top of page

AI-Proof: What Machines Can Never Be — Volume II: Application, Part V: Conscious Strategy, Chapter 22: Own the Layer AI Cannot Disintermediate

6 hours ago
39 min read

Part V asks what is worth committing to when every move is cheap. Chapter 21 set a decade against a week. This second chapter asks what stands between people when a machine can connect them directly, and who keeps it.

Rajan’s nephew said machines would cut out every middleman, Rajan included. Rajan is no middleman: he weaves. But the machines were not emptying the middle. They were moving into it, onto ground that nobody owns.

In Brief: For readers who stand between a maker and a buyer, or depend on someone who does, and wonder what a machine leaves them. Chapter 22 asks what is left between people when a machine can stand between any buyer and any seller. A cloth maker in Coimbatore hears his nephew’s case for cutting out the middle; a designer in Kumasi watches machines make “Kente-inspired” patterns; a customs officer in Chennai, a venture partner in Singapore and a hiring screen in Hyderabad show who stands in the middle now. Kālidāsa’s cloud messenger, Bruno Latour’s mediators, Kenneth Arrow on trust and Max Weber on bureaucracy help answer what no machine can take out of a deal, and why nobody owns what every middle stands on.

Machines move into the middle rather than empty it; a party is not a middle, though it can be replaced, and the trust beneath every middle is kept, never owned.

Opening: The Sunday Call from Singapore

Rajan’s nephew called it disintermediation, and Rajan told him he was wrong.

Not about the word. Karthik explained the word clearly enough. To disintermediate is to remove the middleman: whoever stands between the one who makes a thing and the one who buys it. Machines, he said, would soon find the mills, compare the prices of a thousand suppliers, draft the shipping papers and write to buyers in their own languages. Work that consisted of finding, checking and passing things on would be done by software, quickly and for very little.

It was a Sunday in November 2025. Karthik had joined a venture firm in Singapore in July, and he still telephoned his uncle every Sunday. This time he had asked for a video call, so that he could show a chart. Rajan Subramanian watched it in his office above his factory in Coimbatore. He has run the business since 1997. About 140 people weave cloth and finish it by hand on the floor below, for buyers in fourteen countries. Dye houses in Tiruppur dye the yarn, and some of the cloth goes to a finishing house in Karur.

“It will happen to everything in the middle,” Karthik said.

“Everything,” Rajan said. “Including me.”

“That is what I am telling you, Uncle.”

Rajan had his reply ready. “You cannot download a bale of cotton,” he said. “The machines live in computers. I live in the world.”

He believed it. He was half right.

He was right about himself. Rajan is not a middleman. He does not pass cloth from one person to another. He makes it, on his own looms, and his name goes with it. But the machines did not live only in computers. In March 2025 Rajan had ordered every system the vendors offered in one phone call, and since April cameras had watched the cloth on his own looms. On standard cloth they caught broken threads that tired eyes missed on the night shift.

Karthik was half right too, and he had the better example. Jürgen Weiss, who has bought cloth from Rajan for a firm in Hamburg since 2013, had bought his standard weaves through a sourcing platform since late 2024. Its software matched his orders to mills across South India. It booked their looms, inspected the cloth with cameras at the mill and shipped real bales to Hamburg. You cannot download a bale of cotton. The platform did not need to.

Karthik’s case deserves its full strength before its weakness. Middlemen exist because dealing with strangers costs something. A buyer must find a seller, check that the cloth is what was promised, and trust that the money and the goods will both arrive. Agents, brokers and buying houses grew up to carry those costs, and they charged for carrying them. Machines lower each cost. They find in seconds, they inspect with cameras, and they keep records that a stranger can read. Some middlemen lived on what their buyers and sellers did not know. When a machine tells them, the gain is real, and the loss falls mostly on the middleman. Some middles deserve to go.

What Karthik had wrong was what the platform was. It had not removed a middle. It was one. It stood between Weiss and mills he would never visit, and it was paid for standing there. It could also become a stronger middle than the agents it replaced. An agent knew a few buyers and a few mills. The platform saw every order that passed through it, and it set its own fee.

The platform is the usual case. Machines do not empty the middle. They move into it. By 2025 Google answered many searches with a summary written by a machine. Pew Research Center followed the browsing of 900 American adults in March 2025. When a summary appeared, they clicked a traditional result in 8 per cent of visits, against about 15 per cent without one. They clicked a link inside the summary in 1 per cent.¹ The summary now stands between readers and the people who wrote what it summarises. In September 2025 OpenAI let people in the United States buy goods inside ChatGPT, starting with sellers on Etsy. Its announcement said how it would be paid: “Merchants pay a small fee on completed purchases.” It also kept the seller as “the merchant of record”, who takes the payment and sends the goods.² The middle had a new occupant, and the seller still stood behind the goods. In March 2026 OpenAI scaled the feature back and moved checkout into retailers’ own apps inside ChatGPT. ChatGPT stayed in the middle, as the place where buyers found the goods.

So this chapter asks: When a machine can stand between any buyer and any seller, is anything left between people that cannot be taken out, and can anyone own it?

The print edition’s reply was six “layers” that AI cannot disintermediate: the physical world, community, governance, capital, trust and being chosen. It rested five of them on consciousness, and the sixth, the physical world, on atoms. Governance needed “a self”, it said, and “AI does not have one”. Being chosen was “consciousness recognising consciousness”. This edition keeps the six subjects and tests each one. It does not keep the reason. The six are not of one kind: one is a matter of things, others are institutions and relations, and one is another person’s choice. What they share is that each is a place where a middle can stand, and where it matters who stands there.

The title also needs an argument before the sections begin, because it claims too much twice. The first word to test is “cannot”. Most of what the print edition said a machine cannot do was a claim about doing, and a claim about doing carries a date. One of its sentences said that no machine could walk a mill floor and feel whether the looms were running true. A few lines later it described sensors that already listen to machines in factories.

Only one “cannot” in the chapter holds without a date, and it was hidden near the end: “You cannot disintermediate someone who is not intermediate.” That is true by meaning. To disintermediate is to remove a middle, and a party to a deal, one of the people it is made between, is not a middle. Rajan makes the cloth, so no machine can cut him out of a chain in which he stands at one end.

But look at what that protects. A party cannot be disintermediated; he can be replaced. The platform did not cut Rajan out of anything. It offered Weiss another maker, one whose machines did much of the work. This book calls that substitution: work a person was paid for goes to machines, here a rival’s. Being replaced by a rival is the oldest threat in trade. The title’s “cannot” is true, and it protects no one.

The second word is “own”. The layers the print edition praised are relations: trust, belonging, the standing to decide. Nobody owns a relation. A person holds a part in one: the promises he has made and keeps, his name on what he sends. That part can be kept or spent. It cannot be handed to someone else, and it guarantees no customer.

In a book of 2005, the French sociologist Bruno Latour set out a distinction that helps with the rest. “An intermediary, in my vocabulary, is what transports meaning or force without transformation: defining its inputs is enough to define its outputs.” A mediator is different: “Mediators transform, translate, distort, and modify the meaning or the elements they are supposed to carry.” One of his examples was a machine. “A properly functioning computer could be taken as a good case of a complicated intermediary,” he wrote. “But if it breaks down, a computer may turn into a horrendously complex mediator.”³ In his example a working computer passed things on. Twenty years later the machines in the middle change what they carry even when they work. A summary chooses what to keep. A platform ranks mills, and the ranking decides which mills a buyer sees. A drafted letter speaks for the man who signs it. These machines are mediators in Latour’s sense. His distinction describes; it does not say who must answer. That step is this book’s: a mediator’s changes have to be answered for by someone, because the people at either end receive the changed thing, not the original. When a summary drops the sentence that mattered, the reader acts on what was left. When a ranking hides a good mill, the buyer never knows what he missed. That is why the middle is worth occupying. It is also why it matters who stands behind it.

22.1 Physical Infrastructure: Bales, and Who Carries the Risk

Chapter 15 showed why the atoms give no one an advantage: they slow every rival alike, the platform’s mills included. What they leave open is who carries their risk on the way.

The physical middle is made of people who carry that risk. Years ago a shipping agent in Chennai held a container for Rajan, because he knew Rajan would pay. The agent was a middle, and what he passed on was not information. For some days the container was on his account. Machines can now plan most of what such an agent does. They book the ship, fill in the forms and follow the box across the sea. When the box is lost or late, someone still pays, and someone must answer to the buyer. On the platform’s lots it is the platform, with an insurer behind it. On Rajan’s own lots it is Rajan. For the container in Chennai it was the agent.

So the physical middle splits in two. The carrying of information goes to machines first and fastest. The carrying of risk stays with whoever has agreed to bear it. No machine bears anything on its own account; its makers and users do. A machine can be set to pay a claim at once, from an account someone filled, on terms someone wrote. The machine makes the payment, and the loss is still theirs. The middle that matters is the one where someone has agreed to pay when things go wrong. Paying is not the whole of it. An insurer can pay for a lost bale, but it cannot give back a missed season.

Not every new middle agrees to pay. The checkout inside ChatGPT took a fee on each purchase and left the payment and the goods with the seller. Weiss’s platform took its fee and carried the risk of its own lots. So the question to put to a new middle is not only whether it is useful. It is whether its toll, the fee it takes for standing in the middle, comes with its share of the risk.

Rajan’s own floor shows the rest. His floor supervisor, Muthu, had run the floor for twenty years. On standard cloth the cameras were right more often than he was, and by October he had stopped overruling them there. On the speciality cloth, which is finished by hand and meant to be a little uneven, most of Muthu’s overrules held. The atoms decided neither case. What the cameras had been trained to count as a fault decided one. Muthu’s right to argue with them decided the other, and it was a right he could use. It came in August with half an hour a shift set aside for it, and Rajan’s word that a reasoned overrule would not be held against him. Without the half hour it would have been a right on paper.

22.2 Community and Identity: Who May Speak for a Cloth

The question that matters here is not whether belonging can be automated. It is standing: who may say what a cloth means.

Efua Baidoo runs a collective of eighteen designers in Kumasi, in Ghana. She was born on the coast, among the Fante, and moved to Kumasi for the cloth. The collective makes garments and furnishings from strips of Kente woven by master weavers in Bonwire, the town near Kumasi where Kente is said to have begun.⁴ Most of its buyers are in London, Milan and Cape Town.

Kente is woven in narrow strips on a loom, and the strips are sewn together. Its patterns have names. A museum in Kansas City holds an early twentieth-century cloth in red, yellow and green that its record identifies as Oyokoman. Before the twentieth century, the record says, the design was reserved for the Oyoko clan, whose members were “considered the sole legitimate heirs to the Asante monarchy”.⁵ Weaving Kente has traditionally been men’s work. A study of Bonwire published in 2023 found that “a traditional ban on women entering the loom has proved difficult to overturn conclusively”: more women weave there than ever before, and they still face pressure to stop.⁶

In 2025 Efua came across three design platforms that would generate “Kente-inspired” patterns in seconds, in colours chosen for buyers in Europe and America. Some were beautiful. Her first thought was that they were hollow: a machine could copy the look, but it could not know what the patterns meant.

That thought was half wrong, and the wrong half matters. Much of what a pattern means is written down. Museum records, books and the collective’s own leaflets give the names and the sayings that go with them. As of 2025 a machine with search could find them in seconds, and some of Efua’s younger designers knew less about the names than a machine’s summary did. Her pragmatists, as she called them, wanted the collective to use machines too: for its catalogues, for colourways shown to buyers before any strip was woven, for letters in Italian. On those tasks they were right, and she let them.

What the platforms skipped was not knowledge. The word “knowledge” was hiding three things. First, there is what is written down, which can be retrieved, and a machine can be taught it. Second, there is standing: who may say what a use of a pattern means now, and whether it is right. Third, there are the hands: which weavers in Bonwire wove which strips. The platforms had the first. They had none of the second. They could not have the third, because nobody had woven their patterns at all.

So Efua did not try to stop the machines. She asked three master weavers from Bonwire, men in their seventies, to sit with two of her designers as a panel. The panel does not teach. It says which named patterns may be used for what, and it gives its reasons in writing. A buyer, a weaver or one of the designers can dispute a reason, and the panel must reply. Each piece now carries a short record: the pattern’s name, who wove the strips, and what the panel said about its use. Chapter 20 argued that a record of where a thing came from is testimony. It is worth as much as the people who can be asked about it, and a label nobody may challenge is only theatre. The dispute is there for that reason.

It was tested in February 2026. A hotel group in Cape Town asked for bedspreads in the Oyokoman pattern. The panel refused, in writing: in its view a design once kept for a royal clan should be worn, not slept under by strangers. The buyer disputed the reason. The design was no longer reserved, she wrote, and a museum’s own record said so. The panel replied that it had not said the design was reserved, only that this use was wrong, and it offered three other patterns. The youngest of the three master weavers wrote that he would have allowed the bedspreads. The buyer took one of the other patterns, and a smaller order. Whether the panel was right, Efua does not know. She knows that it gave its reasons, that the buyer could argue, and that the order was smaller.

It costs something. The panel is paid from the collective’s margin. The weavers are paid more per strip, because their names are on the record. The work is slower. Some buyers have gone to the platforms, which are cheaper, and those buyers were not wrong. For a shop’s cushion covers, a pattern that looks like Kente may be all a buyer wants. That is substitution again, and it is a fair choice.

A community is not an intermediary. It is the thing itself. The platforms treated the weavers as a middle between a buyer and a look, and skipped them. For Kente, the weavers and the people who wear it are parties, not a middle, and the meaning of a named cloth is theirs to state. A community can still lean. It favours its own families and its own towns, and if it were used to keep buyers from leaving, it would be lock-in by a warmer name.

But the panel is not the community, and its standing has limits that should be said. Efua chose it. Its three weavers are men, in a town where women now weave against pressure to stop, and nobody who wears the cloth far from Bonwire sits on the panel. The youngest weaver’s dissent showed that even three masters do not speak with one voice. What the panel decides is narrower than what a cloth means: it decides what this collective will make and put its name to. A buyer it refuses can go to another weaver. If its record ever became the only proof that buyers would accept, the panel would become a gate, and it would have to treat every buyer and weaver alike.

Ghana’s law has its own reply to the question of who may speak. Its Copyright Act of 2005 says that folklore “includes kente and adinkra designs, where the author of the designs are not known”. It vests the rights in folklore not in the towns that make it but in the President, “on behalf of and in trust for the people of the Republic”. It forbids anyone to sell copies of works of folklore in Ghana, made there or abroad, “without the permission in writing of the National Folklore Board”.⁷ That is an institution standing in the middle by law. As written, it reaches sales in Ghana. It is hard to see how it could reach a platform that sells patterns to a buyer in Milan. Law and standing answer different questions. The panel’s refusal had no force in law; it was a reason, given to a buyer who was free to reject it.

Did the platforms learn from photographs of Bonwire’s cloth? Efua cannot know, and nobody could tell her. If they did, what Bonwire’s weavers made over generations is earning money for people who never wove it. What is owed to the people whose work trains a machine is a question this book has asked before, and the law has not yet settled it.

In October 2026 the collective is still trading. Its sales are below its best year and above its worst. Its buyers may value the record on each piece, or only like reading it. Efua does not know which.

22.3 Governance Systems: Power Requires Legitimacy

Rajan once spent three days in Chennai sorting out a problem that should not have existed.

It was 2019. His papers for a shipment to Hamburg were in order: the invoice, the packing list, the shipping bill that an exporter files with customs. A new officer at the port, transferred two weeks earlier from Visakhapatnam, would not let the shipment go without a laboratory report on the fibre content of every lot. The officer before him had never asked for one. The rule could be read either way. Delay meant that Weiss would miss a season, and might not order the next year.

Rajan did not call a lawyer. He went to Chennai and sat with the officer. He brought the earlier assessments, not as a weapon but as a place to start. The officer explained his caution. At his last port he had worked on a fraud in which an exporter declared one fibre and shipped another, to claim a larger refund on exports. Rajan understood. He offered test reports from an accredited laboratory for each lot, and the officer accepted them. The shipment moved. Weiss ordered again the next year.

The print edition read this story as proof that governance needs “a self”. The solution, it said, came in “the space between what the rules said and what two conscious beings, negotiating in good faith, decided the rules should mean”. That sentence should worry anyone. An exporter and an officer deciding in private what a rule means is not governance. It is the place where favour lives.

What made the meeting legitimate was something else, and it can be named. The officer had authority under the law to ask for the reports, and he had to answer for asking. He gave his reason. Rajan could have asked for the demand in writing and taken it higher. And Rajan did his own part: he explained his cloth and supplied what the officer’s caution needed. Here legitimacy was a chain of reasons that could be checked, and a way to correct the decision if it was wrong. It did not depend on anyone’s inner life.

It did depend on access. Rajan could spend three days at the port. He had dealt with its officers for years. A young weaver with one shipment and no name in Chennai could have done neither. The meeting that cleared Rajan’s cloth was a privilege of the known. Chapter 15 called where such favour leads regulatory capture, after George Stigler’s argument of 1971.⁸ Nobody in that office intended capture. Meetings like it are where it begins.

In 2020 India’s customs moved the other way: by 31 October, imports were assessed by officers the importer never meets, everywhere except land customs stations and diamond bourses. One aim, the circular said, was “reduced interface between trade and Customs officers”; the reform covered imports, not exports like Rajan’s, but its reason is the one his visit raises.⁹

The German sociologist Max Weber described the ideal behind such a reform a century earlier. Bureaucracy’s nature, he wrote, “which is welcomed by capitalism”, develops more fully the more it is “dehumanized”: the more it removes from official business “love, hatred, and all purely personal” feeling, the “elements which escape calculation”. This, he said, “is appraised as its special virtue”.¹⁰ A faceless assessment is the ideal Weber described, with a computer in the middle. A machine that applies a rule in the same way to every exporter would be its next step. For the exporter nobody knows, it may be fairer than Rajan’s three days. That is the strongest point against the print edition here, and it should be said plainly. The circular states an aim, not a result; whether faceless assessment has been fairer in practice is a separate question, which this chapter does not answer. The faceless way has costs of its own. An officer who is never met cannot be charmed, and he cannot be shown a bolt of cloth either. Everything must go into writing, and a small exporter who writes badly is a stranger of a new kind. A rule applied in the same way to everyone can still be the wrong rule, or weigh most on those least able to meet it. And a reform that ends the private meeting can replace it with a rule that nobody sees.

But a faceless system still owes what made the meeting legitimate. It must give its reasons, and there must be a named place where a wrong decision can be challenged and corrected, and where someone must answer for it. Chapter 15 said that a right of reply works only if the person knows that a claim about her exists, can read it and can reply without risk. A decision is such a claim, and its reason is what the exporter must be able to read. A decision with no face can be fair. A decision that nobody will explain cannot be checked, so nobody can know whether it is fair. Machines make this harder. When an officer follows a machine’s score, the officer, the department that chose the machine and the firm that built it can each point to the others. A machine’s account of why it flagged a shipment is not a reason that anyone answers for. An officer with authority must give the reason and answer for it.

Set beside Efua’s panel, this can look like a contradiction. The panel has names and faces, and the reform takes faces away. But the two stand in different places. The panel decides what one collective will make, and a buyer it refuses can go elsewhere. Customs is a gate that every exporter must pass, and there is no other customs to choose. A party may choose whom it deals with. A gate must treat strangers alike. Both owe their reasons, and both must hear a challenge. The rule reaches the machine middles too. While buyers like Weiss can still go to a mill directly, the platform is one middle among several. It may choose which mills it lists, as any trader chooses whom to deal with. If it became the only road from these mills to buyers abroad, it would be a gate. It would then have to treat every mill alike, and owe each its reasons for a ranking and a way to challenge one.

None of this means that officers should hear no exporters. Rules are better when the people who live under them are heard. What matters is whether the hearing is open to others and written down, or private. Power does require legitimacy. But legitimacy is given, by the law and by those who accept its reasons. Nobody owns it, least of all the exporter who knows the officer.

22.4 Capital Allocation: A Promise, Not Only a Forecast

Money follows many things, and judgement is only one of them.

Before the call ended, Karthik talked about his own work, because it was the other half of his argument. His firm used a screening model that read the decks and accounts of hundreds of companies a month and ranked them by measures the fund had chosen. It was fast and consistent. On most of the questions it was asked, Karthik thought, it did better than most of the firm’s people. His own job, four months in, was to write the notes the partners read.

That autumn the model had ranked a company in Jakarta near the top. The company matched shops that needed goods delivered with the owners of small lorries, and it took a share of each trip. Its numbers were good. The partner who would have led the investment went to Jakarta for two days. She came back and said she would not invest yet. Many of the lorry owners on the platform, she said, were the same few brokers who had always run the trade, now working through an app. They had survived the app, she thought, because they did what it did not. They advanced fuel money to drivers, vouched to shops for owners nobody knew, and paid when a lorry broke down on the road. The app had taken over the matching. The credit, the vouching and the loss had stayed with the brokers. The middle had not gone. It had moved inside the company, and it could move out again. She could not show this in the company’s data, and she said so. The partners passed. Another fund invested a month later.

The model was itself a middle. It stood between the partners and hundreds of companies they would never read, and it decided which ones they saw first. In Latour’s sense it was a mediator: it did not pass the companies on unchanged, it ranked them. The partner’s two days in Jakarta were a way of going round it, to look at what the ranking had been built from.

In November 2025 nobody could know who was right, and Karthik did not pretend to. Part of the partner’s view was evidence: she had stood in the brokers’ yards and counted the lorries. Part of it was a hunch, in a field where results come slowly and noisily. The psychologists Daniel Kahneman and Gary Klein argued when a hunch deserves trust.¹¹ The world must be regular enough to be learned, and the person must have had long practice with feedback. By their test, as I apply it, venture investing offers little of either. The model had evidence too, more of it than she had. An investor in Kampala, Lillian Adeke, once argued for a company her fund’s screen had rejected, as Chapter 7 told. This partner argued against a company her screen liked. It is the same act: saying what you cannot prove, to people who decide.

The print edition gave the doubt to Karthik and called it “an act of consciousness”. That is not what made it count. What made it count was who would answer for the decision. The model could rank. It could not commit the fund’s money, because it was not the fund.

A decision to invest is not a prediction. It is a promise. The forecast can be the machine’s, and often it should be. The commitment belongs to the partners, and they make it with other people’s money. They answer to the fund’s own investors, to whom they must explain each bet, and each bet they refused. If the partner was wrong about Jakarta, she will have to explain to her partners why the fund missed a good company. What the founders get from a fund, if it keeps its word, is more than money: a party that will still be there when the numbers turn.

So the money middle splits, as the physical one did. Matching money to companies is moving to machines, as finding mills did. A fund stays a party because it commits its investors’ money and must explain each commitment to them. Whether the commitment is also a promise to the founders depends on what the fund does later, when the company is in trouble. That is not settled on the day the money arrives.

22.5 Trust Architecture: The Layer Beneath Everything

You can rely on a calculator to produce correct arithmetic. You do not trust it. Chapter 3 described trust, with the philosopher Annette Baier, as accepting that someone could harm you, while not expecting them to.¹² Trust in a machine, it added, is really trust in the people behind it.

This section’s title needs two corrections. “Architecture” suggests that someone designs trust. Nobody does; many people keep it at once. What can be designed are the conditions that make trust reasonable: records, reasons and a place to complain. And “everything” is too wide. Trust is the layer beneath every middle, human or machine.

Start with a poem. In the Meghadūta, “the cloud messenger”, the Sanskrit poet Kālidāsa, who probably wrote in the fifth century,¹³ tells of a yakṣa, an attendant of Kubera, the god of wealth. For neglecting his duty he has been exiled for a year to a hill far from his wife. When the rains come, he sees the first cloud on the hill and asks it to carry a message north to her.¹⁴ Then the poet stops to say what is strange about this:

dhūmajyotiḥsalilamarutāṃ sannipātaḥ kva meghaḥ / sandeśārthāḥ kva paṭukaraṇaiḥ prāṇibhiḥ prāpaṇīyāḥ | / ity autsukyād aparigaṇayan guhyakas taṃ yayāce / kāmārtā hi prakṛtikṛpaṇāś cetanācetaneṣu ||

The verse sets “a cloud, a mixture of smoke, light, water and wind” against messages, which need living messengers. In M. R. Kale’s translation, they can only be conveyed by “beings with organs of sense capable of discharging their functions”. The yakṣa, in his eagerness, did not consider this, “For those that are sorely affected by love are naturally incapable of distinguishing between objects, sentient and insentient.”¹⁵ Arthur Ryder, translating in 1912, put the same thought in rhyme: no lover “could discriminate / ’Twixt life and lifeless things, in his love-blinded state.”¹⁴ The commentator Mallinātha, writing centuries later, glosses the poet’s word for such lovers, prakṛtikṛpaṇa, as svabhāvadīna: in my rendering, by nature helpless.¹⁶

As I read it, the verse draws its line at life. Messages need living messengers, and the yakṣa’s request is an error that his longing explains. Latour would draw no line there. For him, whatever changes what it carries is a mediator, living or not. On the doing, as of 2026, Latour is right. The machines in today’s middles carry messages, summarise them, rank them and choose which ones a reader sees. And the poem does not mock the yakṣa. It gives the cloud the whole route north, river by river and city by city, and then the message, as if the carrying could be done.

Near its end, as I read it, the poem answers its own doubt. The yakṣa fears that rumour may have made his wife doubt him, and a cloud’s word alone would not settle that. So he gives the cloud something to say that only he could know. One night, asleep with her arms round his neck, she woke crying aloud. When he asked her why, again and again, she told him with a hidden laugh that she had dreamed of him with another woman.¹⁵ In a note, Ryder calls this “a secret which, as it could not possibly be known to a third person, assures her that the cloud is a true messenger.”¹⁴ The next verse calls it a token of recognition, abhijñāna:

etasmān māṃ kuśalinam abhijñānadānād viditvā / mā kaulīnād asitanayane mayy aviśvāsinī bhūḥ

In my rendering: “Knowing from this token that I am well, do not, dark-eyed one, let rumour make you distrust me.”¹⁵ Ryder has “Then let no gossip thy suspicions move”.¹⁴ The cloud does not need to be trusted. The message carries a mark of the one who sent it, so her trust can pass the carrier and reach him.

The two lamps answer different questions. Latour asks what a carrier does. Kālidāsa’s verse asks what a message needs. This book draws its line in neither place. It does not draw it at life, because nothing in its argument depends on whether a machine experiences anything. It does not draw it at what machines can do, because that changes every year. It draws it at what a message is. A message belongs to someone, and someone must answer if it arrives wrong. The yakṣa’s words stay his, whatever carries them; if the cloud rained them on the wrong house, the fault would be his to put right. The poem’s own answer is the token. What a message needs from whatever carries it is not life but a way back to the one who stands behind it. The poem is about an exiled lover and a cloud, and the use I make of it is mine.

Read literally, Mallinātha’s gloss allows a second reading. It is mine, not the translators’. A lover may know perfectly well that a cloud is not a person, and ask it anyway, because longing asks whatever is near. The token fits that reading: the exile knows that words brought by a cloud need proof that they are his, and he provides it. The second reading fits the present better. People who send their letters, purchases and introductions through machines mostly know what the machines are. They use them because they are near and cheap, and that is often wise. The danger lies only in forgetting whose words they carry.

What every middle stands on is the trust of the parties around it. The American economist Kenneth Arrow wrote in 1974: “Trust is an important lubricant of a social system. It is extremely efficient; it saves a lot of trouble to have a fair degree of reliance on other people’s word. Unfortunately this is not a commodity which can be bought very easily. If you have to buy it, you already have some doubts about what you have bought.”¹⁷

What can be bought is assurance. Weiss can buy a written guarantee with an insurer behind it, and for cloth from mills he will never visit, that is sensible. Such a guarantee is one of the institutions that, in the economist George Akerlof’s words of 1970, “arise to counteract the effects of quality uncertainty”.¹⁸ It is what people buy when they have doubts, which is Arrow’s point. A guarantee does not make a stranger trustworthy. It lets a buyer deal with him while needing less trust. A platform with a written promise can let a buyer in Hamburg deal with a mill he has never seen. In that way a machine middle can make room for trade between strangers, not only spend the trust already there. It fits how Weiss buys: his standard weaves through the platform, his speciality pieces from Rajan. A standard weave can be written into a guarantee: its count, its weight, its faults per metre. A cloth meant to be a little uneven is harder to write down, because on it a fault and a beauty can look alike. That was the cloth on which Muthu’s overrules held. For it, Weiss relies on Rajan’s name: on trust, not only on assurance.

Trust itself cannot be bought, and nobody who stands on it owns it. Every middle, human or machine, draws on it. A machine in the middle is believed as far as someone stands behind it, so it stands on this ground too. A summary that misreads its sources spends the trust that their writers had earned. Unlike the exile, who chose his cloud, the writers never chose that carrier, so the summary’s makers owe an answer of their own. Since April 2025 Rajan’s letters to his buyers have been drafted by a machine, and he signs them. If one of them promises a date his floor cannot meet, the buyer will not blame the machine. The letter spends Rajan’s trust, not its software’s. Those who answer for what went wrong, by explaining it and putting it right, keep it. Trust is not always good, either. A trader who is trusted is checked less, and trust that nobody checks is where favour lives.

In business talk, a moat is whatever keeps rivals away, like the water round a castle. Trust is not a moat around the castle. Trust is the ground the castle stands on. Nobody owns the ground. Everyone in a trade walks on it and wears it, and some help to keep it. You cannot own the layer the machines depend on. You can keep your part of it.

22.6 Distribution for Employees: Being Chosen

As of 2026 the first choice is often made by a machine, not a human.

The sections so far speak to people who own a floor, run a collective, clear goods or commit a fund’s money. An employee may fairly ask where her layer is. She owns no looms. She signs no customs papers. She commits nobody’s capital.

The honest reply is that an employee owns no layer at all. She is a party to her own work, because her name is on it, and she answers for it. Inside her firm she is usually a middle: between one team and another, between a machine’s draft and the client who reads it. Machines move into those middles as they move into every other. An assistant who once wrote her manager’s letters to clients may now check the letters a machine drafts. She still stands between the firm and its clients. What she passes on is the machine’s words, and whether she has the time to read them is her employer’s decision. If she signs without that time, she becomes what Chapter 1 called a moral crumple zone. Muthu’s right to argue with the cameras worked in part because it came with the half hour.

The print edition’s reply was that her layer is being chosen: the people who decide will choose her because they have watched her think. Being chosen is real, but it is someone else’s act. It runs along old lines, towards the people who are already known. And it now often begins with a machine. In June 2024, at a pharmaceutical company in Hyderabad, a hiring screen rejected a mid-career scientist whom the head of research had recommended after two hours of conversation. The error was seen only because she had a sponsor. Nobody looked again at the candidates who had none. A sponsor is a human middle who vouches. The screen was a machine middle that ranked. A candidate with neither was not seen at all.

So what can an employee keep? Not a position. A position is given or withheld by the people who employ her, and knowing the work well fills a position without granting one. She can keep her name on her work, the explanation she gives when it fails, and the habit of putting right what she can. Those are hers. They keep no job, and nobody may notice them.

Her employer has the other half. A company that lets a machine make the first cut owes the people it cuts a way to be seen by a person, and a way to dispute the result. It owes its own staff a straight account of which middles it is giving to machines, and when. It owes them a hearing before the decision, not after it. Who keeps what the machines save, in hours or in money, is the employer’s choice, not a fact about the machines. Chapter 18 argued both points. This section promises no career. It says only where an employee stands: at one end of her own work, and in the middle of other people’s.

So the chapter’s doubt can be met, though not with the comfort the title promised. Machines do not empty the space between people. They move into it, and often they serve the people at either end better than the old middle did. What no machine can take out of a deal is a party to it, though a party can still be replaced. The one who makes the value, the one who stands between it and the buyer and the one who keeps the gain need not be the same. A new middle can take the toll while the risk stays at the ends. Beneath every middle lies the trust of the parties. It can be bought only as assurance, and nobody owns it. So the question to ask of any middle, old or new, is not whether a machine can do its work. It is who stands behind it, who carries the risk, who gives the reasons, who may challenge them, and who must answer when it misleads.

What would prove this wrong?

My main claim is about what machines do, so it carries a date. As of 2026, machines move into the middle of trade rather than empty it. The summary, the platform and the checkout each put a new party between buyer and seller, and each was paid in fees or in attention.

The claim would be wrong if buyers’ machines began to deal directly with makers’ machines, and nobody stood between them taking a toll or setting the ranking. Picture a buyer’s agent that finds mills on an open register, checks their records against a public standard and pays them through a channel nobody owns. If trade of that kind grew, the middle would be emptying.

A reader can look in three places. The first is the share of each sale taken by whoever stands between the parties, and whether that share falls over the years. The second is who sets the ranking a buyer sees, and whether its rules can be read. The third is who pays when a lot is lost. If the toll falls towards nothing, the rules are open and the risk stays with the parties, I was wrong about the middle. The claim about trust would stand either way, because the parties’ trust lies beneath every trade, with a middle or without one.

Ending: The Message from Chennai

On Monday morning a message came from Chennai. It was from the officer who had wanted the laboratory reports in 2019. He now held a senior post at the same custom house. In six years he had telephoned Rajan perhaps four times, always politely.

A change was planned, he wrote, in how certificates for exported cloth would be checked. What did exporters really think? He would value Rajan’s personal view, between the two of them.

Rajan read it twice. He knew what such a message was worth. An officer who asked his opinion would remember him, and one day a shipment of his might wait on that officer’s desk.

He wrote back before lunch. He thanked the officer. He said he would send his view in writing to his exporters’ association. The association could raise it at the export commissionerate’s trade facilitation committee, where exporters’ bodies meet the commissioner.¹⁹ He would rather the question went to every exporter, he wrote, and not only to those the officer knew. He did not say more than that, and he did not tell Karthik.

He sent it. By evening there was no reply, and Rajan could not tell what the silence meant.

Practice

Draw the middle. Choose one thing you make, sell or do for someone. On one page, draw everyone and everything between you and the person it reaches: who carries it, who vouches for it, and who pays if it fails. Mark the machines, and whoever takes a fee for standing there. Beside each, write who stands behind it, and whom you could ask why it did what it did. Then mark what in the drawing you own, and what you only keep: a promise, your name, a habit of explaining. Beside each thing you keep, note whose hours keep it. If you are employed, mark which links your employer decides, and which you could not change if you wanted to. The page will keep no customer and secure no job. It shows where you are a party and where you are a middle, and which of those a machine could take.

Words to keep

  • abhijñāna: Sanskrit, “a token of recognition”; in the Meghadūta, the private memory the yakṣa sends with the cloud so that his wife will know the message is his.

  • assurance: a promise that can be bought, such as a guarantee, an insurance policy or a brand; what people buy when they cannot yet trust.

  • disintermediation: removing the middle party between the one who makes a thing and the one who buys it.

  • faceless assessment: in Indian customs since 2020, the checking of an importer’s documents by an officer at another customs station, whom the importer never meets.

  • folklore (Ghana): in Ghana’s Copyright Act of 2005, works of the country’s traditions, which “includes kente and adinkra designs, where the author of the designs are not known”; the rights are held by the President on behalf of the people.

  • intermediary: in Latour’s sense, something that passes things on without changing them.

  • Kente: cloth woven in narrow strips, traditionally by men, in Asante and Ewe towns of Ghana; its patterns have names.

  • mediator: in Latour’s sense, something that changes what it carries.

  • Meghadūta: “the cloud messenger”, Kālidāsa’s Sanskrit poem in which an exiled yakṣa sends a message to his wife by a cloud.

  • party (to a deal): one of the people a deal is made between; a party cannot be disintermediated, though it can be replaced.

  • prakṛtikṛpaṇa: Sanskrit, “by nature helpless”; Kālidāsa’s word for lovers who ask a cloud to carry their words.

  • regulatory capture: rules that come to serve mainly the firms they are meant to govern (Stigler).

  • svabhāvadīna: Sanskrit, “helpless by one’s own nature”; Mallinātha’s gloss on prakṛtikṛpaṇa.

  • substitution: work a person was paid for is done by machines instead, his own or a rival’s; the threat a party faces, where a middle faces disintermediation.

  • yakṣa: in Indian tradition, a demigod; in the Meghadūta, an attendant of Kubera, the god of wealth.

Sources

  1. Athena Chapekis and Anna Lieb, “Google users are less likely to click on links when an AI summary appears in the results,” Pew Research Center, 22 July 2025: browsing data from 900 U.S. adults, March 2025; a traditional result was clicked in 8 per cent of visits with an AI summary and about 15 per cent without; a link inside the summary in 1 per cent.

  2. OpenAI, “Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol,” 29 September 2025 (openai.com), launched for users in the United States buying from US sellers on Etsy; reported by TechCrunch, 29 September 2025. The post calls the merchant “the merchant of record across the purchase journey” and says that orders, payments and fulfilment are handled by the merchant. In March 2026 an OpenAI spokesperson said that “Instant Checkout is moving to Apps”, retailers’ own apps inside ChatGPT (Brian Warmoth, “OpenAI shifts checkout plans in its agentic commerce strategy,” Digital Commerce 360, 6 March 2026), and OpenAI scaled the feature back (Mitchell Parton and Julia Waldow, “What went wrong with ChatGPT’s Instant Checkout,” Modern Retail, 27 March 2026, citing The Information).

  3. Bruno Latour, Reassembling the Social: An Introduction to Actor-Network-Theory (Oxford: Oxford University Press, 2005), p. 39 for the definitions of intermediary and mediator, page as cited on the New Savanna blog, 12 August 2015; the printed page itself not seen.

  4. Bernard Osei-Kwabena, “Celebrating Kente,” ModernGhana, 11 June 2005: Kente “originates from Bonwire”, near Kumasi.

  5. Nelson-Atkins Museum of Art, Kansas City, “Kente Cloth” (Oyokoman), Akan (Asante), early 20th century, strip-woven silk, object 2013.2.5, online collection record.

  6. Dede Amanor-Wilks, “Continuity and change: women’s work in the kente economy of Bonwire, Ghana,” Feminist Africa 4, no. 1 (2023), published online 15 May 2023 (feministafrica.net): the study also records the “suspension of the traditional gender taboo on weaving”.

  7. Copyright Act, 2005 (Act 690), Ghana, ss. 4(2), 17, 44(1) and 76 (GhaLII); s. 64(1) requires anyone using folklore beyond the permitted uses to apply to the National Folklore Board and pay a fee. That the Act does not reach sales abroad is my reading: s. 44(1) is limited to sale “in the Republic”, and nothing in the Act gives it effect outside Ghana.

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

  9. Central Board of Indirect Taxes and Customs, Circular No. 40/2020-Customs, 4 September 2020, para. 4 (roll-out of Faceless Assessment for imports by 31 October 2020; its aims include “reduced interface between trade and Customs officers”), and Circular No. 45/2020-Customs, 12 October 2020, Annexure I (31 October 2020: “All India – All Zones - All Imports under Faceless Assessment”, “Excluding Land Customs Stations/Diamond Bourses”); CBIC, Guidebook on Faceless Assessment, 29 October 2020, ch. 2, para. 2.1 (“an assessing officer, who is physically located at a Customs station, which is not the port of import”).

  10. Max Weber, “Bureaucracy,” in From Max Weber: Essays in Sociology, trans. and ed. H. H. Gerth and C. Wright Mills (New York: Oxford University Press, 1946), pp. 215–216 (from Wirtschaft und Gesellschaft).

  11. Daniel Kahneman and Gary Klein, “Conditions for Intuitive Expertise: A Failure to Disagree,” American Psychologist 64, no. 6 (2009): 515–526.

  12. Annette Baier, “Trust and Antitrust,” Ethics 96, no. 2 (1986): 231–260; see Chapter 3.

  13. “Kalidasa,” Encyclopaedia Britannica (online), which places him in the fifth century CE.

  14. Kālidāsa, “The Cloud-Messenger,” in Kalidasa: Translations of Shakuntala and Other Works, trans. Arthur W. Ryder (London: J. M. Dent; New York: E. P. Dutton, Everyman’s Library, 1912), stanzas I to VI and, near the end of the poem, the note beginning “Then is added a secret” and the two stanzas after it (Internet Sacred Text Archive).

  15. The Meghadūta of Kālidāsa, with the Sanskrit commentary of Mallinātha, an English translation, notes, appendices and a map, ed. and trans. M. R. Kale, 8th ed. (Delhi: Motilal Banarsidass, 1974), Pūrvamegha 5 (Internet Archive scan; Kale’s glosses in parentheses fall outside the words quoted); Uttaramegha 51–52, Sanskrit as in the GRETIL e-text based on Kale’s edition (gretil.sub.uni-goettingen.de), with the English rendering of 52 mine.

  16. Kālidāsa, Meghadūta, Pūrvamegha 5, Sanskrit text with Mallinātha’s Sañjīvanī commentary (sanskritdocuments.org). Mallinātha: kāmārtā madanāturāś cetanāś cācetanāś ca teṣu viṣaye prakṛtikṛpaṇāḥ svabhāvadīnāḥ (the site prints cetanāś cetanāś ca, evidently for cetanāś cācetanāś ca). The English rendering of the gloss is mine.

  17. Kenneth J. Arrow, The Limits of Organization (New York: W. W. Norton, 1974), ch. 1, “Rationality: Individual and Social,” p. 23.

  18. George A. Akerlof, “The Market for ‘Lemons’: Quality Uncertainty and the Market Mechanism,” Quarterly Journal of Economics 84, no. 3 (1970): 488–500, section IV, “Counteracting Institutions,” p. 499 (“Numerous institutions arise to counteract the effects of quality uncertainty”); Chapter 23 develops his argument.

  19. Chennai Customs, Chennai-IV (Export) Commissionerate, minutes of the Permanent Trade Facilitation Committee meeting of 20 June 2025 (issued 24 June 2025), chaired by the Commissioner of Customs, Chennai-IV (Export), with the Federation of Indian Export Organisations and customs brokers’ and shipping agents’ associations attending (chennaicustoms.gov.in).

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 six layers (physical infrastructure, community, governance, capital, trust and being chosen) are ones AI cannot disintermediate, as the title says; that Efua’s assurance “was unfakeable precisely because it rested on relationships between conscious beings”; that belonging cannot be automated because an AI “cannot ... care whether you show up”; that governance needs “a self that can be held accountable” and “the community’s recognition”, and that “Participation requires a self. AI does not have one.”; that the customs problem was solved in “the space between what the rules said and what two conscious beings, negotiating in good faith, decided the rules should mean”; that Karthik’s pause “was an act of consciousness” that “no AI architecture currently in existence can replicate”; that trust “is a relationship between selves”, that “The language model cannot betray you because there is no self behind its outputs that has decided to be honest with you”, and that the midnight question “cannot be addressed to a server”; that relationships are “a layer. The deepest one.”; that an employee’s layer is being chosen and that “I want her on this” is “consciousness recognising consciousness”; that no machine can feel “whether the equipment is running true”, alongside a paragraph on sensors that already do; that such sensors “have not yet reached the textile mills of Tiruppur”; that “the cost of intelligence is collapsing toward zero”; and that Rajan’s discovery was “that his answer was himself”. It also named Uber and Amazon as removers of the middle, though both became a new one; gave Karthik an age (“twenty-six”) after 2024; made Rajan one of the “brokers and agents” a start-up would cut out, though he makes cloth; put his dyers in Karur, where he has a finishing house; said his AI adoption “nearly cost him his best workers” and that Muthu “had nearly walked away”, where three inspectors left and Muthu went silent; gave the Hamburg buyer “fourteen years” and said he “never once asked for a discount”; listed a customs officer who “called him last month to ask his opinion”; mixed a buyer’s label (Oeko-Tex) with customs papers; made the Kente elders “women in their seventies who had learned the patterns from their mothers”, though Kente weaving has traditionally been men’s; gave a pattern a different meaning “at a funeral in Cape Coast”, unsourced; set Efua Mensah in Jamestown, Accra; gave Priya Chandran, the test lead in Bengaluru, “lunches”, a “task force”, a “shower epiphany” and “three other candidates ... equally qualified”, none of them in this edition’s account of her; planted a visit (“That story belongs elsewhere”, twice) that a later chapter contradicts; and closed on a conversion, with jasmine at the window.

  • What I now hold: machines do not empty the middle; they move into it (the platform, AI summaries, checkout inside a chatbot). The title’s “cannot” holds only by meaning: a party is not a middle and cannot be disintermediated, but it can be replaced, so the “cannot” protects nothing. “Own” fails too: trust, belonging and the standing to decide are relations, and a person keeps or spends his part in them. Machines in the middle are mediators in Latour’s sense: they change what they carry. The physical middle splits into carrying information, which machines take first, and carrying risk, which stays with whoever agreed to bear it. A machine can make a payment; the loss stays with its makers and users. What a cloth means splits into what is written (a machine can retrieve it), standing (who may say what a use means now) and the hands (who wove it); a record of origin is testimony and must be open to dispute. Legitimacy is a chain of reasons that can be checked and a way to correct a decision; access is not legitimacy, and the face-to-face fix that served Rajan was a privilege of the known; faceless assessment (2020), the impersonal ideal Weber described, may be fairer to strangers, provided someone gives reasons and a wrong decision can be challenged. A decision to invest is a promise made with other people’s money; a hunch is a reason to look, sorted by Kahneman and Klein’s test. Assurance can be bought (one of Akerlof’s counteracting institutions); trust cannot (Arrow). Trust is the ground every middle, human or machine, stands on: spent by any middle that misleads in someone’s name, kept by those who explain and put right, owned by nobody. Kālidāsa’s verse on the cloud draws the line at life, as I read it, and Latour at nothing; this book draws it at what a message is: it belongs to someone, and someone must answer if it arrives wrong. An employee owns no layer; being chosen is someone else’s act and often begins with a machine; an employer owes those it screens a way to be seen and to dispute. A new middle may take the toll and leave the risk at the ends. In the partner’s view, the Jakarta brokers survived the app because it took only the matching; nobody could yet know whether she was right. A party may choose whom it deals with, where a gate must treat strangers alike; both owe reasons and must hear a challenge. A machine middle that became the only road to buyers would be a gate too, and would owe reasons for its rankings and a way to challenge them. A right to argue is worth the time that comes with it. The poem’s own token of recognition, as I read it, shows what a message needs from its carrier, a way back to the one who sent it; where the sender never chose the carrier, the carrier’s makers owe an answer of their own.

  • What changed my mind: the definition of disintermediation itself, which makes the coda’s “cannot” true and empty; the print’s own examples, which became new middles; three outside readings of an earlier version of this chapter, in October 2026, which found that a new middle can take a toll and leave the risk at the ends, that the Jakarta story needed a reason why the brokers survived, that the panel and the faceless reform pulled against each other and that the panel’s own standing went unexamined, that Latour describes and does not prescribe, and that the employee section left out time and voice; the token of recognition near the end of Kālidāsa’s poem; Pew’s study of AI summaries and OpenAI’s checkout announcement, both of 2025; Latour’s intermediaries and mediators, with his own example of the computer; Kālidāsa’s cloud, read with Kale, Ryder and Mallinātha; Arrow on trust that cannot be bought and Akerlof on institutions that sell assurance; India’s faceless assessment and Weber on bureaucracy; the study of women and the loom in Bonwire (2023) and Ghana’s Copyright Act of 2005; the audit’s findings that five of the six sections repeated an earlier chapter and that the coda repeated the residue another chapter had ruled out; the canon for Rajan, Karthik, Weiss and Muthu.

  • What survives: “Rajan’s nephew called it disintermediation, and Rajan told him he was wrong.”; “Everything ... Including me.” and “That is what I am telling you, Uncle.”; “You cannot download a bale of cotton.”; “He believed it. He was half right.”; the six subjects and their numbers; “a community is not an intermediary. It is the thing itself.”; the customs officer, his caution “rooted in a fraud case”, and Rajan bringing the earlier papers “not as a weapon”; “A decision to invest is not a prediction. It is a promise.”; the concession that human judgement does not universally beat algorithmic analysis; the calculator; “Trust is not a moat around the castle. Trust is the ground the castle stands on.”, now with nobody owning the ground; the 22.5 title, “The Layer Beneath Everything”, with “own the layer” narrowed to “keep your part of it”; “You cannot disintermediate someone who is not intermediate.”, now carrying the argument by meaning.

  • Section titles in the print edition: 22.1 was “Physical Infrastructure: The Atom Advantage”, 22.2 was “Community and Identity: Belonging Cannot Be Automated”, 22.4 was “Capital Allocation: Money Follows Judgement”, and 22.6 was “Distribution for Employees: Being Chosen by Humans”. The other two keep their print titles.

Recent Posts

See All

Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
bottom of page