went where, not a verdict on a man.

By September each side had met a failure, though not the same kind. A platform lot arrived off-shade. The mill’s cameras had passed it; Weiss’s own inspection in Hamburg rejected it on arrival. The platform refunded and replaced it within its stated time. Nobody phoned first, and nobody needed to: the money and the new lot came when they were promised. One of Rajan’s lots had a weaving fault on two of his own looms, where the checks had drifted out of line. His floor caught it before the cloth left, at the cost of five days. Rajan phoned Weiss before shipping and re-wove the lot at his own cost. He had done for Weiss what Senthil had done for him in February 2025.

The two failures tested different things. One showed that the platform’s remedy worked when its inspection did not. The other showed that Rajan’s floor could catch a fault before it reached the buyer, and that catching it cost time. The twelve per cent was a visible number. The cost of a failure was not. It depended on when a fault was found, whether a replacement came in time, and what Weiss’s own customers would lose if the cloth arrived late. One season and two incidents are informative. They are not a controlled test of which supplier is more reliable, and a result inside a story is not proof of the choice.

Weiss had his cloth either way. The platform’s remedy was certain and fast, and came in money and a replacement lot. Rajan’s was slower, and came with an explanation and a man who could be asked what would be different next time. By October 2026 Weiss had not decided. Nor should the story decide for him.

The print edition ended this section with its deepest claim. Every moat, it said, is built from “exposure”, and a machine cannot build one because “it lacks the capacity to be at risk”. Exposure is not the line. The platform is exposed: its insurer pays when it fails. Rajan is exposed. So is a nurse who signs a chart she had no time to read, and she has no power over what she signs. Exposure without the power to explain, to undo and to put right is only risk.

And answering is not a moat. By Buffett’s own test it cannot be: it must be done again for every lot, by the person who offers it, and a rival can offer another kind. It also ends with the person. If Rajan’s firm is to outlast him, others on his floor must be given the authority to call first, explain and put right in its name. Then it offers what the platform offers, a company’s answering through its people, and the jobs of the people who work there will depend on whether that is enough. Chapter 14 held that surviving is not the same as deserving to exist. Calling first does not make Rajan’s business deserve to exist; it gives some buyers, for some cloth, at some price, a reason to choose it this season. It gives him no right to be chosen at all. Weiss owes Rajan what he gave him in February: an honest call before he decides.

So the chapter’s doubt can be settled, though not comfortably. A machine helps others do the work for less. What is left is reasons: a record people sign, a network, a habit of calling first, looms and hands near the cloth. A company can do much of what Rajan’s habit does. It can tell a buyer first, explain what went wrong, replace the cloth, pay, and give him a named person to call. What it cannot do is make Rajan’s history its own: the plain fact that this man made these decisions, on these looms, and stands behind them now. That is all this chapter means by the person who answers. It is a fact about who he is. It is not a capacity a machine lacks and might one day gain, and not a service nobody else can supply. It may matter to a buyer. It does not make Rajan’s assurance better than the platform’s, and it gives him no right to the order. The first never holds by meaning: a machine will never be you. Being you is not a better product. It can only be offered, again, each time.

Ending: The February Dusk

On the evening of the call from Hamburg, Rajan sat in his office while the February dusk came down over Coimbatore. The floor below was dark. The looms were still.

He picked up his phone and called Senthil. It was not about business. Senthil’s mother was to have an operation on her eyes the next week, and Rajan wanted to know which hospital, and whether she was afraid. They talked for twenty minutes. Senthil asked about Rajan’s knees. They discussed whether the monsoon would come early.

The call advanced no order and solved no problem. No system that scored suppliers would have counted it, and Rajan did not make it to be counted.

He put the phone down. In the morning there would be cloth to look at and a call to a buyer in Stockholm about a delayed order. There was also a reply to Hamburg that he had not yet written.

Practice

The first hour after. Choose one piece of work that leaves your hands with your name, or your firm’s, on it. Write down what you would do in the first hour after learning that it had failed. Whom would you tell, before they found out for themselves? What would you explain? What could you undo, and what would you pay or put right? Beside each step, write whether you have the time and the authority to do it, and if you do not, who does. Then mark which steps a written guarantee or an insurer could do instead, and which need someone who knows how the work was made. If you work for someone else, note whose hours each step would take. The page will not win or keep a customer. It shows what kind of answer you are offering, and whether anyone has given you the means to offer it.

Words to keep

• distribution: how a business reaches the people who buy from it; in this chapter, being chosen because someone recommends you.

• kaṇṭakaśodhana: “the removal of thorns”; the book of Kauṭilya’s Arthaśāstra that protects people from those who harm them, dishonest artisans and traders among them.

• liability: being bound by law or contract to pay when something goes wrong. It can be bought, through contracts and insurance.

• liability shield: in ordinary use, something that protects a person or firm from having to pay when things go wrong. The print edition used the phrase for the opposite: taking liability on.

• moat: in Buffett’s sense, what protects a business’s returns from its rivals: a lower cost, a famous brand; others would add a licence nobody else can get.

• naṭuvunilaimai: Tamil, “standing in the middle”; impartiality. The title of the Tirukkuṟaḷ’s chapter 12, which ends with a verse on traders.

• network effect: a product becoming more useful to each person as more people use it.

• probity: honesty in dealing; in Smith’s account, the product of frequent dealing and the fear of losing one’s character.

• regulatory capture: rules written and run mainly for the benefit of the firms they are meant to govern (Stigler).

• switching cost: what a customer loses by leaving.

Sources

1. Warren E. Buffett, letter to the shareholders of Berkshire Hathaway Inc. for 2007 (February 2008), p. 6 (berkshirehathaway.com). Buffett had used the image before, as in the 1993 letter’s “a protective moat around their economic castles”.

2. Tirukkuṟaḷ 118 and 120 (chapter 12, “Uprightness of Heart”; Tamil naṭuvunilaimai), trans. V. V. S. Aiyar, The Kural or the Maxims of Tiruvalluvar (first published Madras, 1916; second edition, undated, Sheramadevi: Bharadvaja Ashrama), checked against the Tamil Digital Library copy of the second edition. For “The glitch” in 15.2: Gary Bolton, Ben Greiner and Axel Ockenfels, “Engineering Trust: Reciprocity in the Production of Reputation Information,” Management Science 59, no. 2 (2013): 265–285, as in the authors’ working paper of 25 March 2012 (ockenfels.uni-koeln.de): of 742,829 eBay users who had received at least one feedback, 67 per cent had 100 per cent positive and 80.5 per cent more than 99 per cent; eBay’s detailed seller ratings (spring 2007) were blind, one-sided and anonymous, “many traders use the new system to avoid retaliation”, and this improved informativeness; from spring 2008 sellers could leave only positive feedback. The estimate that buyers are at least mildly dissatisfied in about 21 per cent of transactions, and the argument that many give no feedback for fear of retaliation, are Dellarocas and Wood’s (2008), as the paper reports them (checked 4 November 2026).

3. G. U. Pope, The Sacred Kurral of Tiruvalluva-Nayanar (London: W. H. Allen, 1886), chapter 12, “Impartiality”; text as in Project Madurai (South India Saiva Siddhanta Works Publishing Society reprint, 1962, which also carries the translations of W. H. Drew, John Lazarus and F. W. Ellis).

4. Tirukkuṟaḷ 120, vāṇikam ceyvārkku vāṇikam pēṇip / piṟavum tamapōl ceyiṉ, with the commentary of M. Varadarasan (Mu. Va.), which reads the verse as defining the right way of trade; the commentaries of Solomon Pappaiah and Puliyur Kesikan read it the same way (thirukkural.net). The English rendering of the gloss is mine.

5. Adam Smith, Lectures on Jurisprudence, report of 1766; text as in Lectures on Justice, Police, Revenue and Arms, ed. Edwin Cannan (Oxford: Clarendon Press, 1896), Part II, Division II, §17, “Of the Influence of Commerce on Manners”. The Glasgow edition (Oxford: Clarendon Press, 1978; Indianapolis: Liberty Fund, 1982) prints “20” for “twenty”.

6. George J. Stigler, “The Theory of Economic Regulation,” Bell Journal of Economics and Management Science 2, no. 1 (1971): 3–21.

7. Regulation (EU) 2026/1744 (the “AI Omnibus”), published 24 July 2026, in force 27 July 2026; Annex III high-risk duties from 2 December 2027, product-embedded high-risk systems from 2 August 2028 (National Law Review, 31 July 2026).

8. Ministry of Electronics and Information Technology, India AI Governance Guidelines, 5 November 2025; summarised by AZB & Partners, 11 November 2025.

9. Moffatt v. Air Canada, 2024 BCCRT 149 (Civil Resolution Tribunal of British Columbia, 14 February 2024), para. 27; see Chapter 8.

10. For example Mata v. Avianca, No. 22-cv-01461 (S.D.N.Y. 22 June 2023, Judge P. Kevin Castel): the lawyers and their firm were ordered to pay $5,000 (LawNext, June 2023); for the count of such cases, the Charlotin database, as in Chapter 8.

11. Kauṭilya, Arthaśāstra IV.1, “Protection of Artisans,” in Book IV, “The Removal of Thorns,” trans. R. Shamasastry, Kautilya’s Arthashastra (Bangalore: Government Press, 1915).

12. Guido Calabresi, “Optimal Deterrence and Accidents,” Yale Law Journal 84 (1975): 656–671, at 666; the book is The Costs of Accidents: A Legal and Economic Analysis (New Haven: Yale University Press, 1970).

13. Directive (EU) 2024/2853 on liability for defective products (23 October 2024): software, including AI systems, counts as a product; to be transposed by 9 December 2026. Article 2(1) as corrected by the corrigendum of 7 May 2026: it applies to products “placed on the market or put into service after 8 December 2026”. Article 6(1): compensation only for death or personal injury (including medically recognised damage to psychological health); damage to property other than the defective product itself and property used exclusively for professional purposes; and destruction or corruption of data not used for professional purposes.

14. European Commission, notice of withdrawal of the proposed AI Liability Directive, Official Journal C/2025/5423, 6 October 2025.

15. “Insurers launch cover for losses caused by AI chatbot errors,” reported 11 May 2025 (Armilla AI’s reprint of the Financial Times report): policies underwritten by several Lloyd’s insurers, including Chaucer, covering damages and legal costs when an AI tool performs below expectations.

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 the strongest moats “are built from things that require consciousness”; that trust “is a conscious phenomenon. It requires two selves” and that “An AI can simulate reliability. It cannot earn trust”; that distribution has a “consciousness requirement”; that trust “holds longest”, that the strongest moats “have become more valuable”, that “worlds cannot be replicated on a timeline that matters” and that Rajan’s knowledge of enforcers “cannot be downloaded”; that a machine “cannot build” these moats because “it lacks the capacity to be at risk … It has nothing to lose”; that “the exposure was the moat”; that Rajan’s calls “with no transaction in mind” were “investments in distribution whose compound returns Rajan will never trace to their source”; that the ledger’s data was the founder’s moat; that Rajan’s hands were “the purest form of this moat”; that “the value is shifting from generating work to guaranteeing it”; that companies should make regulators “consult rather than constrain” them, and that “Rajan knows the people who enforce them”; that the call to Senthil “was the competitive advantage”. It also never asked whose wages would pay for a matched price; made Rajan a broker who sources fabric and clears customs; had Weiss learn of platforms for the first time; told the speciality line and the rollout in the wrong order, with a buyer’s call from Tokyo; gave Rajan “twenty-seven years” in 2026 and Weiss “a decade”; gave Senthil and Rajan “shared meals” and “a factory fire in 2014” in place of Chapter 7’s years of doubt; gave Senthil a daughter applying to engineering colleges, an echo of Rajan’s own daughter; named the founder’s companies ClarityAI and TradeGrid (both real company names) and a Stuttgart buyer Werner Hoffmann; dated the moat image to Buffett’s “shareholder letters of the late 1980s”; and described the EU’s AI Act and India’s guidelines as they stood before the changes of 2025 and 2026. Its section titles made each subject a moat: “The Moat You Build by Listening”, “by Connecting”, “by Being Chosen”, “by Being Believed”, “by Understanding Power”, “by Belonging”, and “The Physical Moat: Atoms vs. Bits”; this edition retitles 15.1 to 15.7 and keeps 15.8’s title, argued with.

• What I now hold: the competitor is never a machine but a company using machines, with its own balance sheet, insurer and staff; the contest is between kinds of answering. By Buffett’s own test (“A moat that must be continuously rebuilt will eventually be no moat at all”; a business that depends on a great manager is excluded), what Rajan has is not a moat but a practice; the title is kept and argued with. A ledger is valuable because each entry has someone behind it (Chapter 8’s testimony); its knowledge was other people’s, and Chapter 20 asks what they are owed. A network effect is a reason to stay that does not depend on the company being good; a network that writes about outsiders owes them a reply. Being chosen through recommendation is honest about a habit and favours the already known (Chapter 7.6). The Kuṟaḷ, on its modern Tamil commentators’ reading, makes honest trade a virtue; Smith makes it a structure, which a platform has in the highest degree; the English translators make the verse a promise. Rules reward records, not acquaintance; friendship with enforcers is capture (Stigler). Belonging used as a moat is lock-in, and it leans (Kuṟaḷ 118). Atoms slow every rival alike; looms bring control and obligations, and a price cut falls first on the floor. Kauṭilya lays liability on the hand that did the work; Calabresi on whoever is best placed to weigh what failure costs against what preventing it costs, and the law is moving partly the way his question points. A signature without time to check is a crumple zone. A model cannot sell its own answering; a company can sell answering for its machines; one person can sell only his own. Answering is a reason for some buyers, for some cloth, at some price, and never a right to be chosen. A company can do much of what Rajan’s answering does (tell first, explain, replace, pay, give a named person); what it cannot do is make his history its own, and that is a fact about who he is, not a better assurance or a capacity a machine lacks. Consent to write is not consent to every use. A right of reply works only if the person knows, can read and can reply safely. A recommendation carries a field of view. Smith and the Kuṟaḷ are not a choice between virtue and self-interest, and neither wins. Documentation can make conditions visible; a small firm needs simpler means, not an exemption. Responsibility follows control: the hospital answers too. The two failures were different in kind, and a season is informative, not a test. Calling first gives a reason this season; it does not make a business deserve to exist. Whoever runs the machine, the buyer still chooses whom to trust. A machine’s summary of signed entries is the company’s statement. A reply is not a correction. Part of a habit can be written into terms of sale, backed by the seller’s own money. Calabresi’s question can point at the signer, where checking is cheap. One person’s answering ends with him; a firm that outlasts its founder answers as a company does.

• What changed my mind: two readers’ warnings that “the person who answers” could become a new residue that cannot be copied; Buffett’s 2007 letter, read in full; Chapter 3’s correction of “nothing to lose”; Chapter 5’s line between liability and answering; Chapter 7’s account of Senthil and of the kinship premium; Chapter 8’s testimony and Moffatt; Chapter 9’s canon for Weiss and the platform; the Tamil commentators on Kuṟaḷ 120; Smith’s lectures; Kauṭilya IV.1 and Calabresi; the Lloyd’s policies of 2025; the EU’s product liability directive and the withdrawal of its AI liability proposal; the audit’s canon and fact findings; the discovery that TradeGrid is a real Nigerian marketplace.

• What survives: Jürgen’s call and “I owe you an honest conversation”; “the next one will be harder for me to refuse”; the Stuttgart buyer’s “when something goes wrong he calls you before you call him”; the handwritten note; Balogun’s knowledge “gone the moment it is said”; consent built into the record; the Accra trader and the dye house in Lomé; passengers and members; “A moat is a rate of deepening, not a wall”, now read against Buffett; the eight subjects and 15.8’s liability shield, narrowed; the shipping agent who held a container; the February dusk, the knees and the monsoon.

Next in the series: Chapter 16: Embed AI Deeply, Don’t Bolt It On.