# AI-Proof: What Machines Can Never Be — Volume III: Transformation, Part VII: The Conscious Blueprint, Chapter 36: The Conscious Organisation

By Nithin Palal · 21 November 2026 · 47 min read
Section: Artificial Intelligence · Series: AI-Proof: What Machines Can Never Be
Source: https://nithinpalal.com/ai/ai-proof/chapter-36-the-conscious-organisation

> A plant near Lyon got four entries in four months on a form for concerns about its machines. A night technician had raised a doubt in March. An organisation is not conscious: people speak where speaking is answered, and named people answer.
*Part VII looks at what is worth doing on purpose. After a year of practice, working alone and the start-up, this last chapter of Part VII asks how a large organisation comes to know what its machines are doing, and who answers in its name.*

← Previous: [Chapter 35: The Conscious Startup](https://nithinpalal.com/ai/ai-proof/chapter-35-the-conscious-startup) · Series hub · Next: [Part VIII: The 2035 Human (overview)](https://nithinpalal.com/ai/ai-proof/part-8-the-2035-human) →

**A plant near Lyon got four entries in four months on a form for concerns about its machines. A night technician had raised a doubt in March. An organisation is not conscious: people speak where speaking is answered, and named people answer.**

**In Brief: For readers who work inside large organisations that now run partly on machines. Chapter 36 asks how a large organisation comes to know what its machines are doing, and who answers in its name. A vice-president from Hyderabad takes her site’s practices to a plant near Lyon, where they fail; a night technician’s doubt about a release system is tested by a model; the chief executive in Basel learns what never reached him; a customer service team is weighed against a voice agent. The monks’ rule that reads silence as purity, and Hobbes on how many people become one person, help show how an organisation hears, who speaks for it, and who pays for both.**

**An organisation is not conscious and need not be; it knows only what its people have learned is worth telling it, and it answers only through people who act in its name and can be found, so the telling and the answering must both be arranged, and paid for, before a machine goes wrong.**

## Opening: The Meeting Room near Lyon

The meeting room at NovaChem’s plant near Lyon smelled of coffee that had stood too long and of carpet cleaned too recently. Meera Reddy noticed it in the first minute and then stopped noticing. She was watching the room.

It was February 2026. Meera is vice-president of operations at the Hyderabad site of NovaChem, a drug company with its head office in Basel. She had been with the company for fifteen years, the last seven as vice-president, and about four hundred people work in her operations. Basel had asked her to write down what her site had built around its machines since 2023, so that other sites could try it. She had called the document “Conscious Operations”. Lyon, the plant that makes NovaChem’s medicines for Europe, would try it first, from March.

The site director, Isabelle Moreau, asked how many people the new work would need. A manufacturing lead asked what records the reviews would have to keep. The head of quality asked how a new form for concerns would fit the system the plant already used for deviations. The questions were specific. They were professional. They were, every one of them, about machinery.

Meera recognised the room the way you recognise your own handwriting in a document you do not remember writing. In 2023 a consulting firm she had hired from Mumbai had laid four new tools over her own site’s work, and a chemist there had told her in a corridor that the result was “not ours”. Now Meera had flown from Hyderabad with a document, to do to Lyon what she had once done to her own site.

An organisation is made of people, and each of them notices things. Its machines notice things too, and some notice more than any person could. The organisation itself notices nothing. It has no eyes of its own and no mind to keep what the eyes see. This chapter asks: **An organisation is not conscious. When its machines go wrong, how does it come to know, and who answers in its name?**

The print edition took a position on this that this edition must change. “No. Organisations are not conscious,” it said, and that much was right. Then it went on: when an organisation’s people are awake, “the organisation behaves as if it can see. This is not a metaphor.” It is a metaphor, and it hides the work. An organisation does not see. It is told, or it is not told. People tell where telling has changed something before, and they keep quiet where it has not. And an organisation answers only through people who act in its name and can be found, and only for what those people could know. Making both possible costs hours and money, and someone has to decide to pay them.

So the chapter’s title needs an argument, once. It keeps a word this book no longer lets carry any weight. A few philosophers have argued that a large enough group might, in some sense, be conscious.¹ This chapter takes no side, because nothing in it depends on how that is settled. What an organisation can be is something that hears and something that answers. Both are done by people, and both can be designed well or badly. It does not need a mind to be held to account. It has power, it makes decisions, and it spreads their costs onto people, and that is enough.

## 36.1 Diagnosing What an Organisation Hears

The pilot ran from March to June 2026. Lyon installed three things from Meera’s document. Each machine system got a named person at every point where its work passed to a person. A form let anyone raise a concern about a system. And each team was asked, once a week, the question Meera had used in Hyderabad: “What did you notice this week that nobody asked you to notice?”

By the end of June the form held four entries. The reviews happened on schedule, and the panels treated them as audits, ticking a checklist against each system’s outputs. The weekly question was tried twice and dropped. People found it strange. Not because they had nothing to say. Because nothing in their working lives had ever rewarded the act of saying it. That was true at least of their working lives at that plant.

The pilot failed quietly and politely, with everyone doing what was asked of them. Many hands made it. Basel approved a template. Meera wrote it as one. Isabelle ran it because Basel had asked her to. Nobody in Lyon had been asked what the plant needed. The document had travelled. What had made people in Hyderabad take its questions seriously had not: the replies people there had learned to expect, and the hours a budget had paid for. A practice moves with the conditions that made it work, or it arrives as a form.

At the end of June Meera telephoned Isabelle, and asked her one question: “What are your people already noticing that no one is asking them about?” Isabelle went and asked. What she heard came from the night shift.

Every batch of medicine the plant makes leaves a record: the weights, temperatures and times of each step. Since 2024 a machine system had read each record before release and flagged anything unusual. Then a person signed. Under the European Union’s medicines law, a qualified person must certify that each batch “has been manufactured and checked in compliance with the laws in force” where it was made and as its licence requires, and record it in a register.² The system checks; the qualified person signs.

A quality technician who had worked nights at the plant for years had noticed something about one product. When it was made at night, the system’s reading of its records came back clean more often than he believed it should. He read some of those records himself, out of habit, and saw small things he would have raised by day. In March he told his shift supervisor. The supervisor pointed out that the system had been validated, which was true. Nothing was written down. When the form arrived in March, he did not use it.

His supervisor was not a coward, and neither was he. A deviation raised against a validated system opens an investigation. It can stop production, it brings auditors’ questions, and it falls on the night shift’s own records. The supervisor had a release plan to meet. The technician could not prove what he had seen, and the system had been tested by people paid to test it. Each had an ordinary reason to wait. In practice, the arrangement gave neither of them a reason not to.

The rules on paper asked for more than waiting. EU guidance on computerised systems says: “All incidents, not only system failures and data errors, should be reported and assessed.”² The technician had reported his doubt aloud, to the person above him. The assessment was one sentence about validation, and nobody wrote it down. Whether a doubt counts as an incident is itself a judgement, and nobody recorded that one either.

In July the quality team asked a model to compare two years of the system’s flags, by product and by shift. It took an afternoon, and this call goes first to the machine. Someone in the quality team still had to choose what to compare, and to check the counts against the plant’s own records. Most of the gap between night and day was the product mix: the plant ran its simplest products at night, and they had fewer deviations to flag. On those, the system had been right, and so half of what worried the technician was wrong. The other half was real. For the one product he had named, small deviations were flagged less often at night than the same deviations by day. The data the system had been tested on held few night runs of that product. No batch had left the plant outside its limits. The check had been thinner, not wrong in its results. In August the plant’s qualified person decided that, until the system was tested again, a person would read in full the records of that product’s night batches before she signed. It cost night-shift hours, and that product’s batches were released a day later. Part of the model’s finding could be checked without it, since the plant’s schedules showed which products ran at night. The rest was what the retest would test. A difference in flags was a reason to look, not yet proof of a defect. Nobody at the plant had held the whole of it. The technician knew the product, the quality team the validation, and the qualified person her duty to sign. The system’s makers knew what it had been tested on, and what they would show the plant depended on NovaChem’s contract with them.

A second reading is harder than it sounds. Lyon, like other plants with such systems, reviewed its records by exception: the screens led a reader to what the system had flagged, not through the whole record.² Reading a record from end to end, at night, was a skill the system had let lapse, except in the few who kept it up out of habit. A fallback that nobody has practised, Chapter 30 found, is not a fallback. A plant that stops trusting a machine needs more than the right to stop it. It needs people still able to do the work by hand, and someone named to decide what evidence will let the machine back.

What did this say about Lyon? Three things can be asked of any organisation, and they say more than its forms do. Do people argue about ideas, or only about whose decision something was? When something goes wrong at the bottom, how far is it softened on its way up, and in whose words does it arrive? And what happened to the last person who said, out loud or in writing, that something was wrong? At Lyon, nothing had happened to the last person who said it, and everyone knew. A fourth thing matters as much: whether a concern can go round the person whose own decision it questions. At Lyon the new form offered such a route, but few yet trusted it. The technician’s doubt went to a supervisor with a release plan to meet. There is a limit to asking. Some of what a manager knows is held in trust and does not belong on a dashboard. Finding out what an organisation hears must never become watching what its people say.

An old rule makes the point sharply. Every fortnight, monks of the Buddhist order meet to recite their rule, the *Pātimokkha*. As the recitation opens, the one reciting tells the assembly how silence will be read. In T. W. Rhys Davids and Hermann Oldenberg’s translation of 1881: “Whosoever have incurred a fault, let him declare it! If no fault have been incurred it is meet to keep silence! Now, venerable Sirs, it is by your silence, that I shall know whether you are pure.”³ The question is put three times. A monk who stays silent while he remembers a fault “is guilty of uttering a conscious lie”. And then: “For a fault, when declared, shall be light to him.” Bhikkhu Brahmali’s modern translation gives the same lines as “If you are silent, I will regard you as pure”, “lying in full awareness” and “When it’s revealed, he will be at ease.”⁴ The Pāli word for the lie is *sampajānamusāvāda*: a lie told knowingly. The 1881 translators’ “conscious” means exactly that, knowing.

The rule is about a monk’s own faults, confessed among equals, and I borrow only how it treats silence. It reads silence as purity, and within its own assembly it has earned the right to. The question is asked three times, so nobody can miss it. Each is held to his own silence. And speaking brings something: the fault declared is made light, *phāsu*, an ease. Lyon read its silence as purity too. Four entries meant nothing was wrong. But the question was asked once a week, and dropped. The technician was asked to report a machine’s fault, not his own, so nothing in speaking could make anything light for him. Saying it would have made more work, for himself and for others. Lyon had not earned the right to read its silence as anything.

The difference runs deeper than the asking. The monk declares a fault that is his own. The fault the technician saw, if it was one, belonged to a system that others had built, tested, bought and learned to trust. So the duty to speak cannot rest on him alone. Nor should an organisation borrow the rule’s demand that every fault be declared. In a company, that demand turns doubts into the company’s property and colleagues into informants. A speaker can have good reasons not to tell a particular person: a patient’s privacy, a colleague he does not want to expose unfairly, fear of what follows. A good channel lets him keep his name private when he asks, and does not run only through the person he is worried about.

## 36.2 The Leadership Problem: You Cannot Lead What You Cannot See

Friedrich asked Meera to come to Basel. It was the middle of September 2026, and he is NovaChem’s global chief executive. He is a chemist by training, and his office is ordered like a laboratory. He had read her report on the Lyon pilot, and he wanted to talk about one sentence near its end: “The practices were installed. The consciousness was not.”

“What does that mean?” he asked. “You say yourself that consciousness is not something you install. So what did Lyon lack?”

Meera had written the sentence in July, and she could not defend the word “consciousness” in it. The people at Lyon had noticed plenty. What they lacked was any evidence that saying so would change something. “A memo from the CEO is not permission,” she said. “A scheduled review is not permission.” Permission, as she meant it, was what happens when someone raises a concern and the reply makes the next person more likely to speak, not less. A mandate could require a reply. It could not make people believe in one.

Friedrich defended his company, and much of what he said was true. NovaChem’s lawyers had settled in the spring which of its systems the European Union’s law on AI covered. Its staff surveys were among the best in its industry. “I do not think this company is asleep,” he said.

He asked for an example where NovaChem had failed to see something that mattered. She gave him the case of the model that sorted reports of side effects from NovaChem’s clinical trials around the world. In July 2025 it had been found giving less weight to reports from some trial sites, in Nigeria, Kenya, India and the Philippines, whose records were less complete. The model had been right about the records and wrong about the patients.

“Who caught it?” Friedrich asked.

The truth was not what he expected. A quarterly audit had caught it, run by software: a machine checking a machine. Meera had asked a question afterwards, whether the company was treating all its patients equally. The decision to change the threshold had been taken in August 2025 by NovaChem’s own drug-safety physicians, in Basel. Nobody had told Friedrich. In the summary that reached him, it had become “a methodology update in drug safety”.

Nothing in that phrase was false. A person in the drug-safety office had written it, honestly and briefly, and a change of threshold is a change of method. It had simply not been written for someone who needed to know what it meant. Friedrich asked how many other decisions had reached him as a phrase. Meera knew only her own site’s answer, from a reading she had asked a model to make before she came. For the rest of the company, nobody in the building knew.

Thomas Hobbes, the English philosopher, asked in 1651 how many people can become one. His reply was: by representation. “A multitude of men are made ‘one’ person, when they are by one man or one person represented; so that it be done with the consent of every one of that multitude in particular. For it is the ‘unity’ of the representer, not the ‘unity’ of the represented, that maketh the person ‘one.’”⁵ The one who acts for them is the actor, and those who own what he does are its authors: “he that owneth his words and actions, is the ‘author’: in which case the actor acteth by authority.” Hobbes wrote about the commonwealth and its sovereign, not about a company, and I borrow only his account of how many become one. On that account, Friedrich is the closest thing NovaChem has to a single person. When he speaks, the company speaks. He knows only what reaches him.

Hobbes allowed the representer to be an assembly as well as one man. Then, he wrote, “the voice of the greater number must be considered as the voice of them all.” In Swiss law, NovaChem’s board is that assembly. It represents the company, and it may hand that task to managers, as it has to Friedrich. The law also makes the company a person of its own, with rights and duties, able to act once its governing bodies are appointed.⁵ So NovaChem can be sued and made to pay. But it acts only through those bodies, and they are people. Someone decides to settle, someone signs, someone writes the account.

Hobbes’s picture has one representer. A company has many, and each answers for something different. Lyon’s qualified person certifies batches under a duty the law places on her, not one Friedrich lends her.² The board answers for supervision, Friedrich for running the business, a site director for her plant. So NovaChem is three things at once: a person in law, a set of routes by which decisions get made, and an arrangement that spreads gains and burdens among people. None of the three needs a mind. Being the first does not make it fair in the third.

Then Meera said what she had come to say. “I cannot do what you are asking. Not the way I did it in Hyderabad.” She had been asked to make other sites see as hers did. But even in Hyderabad she had seen little with her own eyes. A chemist had told her the first tools were not theirs. A customer service director had asked who would answer when her team’s design went wrong. Others had told her, to her face, what her own rules had cost them. Each time, what Meera knew had come from someone who decided to tell her. “Every site needs someone who is there,” she said. “I can give them the questions. I cannot give them the eyes.”

There was more that she owed him. Since December 2025, when a governance team of three began reviewing the site’s machine systems, every review of a new or changed system in her operations had ended with her signature beside the system owner’s. By the spring she had stopped reading some of them before she signed. In March she had signed one without reading it, and in August a member of the governance team had asked her about it, and she could not remember it. Friedrich asked what she would change. She did not yet know.

Her confession belonged to the same story as the night shift’s silence, told from the other end. Below, a doubt was said aloud and never written down, because saying it seemed to change nothing. Above, a signature went on unread because there were more papers than hours. About thirty reviews had reached her desk since December, each a document of many pages. A signature can carry only as much as the reading behind it. Beyond that, it records that a procedure was followed, not that anyone judged. From below, the two look the same. What she changed, in October, is told in Chapter 37.

The two old texts pull against each other here. In Hobbes, the many are one through the person who represents them, and the represented own what he does. Responsibility runs from the top. In the monks’ rule, the assembly is pure only if each member speaks of his own fault, and nobody speaks for him. There it runs from each person. Both make a group one by a procedure, not by a shared mind, and that is this chapter’s reply to the word in its title. But each has only half of what an organisation needs. Hobbes’s representer can stand behind only what he is told. The monks’ assembly has nobody to answer for the whole. And NovaChem’s staff consented to nothing in Hobbes’s sense: they were hired, not asked. So I refuse the direction of ownership Hobbes gives. The people on the Lyon night shift are not the authors of Basel’s choices. An organisation needs both halves: someone who answers in its name, and people whose saying reaches him.

### The other side

Some philosophers would say this chapter makes the organisation too thin. Christian List and Philip Pettit argue that a company can be an agent in its own right. It forms judgements, makes choices and can be held responsible as a company. Their sharpest example comes from two legal scholars. A court of three judges must decide whether a defendant is liable. Each judge finds him not liable, each for a different reason. Yet if the court votes on each question separately, a majority says yes to every one, and the court finds him liable. That verdict is the court’s, and no judge holds it. Pettit adds that a company can be fully fit to be held responsible when its members are not, because each of them may have an excuse the company lacks.

On Lyon, they partly win. Everyone did what was asked, and nobody chose the pilot’s failure. A company can produce what none of its people decided. That strengthens this chapter’s demand rather than defeating it. Even such a company learns only what its people and procedures carry to it, and it explains, pays and changes through people. List himself allows that members can still be responsible for the parts they played. If a company can do what nobody chose, someone must be named in advance to watch for it.

## 36.3 Building a Culture That Replies

“What do you need from me?” Friedrich asked. Meera asked for two things, and neither was a programme.

The first was a reply. The governance team in Hyderabad can require the owner of any system to reply in writing to a finding and to test the system again. It can also suspend the site’s use of a system while it waits. Meera asked for something smaller and wider. Anyone, at any site, who raised a concern about a machine system should get a reply in writing, within a month, from a named person. It would say what had been done and why, or why nothing had been done. A month is the longest anyone should wait for an ordinary concern. A serious one gets its reply sooner, and if a full reply must wait, the person who raised it is told why. A doubt about a batch about to leave still goes to quality at once, by the routes a plant already has. The reply does not replace the plant’s system for deviations. It covers what that system misses: a doubt not yet firm enough to be a deviation, like the technician’s in March. A reply is not agreement, and it puts nothing right by itself. It is the only evidence a person ever gets that saying something was worth the risk. It does not protect her, though. Whether speaking costs her anything depends on her next review and her next shift, which no reply reaches. The monks’ rule promised that a fault declared would be light. An organisation cannot promise that. It can promise that speaking will not vanish.

Friedrich saw the cost at once. A rule like that, he said, would bring a flood. People would write concerns to protect themselves, so that when something went wrong they could say they had warned. The replies would take hours from the people best able to fix things, and soon they would be written from a template, perhaps by a machine. He was partly right. A reply can become a form, and a concern can be raised only to cover its writer. Meera’s reply was narrower than a defence. A flood would itself be information, because it would show where people did not trust the ordinary routes. And a reply written from a template would fail the only test that mattered, which was whether the next person spoke. A good reply also says what was checked and what was not, so that nobody reads it as more than it is. Neither of them could say in advance which way it would go.

A reply is the start of answering, not the whole of it. After a concern, someone must examine the evidence, and someone with the power to decide must decide. The speaker should get reasons she can read. Where something went wrong, it should be put right, and whatever let it happen should change. A reply can be prompt and empty. So a good one also says where to go if the speaker thinks it wrong. That should be someone outside her own line, who can ask for the evidence and keep the matter open. A right to speak with no right to challenge the reply looks like hearing and leaves every decision where it was.

Some of this the law already asks. Since 2022, French law has given effect to an EU directive of 2019 that protects whistleblowers. The directive requires employers of fifty or more people to run an internal channel for reports of breaches of EU law in listed areas. Its areas include public health, and the EU’s medicines law is on its list. In France, the channel must acknowledge a report within seven working days and give feedback within three months. The law forbids retaliation, keeps the reporter’s name confidential, and lets a reporter go straight to an outside authority. Indian company law asks listed companies, and some others, for a “vigil mechanism” with safeguards against victimisation. Such channels are built mainly for suspected wrongdoing and, in France, threats to the public interest. The technician’s doubt might not have counted as either, and most of what a night shift notices is neither. Meera’s reply rule reaches the doubts that the law’s channel was not built for.⁶

Forbidding retaliation is the easy part. Retaliation is seldom a dismissal. It is a worse shift, a cooler review, a promotion that does not come, a reputation for being difficult. A policy can forbid all of it while the managers who make speaking costly are promoted. So someone outside the speaker’s line should look, a year on, at what happened to the people who spoke. Who pays for acting on a concern matters too. If the hours come from the budget of the site where it was raised, the speaker’s own colleagues pay for his speaking, in their shifts and their training. Resentment in a break room can silence the next speaker as surely as a supervisor’s frown. Paying from a central reserve, or from the system’s maker where the contract allows, keeps that price away from the people nearest the speaker.

The second request was for the board, and it belongs to the next section.

Culture is not the poster in the lobby. Culture is what happens when the poster’s author has gone home. A new technician does not learn from a values statement whether to speak. She learns it from what happened to the last person who did. A speak-up policy builds a channel: a route, a named reader and a deadline. Culture is whether the last person who used it is still trusted, still promoted and still on the shift she wanted. This is why a silence like the night technician’s is not always a failure of the person, or of the system’s specification. It can be learned from an arrangement: a validated system, a supervisor with no route upward, a form that went nowhere. Silence learned that way is changed by someone with authority, not by a better form. The political philosopher Dennis Thompson has argued that organisations should name, in advance, the people who must watch how responsibility is arranged and change it when it fails.⁶ A reply with a name on it is one small piece of that. At Lyon one person noticed, another heard, a third tested, a fourth signed and a fifth paid. That nobody did all of it does not mean the outcome has no owner.

Lyon found its own form before Basel decided anything. Isabelle added one item to the end of each team meeting: “Is there anything the systems are doing that we should be paying closer attention to?” It asked less of people than Meera’s question had, and it fitted a plant that already ran its work by meetings. It cost each team ten minutes a week. Whether it will be used for long, nobody yet knows. A question asked every week can become a form too.

The hardest part has no form at all. The technician who notices a thin check has prevented something that never happens, so there is nothing to see and nobody to thank. A check that confirms a machine was right is the commonest kind of check and the least paid. Half of the technician’s worry had been wrong. If an organisation rewards only the catches that turn out right, people learn to speak only when they are sure, which is too late. What has to be protected and recognised is the honest saying, judged by whether it came in time to someone who could act, not by whether it proved right. A person must be able to say: I noticed something, I cannot explain it yet, and I may be wrong. Counting reports would only teach people to write more of them. The same holds for what people know. An appraisal that praises the irreplaceable teaches people to keep their knowledge to themselves, as Chapter 33 found, and leaves the company carrying a risk of its own making.

Hearing is also not remembering. A concern answered and then forgotten has been handled, not learned from. People leave, systems are updated and makers change, and a failure once understood becomes a line in a database. What has to be kept is more than the record: what was thought at the time, what evidence changed it, what was done, and whether it worked. The name on a reply should be a role’s as well as a person’s, so that open matters pass to whoever holds the role when she goes. Lyon’s finding is a test of this: whether a doubt raised at one plant changes how every site that uses the same system checks it.

## 36.4 The Board’s Responsibility: Governance as Attention

The print edition said that for a drug company using high-risk AI in the European Union in 2026, “board-level AI governance is not a visionary act. It is a legal requirement.” That is wrong. The EU’s AI Act puts its duties on the companies that provide and use systems, not on their boards.⁷ Its duties for the high-risk uses listed in its third annex apply only from 2 December 2027, after the law was amended in 2026.⁸ None of the systems in Meera’s operations falls in that list, as NovaChem uses them. The hiring system its plant near Lyon uses will, from December 2027, and even then the duties fall on NovaChem as its user, not on a board, let alone one that sits in Basel, outside the Union. A system outside the Act’s high-risk list is not therefore low in risk, or free of other law. The release system at Lyon falls under the medicines law and its rules for computerised systems.

The law that binds that board is Swiss, and it says something better. The board may hand over the management of the business “to individual members or third parties”, as its rules allow. Some duties it may not hand over. Among the duties the Code of Obligations calls “non-transferable and inalienable” is “overall supervision of the persons entrusted with managing the company, in particular with regard to compliance with the law”.⁹ The law draws a line much like the one this book draws, between handing work to someone and handing over the right to answer for it. The board may delegate the doing. It may not delegate the watching. That does not mean inspecting every system, which no board could do. It means keeping the arrangements by which it would learn that something was wrong.

What does watching mean for a board that meets a few times a year? A case from another country shows what it looks like when it is missing. In September 2021 a court in Delaware let a claim by shareholders of Boeing against its directors go forward. It concerned the 737 MAX, whose automated flight-control software had played a part in two crashes that killed 346 people. The software was automation, not AI. The court’s account of the board, drawn from the shareholders’ complaint and the company’s own documents, was this. “None of Boeing’s Board committees were specifically tasked with overseeing airplane safety, and every committee charter was silent as to airplane safety.” “Airplane safety was not a regular set agenda item or topic at Board meetings.” “Management’s periodic reports to the Board did not include safety information.”¹⁰ At that stage the court took the shareholders’ account as true; it found no director liable. The case was settled. The settlement, approved in February 2022, included $237.5 million, paid to Boeing by its directors’ insurers, and, among other changes, an ombudsperson to whom the employees who certify Boeing’s aircraft on the regulator’s behalf could take their concerns.¹¹ The board had not lacked information in general. It had lacked a channel by which the hard kind could reach it.

Delaware’s courts have their own version of the duty to watch, which binds the boards of companies formed in that state. They call it the *Caremark* duty, after a case of 1996. As later restated, directors can fail at oversight in two ways. They can fail to put in place any system for hearing about a central risk. Or, having one, they can consciously fail to watch it, for instance by ignoring the warnings it brings, which lawyers call red flags. “Consciously” here means what it meant in the monks’ rule: knowingly. In 2019 the state’s supreme court let shareholders of Blue Bell, an ice-cream maker, go forward against its directors. The board had had no board-level system for hearing about food safety before a listeria outbreak killed three people. The Boeing claim went forward on both counts: no system for airplane safety, and a board that, on the shareholders’ account, ignored the first crash.¹⁰ Naming a board committee to oversee AI, as many large companies now do, answers the first kind on paper. The second is harder to see. NovaChem’s summary was honest, and nobody ignored a warning. The warning simply arrived as “a methodology update”. Delaware’s rule reaches a board with no channel, and a board that looks away. As I read the cases, it does not reach a board that never sees the warning because it came in accurate, bland words. That gap is where attention has to work.

So governance is a form of attention, in a plain sense. A board attends with its agenda, its committees and the reports it asks for. A board can perform governance without practising it. It receives the reports, confirms that the rules were followed and keeps the papers that auditors will one day ask for. Practising it means asking for the reports that hurt. A board that signs a summary it did not question has its name on a decision without having made it.

This call, too, goes to the machine. Before Basel, Meera had given a model her site’s quarterly reports for 2025 and the summaries that went from them, and from the offices they fed, to Basel. In an afternoon it marked the places where a delay or a decision had been softened on the way up. One of them was the drug-safety threshold. Reading was not what the board had lacked. As of 2026, a machine can read every report an organisation writes. What it cannot be given is the board’s duty to ask.

Better hearing cures only one failure, which is not knowing. A board can also know and not act, because acting would cost a target, a supplier or a public claim. That is not a failure of hearing, and no channel cures it. It is a decision, and it should be recorded as one, with a name and reasons. An organisation can also learn not to hear. Its reports may be accurate and its dashboards full, while every hard fact arrives in the words that disturb least.

Hobbes helps to see a harder limit. A board represents the company’s owners, who chose it, and it is bound to them. The patients who take NovaChem’s medicines, and the people in its trials, chose nobody on that board. They are not its authors in Hobbes’s sense. Yet its decisions fall on them first. A board that decides things for people it does not represent owes them at least a hearing. At a drug company they already speak, in one form: the reports of side effects that doctors in its trials send in. The model that weighted some of those reports less was, in effect, the organisation hearing some patients less well than others. So the board’s attention must reach beyond its own staff. It can ask, each year, whose reports were read least, and why. It can also ask whose harms are least likely to become reports at all, such as those of patients far from any doctor who would write one.

Meera’s second request was this. The board should be told, by name, every time a machine system somewhere in NovaChem was stopped, and every time someone asked for a stop and was refused. The board does not hold the power to stop a system. At her site that power is the governance team’s, and its terms are written down. But a power to stop that nobody above ever hears about is easy to leave unused. Someone at the highest level should know when the word was said. In a company of NovaChem’s size, stops will be many and mostly dull. A board needs them as a pattern, by site and by system, with the refused stops on top, not as a stream. Friedrich said he would put the proposal to the board at its meeting in December. What the board will decide is not known.

A board’s attention costs the board little. Acting on what it hears is paid for lower down: in night hours, in a release delayed, in money taken from somewhere else. A board that asks for the hard reports should also ask who will pay for acting on them, and see that it is not always the people who spoke.

## 36.5 When to Hire Humans and When to Deploy Machines

Friedrich had a question of his own, and it was about money.

Basel had a plan for customer service at every NovaChem site. From 2027 a voice agent, a machine that speaks with callers, would take the routine calls. Is a medicine in stock? When will a delivery arrive? What does an invoice mean? People would keep the rest. In Hyderabad the calls come to a team led by Shalini, NovaChem’s customer service director, who has been with the company for nine years. Under the plan her team would shrink to about a third of its size. Basel had already written the saving into its budget for 2027.

This call goes to the machine first. In a trial at another site, callers rated the voice agent about as well as they rated the agents. It picked up at once, at any hour, and nobody waited in a queue. For a pharmacist calling at night about a delivery, it is better than a person who is not there. An average rating, though, says nothing about the callers who hung up, or who asked for a person and could not reach one.

Meera had brought a map. For years she had drawn maps of her site’s work on the whiteboard in her office. She used four colours: blue for data, green for decisions, red for failures, black for the handoffs between people. In early September she had drawn Shalini’s calls this way. At the foot of the map she had written her rule: “Mark every point where a person must be present because the cost of their absence is larger than any efficiency gain.”

“Whose cost?” Friedrich asked.

It was the right question, and the rule needs it. The cost of a person’s absence and the gain in efficiency are not paid by the same people, and they are not counted in the same money. The gain would show in Basel’s budget. The cost would fall on callers, and on agents. It would be paid in different coin: in waiting and mistakes for callers, in lost work and harder days for agents, and in skills the company would stop forming. So the rule is not a sum, for a machine or for a person. Each mark on the map is a decision about whose costs count, and someone must answer for it. “Present”, too, means more than a person in a chair. It means a position held: someone a caller can reach, who can explain, and who can put something right. A caller should be able to ask for that person without giving a reason, and reach someone who already has what she said.

Then Meera made her case, and she called it a forecast, because she could not prove it. Shalini’s team had already lived a small version of the plan. Their routing sent the angriest callers straight to the most experienced agents. Shalini could change the routing rule, but not the calls-per-hour target set above her, and the target had made the angriest callers the whole of her best agents’ day. The plan would make that everyone’s day. The easy calls were where new agents learned the work, and where everyone rested between the hard ones. Not every easy call teaches anything, and rest could be bought more cheaply, with a lower target and set breaks. What is harder to buy another way is the learning, and the ear. Managing your own feelings so that a frightened caller feels calmer is work. The sociologist Arlie Hochschild called it emotional labour, and it is tiring.¹² How tiring depends partly on how it is done. A review of 95 studies, published in 2011, found that faking or hiding feelings at work, which researchers call surface acting, went with poorer well-being. Trying to feel what the job asked, called deep acting, was only weakly linked to it.¹² The studies measured links and could not show causes. A team that does nothing else will lose its best people first, Meera said. The plan would be left with a machine for the easy calls and beginners for the hard ones. Some calls that sound routine are not. Shalini’s team already sends any caller who names a patient, a medicine and a date to a trained agent at once. That rule catches callers who name all three. It cannot catch a caller who asks only when a delivery will come, calmly, though a patient is waiting behind the question. An agent who has taken a thousand ordinary delivery calls is likelier to hear when one is not ordinary, though Meera could not prove that either. And Shalini had once asked Meera a question that still had no full answer: “If my team designs this, who answers when it goes wrong: my team, or you?”

Two groups were not in the room. The agents whose work the plan would take would not be asked. Nor would the callers. Nobody had asked a single caller which calls they wanted a person to take.

So the question of when to hire a human is not which tasks are left over once the machine has taken its share. The plan settled that well: the routine calls could go. But a role is not the sum of its commonest tasks. Shrinking the team to a third assumed the routine calls were two-thirds of the work. The agents also taught new colleagues, caught mistakes made elsewhere and knew which customers had been let down before. It is a question about positions and about learning. Someone must hold the place a frightened caller can reach, with the time and the authority to help. A voice agent can hold the conversation. When it fails, the caller’s problem still belongs to someone in the company. And someone must have been formed, on easier calls, to be able to hold it. A company that keeps people only for the hardest work spends the skill it once paid to build, and stops building it. Shalini named the other half when Meera told her what the plan said. Whatever Basel decided, some of her agents would lose their work in 2027 or 2028. They were owed an account of what was coming, early and in writing, not a year of silence and then a letter. Shalini asked for that account, and Meera put the request to Basel. An account was only the least they were owed. Notice, money and help into other work may be owed as well, under the law where they work and by those who wrote the saving into the budget.

The decision was not Meera’s. She said what she expected and what she could not prove. The next week Friedrich wrote to her. For one year Hyderabad’s agents would keep the routine calls during the day; the voice agent would take nights and overflow. The difference against Basel’s plan would come from her own operations’ budget, not from Basel. At the end of the year, a measure Basel would choose would decide what happened next. Which measure matters as much as the year. Cost per call alone would swap one narrow target, calls per hour, for another. A fair measure would also count callers who gave up, calls that came back, agents who left and new agents formed.

Meera named in writing what that money would not buy. In 2025 she had asked Basel for two more people to review the safety reports from the trial sites, and Basel had funded one. The rest of that work had come from her teams’ hours. She had meant to fund the second reviewer herself in 2027. Now she could not, and the case-processing team would keep carrying those hours for another year. She may have accepted a quieter harm to ease a louder one. Shalini’s agents were in front of her, with a plan and a date. The reviewer who will not be hired, and the reports that reviewer would have read, had nobody in the room. Naming the cost she moved onto others did not justify moving it. It let others judge whether she should have. Whether the agents will be better kept, and whether her forecast was right, nobody yet knows. If her forecast turns out right, that will not prove the decision was good. If it turns out wrong, that will not prove it bad.

## 36.6 The Company in 2035

What follows is a projection, and it is mine.

The idea that organisations should move beyond the old chain of managers is not new. Holacracy, devised by the software entrepreneur Brian Robertson, replaced managers with self-governing circles.¹³ Its best-known adopter, the online retailer Zappos, lost 18 per cent of its staff to a buyout offer in 2015 as it moved to the new system, and by 2020 it had been “quietly moving away from holacracy”, in a reporter’s words.¹⁴ Frederic Laloux’s *Reinventing Organizations* (2014) described companies run by self-managing teams.¹⁵ Its leading example, Buurtzorg, the Dutch home-care organisation a nurse named Jos de Blok started in 2006, was still working through self-organised teams in 2026.¹⁶ So the print edition’s claim that these experiments “largely failed at scale” is too strong. What the record shows is narrower. Changing the chart changes who is told what. It does not decide whether anyone answers, and a network of teams still needs someone who sees when two good local decisions add up to a bad one. A company can spread its decisions widely and still keep one place that watches across them.

My projection, for what it is worth, is this. By 2035, most large companies that use machines in decisions about health, money or work will have a named person for each such system. Their boards will hear when one is stopped. In 2025, of the eighty Fortune 100 companies one study read, about half said AI risk was among the risks their board oversaw. About two in five said a board committee oversaw AI.¹⁷ Both were three to four times the 2024 shares, but they count what companies wrote in filings, not systems stopped or owners named. I expect the change because laws, insurers and buyers’ contracts will ask for it, not because companies will have grown wiser. Many will meet it first in its weakest form: a name on a form, without the time, the evidence or the power to stop. A name on a form is not yet someone with the means to act. The change I mean is a named role with all three, and a successor when its holder leaves. My range is wide. It may come by 2030, or only after the kind of failure that makes a court ask what the board knew. What would prove me wrong is a world in which such failures are reported to boards as ordinary operating losses, with no name attached, and nobody minds. Nothing in this projection promises that a company which hears well will do better than one that does not.

This is the last chapter of Part VII, the part of the book about what to do. Each of its chapters looked for something worth doing on purpose, and each found something said in advance, where others could see it and hold someone to it, at a price: not a recipe. A year of practice changes something only when what a person finds becomes an act that others can see and contradict. Working alone, a person sells a promise whose price says what she will stand behind, and whose size must fit the hours she has to keep it. When machines can build a founder’s product and copy his manner, what is left to found is a few people allowed to see differently and held to one stated promise. And an organisation hears only where speaking has been met with a reply, and answers only through people who can be named. Together they change what Part VI meant. Part VI found that the account a person gives of her life is given to others, who may ask why. Part VII turns that account forward. What is worth doing on purpose is a promise said before the work, small enough to keep, open to contradiction, and paid for out of someone’s named hours. A person can make such a promise, and so can a small firm. An organisation does not practise; its people do, so it cannot make that promise for them. It can only make practising possible for its people, or make it impossible.

## Ending: A Page at the Start of the Shift

In the last week of September, Friedrich’s written decision reached Isabelle Moreau in three parts.

From January 2027, anyone at any NovaChem site who raised a concern about a machine system would get a written reply, within a month, from a named person. Each site would pay for the hours from its own budget. The release system that read batch records would be tested again by its makers at every site that used it. The tests would include night runs, and the sites would share the cost. Until Lyon’s test was done, the second reading of that one product’s night batches would continue, and Lyon would pay for it.

Isabelle did the sums at her desk. Since August the night shift had done the second reading in overtime. Kept up until the retest, it needed more night hours than the night shift had. For the rest of the year, two day technicians would move to nights. Neither had asked to. The night premium would come out of money set aside for training. The plant would release less of that product than it had promised Europe’s distributors for the last quarter of the year, and she wrote to them that week to say so.

Then she wrote one page to the night technician. It said what the model had found, including that half of what he had feared was wrong. It said what the qualified person had decided and what it was costing. It said that his name was on the record as the person who had raised it, in March, and that he should have had a reply then. She signed it, and gave it to him herself at the start of his shift.

### Practice

*What happened to what you said.* Take the last time you told your organisation something you thought mattered about a machine or a process. Write down to whom you said it, in what words and when. Write down what happened next, and whether anyone ever told you. If nobody did, ask, but only if asking is safe for you. If it is not safe, write down what makes it unsafe; that is the most useful thing you will find. If you lead people, take the last summary you sent upward and set it beside the report it summarised. List what you left out or softened, and why, and send the list to the person who wrote the report. Then find out who paid, in hours or money, for acting on the last concern your team raised. This takes an hour; if the hour is your employer’s, note who gave it. It promises nothing about your job, and it may cost you some goodwill. It shows whether your organisation can hear, and whether you are part of the reason it cannot.

## Words to keep

- **author (Hobbes):** the one who owns what another does in his name; the one who acts for him is the *actor*.
- **board of directors:** the people who govern a company on behalf of its owners and supervise those who manage it.
- **Caremark duty:** in Delaware law, directors’ duty to oversee; they fail it by having no system for hearing about a central risk, or by knowingly ignoring the warnings a system brings.
- **deep acting and surface acting:** in research on emotional labour, trying to feel what a job asks (deep), and showing feelings you do not have or hiding those you do (surface).
- **emotional labour:** Hochschild’s term for managing your own feelings and face so that others feel what the job requires.
- <strong><em>Pātimokkha</em></strong><strong>:</strong> Pāli, the rule of the Buddhist monastic order, recited every fortnight before the assembled monks.
- <strong><em>phāsu</em></strong><strong>:</strong> Pāli, ease or lightness; what the rule promises to the monk who declares a fault.
- **qualified person:** in EU medicines law, the person who must certify that each batch was made and checked as the law requires before it is released.
- **red flag:** in the law on a board’s oversight, a warning that reached the directors and called for them to act.
- **representer (Hobbes):** the one person or body through whom many people are made one.
- **review by exception:** checking a record by looking at what a system has flagged in it, rather than reading the whole record.
- <strong><em>sampajānamusāvāda</em></strong><strong>:</strong> Pāli, a lie told knowingly; in the 1881 translation, “a conscious lie”.
- **voice agent:** a machine system that speaks with callers and answers their questions.

## Sources

1. Eric Schwitzgebel, “If Materialism Is True, the United States Is Probably Conscious,” *Philosophical Studies* 172 (2015): 1697–1721.
2. Directive 2001/83/EC of the European Parliament and of the Council on the Community code relating to medicinal products for human use, Article 51(1)(a) (“each batch of medicinal products has been manufactured and checked in compliance with the laws in force”) and 51(3) (certification “in a register or equivalent document”); EU Good Manufacturing Practice, Annex 16, on certification by a qualified person and batch release. The clause continues “in that Member State and in accordance with the requirements of the marketing authorization” (legislation.gov.uk copy of the original text). Article 51 was in force in October 2026; the directive that will replace it had not yet been published in the Official Journal (Baker McKenzie, 12 October 2026). EU Good Manufacturing Practice, Annex 11, “Computerised Systems” (revision 1, January 2011; in operation since 30 June 2011), section 13, “Incident Management”: “All incidents, not only system failures and data errors, should be reported and assessed.” This is the text in force in 2026. A revised Annex 11 was put out for consultation from 7 July to 7 October 2025, but it had not been adopted by the end of October 2026: the Commission’s EudraLex Volume 4 page still listed the January 2011 revision (health.ec.europa.eu, checked 31 October 2026). On the qualified person’s own duty: Article 51(1) makes the qualified person responsible “without prejudice to his relationship with the holder of the manufacturing authorization”, and under Article 52 Member States see that the duties are fulfilled. On review by exception: ISPE’s definition, “an approach in which manufacturing and quality data are screened to present or report only critical process exceptions”, is quoted in Johan Zebib, “Review by Exception: Connecting the Dots”, *Pharmaceutical Technology*, 12 November 2019.
3. *Vinaya Texts*, Part I, trans. T. W. Rhys Davids and Hermann Oldenberg, Sacred Books of the East 13 (Oxford: Clarendon Press, 1881), “The Pâtimokkha”, Nidâna, p. 2 (sacred-texts.com). The passage is in the Mahāvagga, II.3.3.
4. *Mahāvagga*, Khandhaka 2 (the observance-day chapter), 3.3, trans. Bhikkhu Brahmali, SuttaCentral (segments pli-tv-kd2:3.3.9–16, bilara-data); Pāli from the Mahāsaṅgīti edition on the same site: *tuṇhībhāvena kho panāyasmante parisuddhāti vedissāmi*; *sampajānamusāvādassa hoti*; *āvikatā hissa phāsu hoti*.
5. Thomas Hobbes, *Leviathan* (London, 1651), chapter 16, “Of Persons, Authors, and Things Personated”, quoted from the Standard Ebooks text (based on the Online Library of Liberty’s transcription), which modernises capitals and punctuation and puts in quotation marks words that the 1651 edition sets in capitals: there, “A Multitude of men, are made One Person … For it is the Unity of the Representer, not the Unity of the Represented, that maketh the Person One.” The same chapter on an assembly: “And if the representative consist of many men, the voice of the greater number must be considered as the voice of them all” (eBooks@Adelaide text, also modernised). Swiss Code of Obligations (SR 220), Art. 718(1): “The board of directors shall represent the company externally”; Art. 718(2): it “may delegate the task of representation to one or more members (managing directors) or third parties (executive officers)”. Swiss Civil Code (SR 210), Art. 53: “Legal entities have all the rights and duties other than those which presuppose intrinsically human attributes, such as sex, age or kinship”; Art. 54: they “have capacity to act once the governing bodies required by law and their articles of association have been appointed” (English translations, which have no legal force; lawbrary.ch, checked 30 October 2026). On group agency (for “The other side”, 36.2): Christian List and Philip Pettit, *Group Agency: The Possibility, Design, and Status of Corporate Agents* (Oxford: Oxford University Press, 2011; ISBN 978-0-19-959156-5), not checked page by page; their view as stated in Christian List, “Group Responsibility” (PhilArchive, February 2019, minor edits October 2021, checked 31 October 2026): a group agent is “an organized collective that constitutes a goal-directed, intentional agent in its own right”; the three-member court, after Kornhauser and Sager (1986), as List cites them; “group responsibility is not generally reducible to individual responsibility”, while members may bear responsibility for the roles they played. Philip Pettit, “Responsibility Incorporated,” *Ethics* 117, no. 2 (January 2007): 171–201, doi:10.1086/510695, abstract (PhilPapers): “it is often possible for a corporate entity to be fully fit to be held responsible, when this is not true of the individual members; they may be able to make excuses that are not available at the corporate level”.
6. Dennis F. Thompson, “Designing Responsibility: The Problem of Many Hands in Complex Organizations,” in Jeroen van den Hoven, Seumas Miller and Thomas Pogge (eds.), *Designing in Ethics* (Cambridge: Cambridge University Press, 2017), pp. 32–56, on prospective design responsibility and named overseers (paraphrased). On reporting channels: Directive (EU) 2019/1937 on the protection of persons who report breaches of Union law, Articles 8(3) (private entities “with 50 or more workers”), 9(1)(b) (“acknowledgment of receipt of the report to the reporting person within seven days”), 9(1)(f) (feedback within “a reasonable timeframe ... not exceeding three months”), 16 (the reporter’s identity not disclosed without consent) and 19 (prohibition of retaliation), with the Annex, Part I.H, on medicinal products (legislation.gov.uk copy of the original text); in France, Law No. 2022-401 of 21 March 2022 (the loi Waserman) and Decree No. 2022-1284 of 3 October 2022: seven working days to acknowledge, feedback within three months, and reports may go directly to an outside authority (Jones Day, October 2022); a whistleblower in France may report a crime, an offence, “une menace ou un préjudice pour l’intérêt général” or a breach of the law (Law No. 2016-1691, Article 6, as amended in 2022; Berton &amp; Associés); in India, Companies Act, 2013, section 177(9) and (10): a “vigil mechanism” for listed companies and prescribed classes of companies, for directors and employees “to report genuine concerns”, with “adequate safeguards against victimisation”; the prescribed classes are companies that accept deposits from the public or have borrowed more than fifty crore rupees from banks and public financial institutions (Companies (Meetings of Board and its Powers) Rules, 2014, rule 7(1)).
7. Regulation (EU) 2024/1689 (the Artificial Intelligence Act): its duties fall on providers and deployers as such (Articles 16 and 26); none is addressed to a board of directors (artificialintelligenceact.eu, checked 23 October 2026).
8. Regulation (EU) 2026/1744 of 8 July 2026, amending Regulation (EU) 2024/1689: in force 27 July 2026; Annex III high-risk duties from 2 December 2027 (recital 40; EUR-Lex). See Chapter 25.
9. Swiss Code of Obligations (SR 220), Art. 716a(1)(5) (the board’s “non-transferable and inalienable duties”, including “overall supervision of the persons entrusted with managing the company, in particular with regard to compliance with the law, articles of association, operational regulations and directives”) and Art. 716b(1) (delegation of management “to individual members or third parties”), in the version in force since 1 January 2023; English translation, which has no legal force (lawbrary.ch; fedlex.admin.ch).
10. *In re The Boeing Company Derivative Litigation*, C.A. No. 2019-0907-MTZ (Del. Ch. 7 September 2021) (Zurn, V.C.), memorandum opinion on the motion to dismiss, slip op. at 13, 17 and 19 (the three quotations) and 65 (“accepts well-pled allegations as true”); the motion was granted in part and denied in part, and the oversight claim against the directors went forward; the facts are drawn from the complaint and documents produced to the plaintiffs, and “the Court ... accepts well-pled allegations as true”. The two crashes, Lion Air Flight 610 (29 October 2018) and Ethiopian Airlines Flight 302 (10 March 2019), “killed a total of 346 people”; MCAS was “the software designed to automatically push the airplane’s nose down in certain conditions” (House Committee on Transportation and Infrastructure, press release on its final report, 16 September 2020). On the duty to oversee: *In re Caremark International Inc. Derivative Litigation*, 698 A.2d 959 (Del. Ch. 1996) (“an utter failure to attempt to assure a reasonable information and reporting system exists”); *Stone v. Ritter*, 911 A.2d 362 (Del. 2006), restating the two ways (the directors “utterly failed to implement any reporting or information system or controls; or ... having implemented such a system or controls, consciously failed to monitor or oversee its operations”); *Marchand v. Barnhill*, 212 A.3d 805 (Del. 2019) (Blue Bell Creameries: no board committee, process or protocol for food safety; three people died in the 2015 listeria outbreak). In *Boeing* the court held both theories adequately pled: no reporting system for airplane safety, and, in the alternative, a board that ignored the Lion Air crash (Potter Anderson case summary; Jones Day, August 2019, on *Marchand*).
11. Settlement of the Boeing derivative litigation: Stipulation of Settlement (Office of the New York State Comptroller), Definition 8 ($237.5 million “to be paid by the Insurers to Boeing”) and Exhibit A, §VIII (an Ombudsperson Program within the Chief Aerospace Safety Officer’s organisation, “providing an additional channel for Unit Members in the FAA’s ODA program to raise work-related concerns”, in effect for at least five years); final approval on 23 February 2022 (Associated Press, 24 February 2022; Bloomberg Law, 23 February 2022); final order and judgment 22 March 2022.
12. Arlie Russell Hochschild, *The Managed Heart: Commercialization of Human Feeling* (Berkeley: University of California Press, 1983), p. 7. See Chapter 13. U. R. Hülsheger and A. F. Schewe, “On the costs and benefits of emotional labor: a meta-analysis of three decades of research,” *Journal of Occupational Health Psychology* 16, no. 3 (2011): 361–389, doi:10.1037/a0022876; abstract: “a final sample of 95 independent studies”; “substantial relationships of emotion-rule dissonance and surface acting with indicators of impaired well-being”; “deep acting displayed weak relationships with indicators of impaired well-being”. The figures are correlations.
13. Brian J. Robertson, *Holacracy: The New Management System for a Rapidly Changing World* (New York: Henry Holt, 2015).
14. Aimee Groth, “Zappos has quietly backed away from holacracy,” *Quartz*, early 2020 (cited as 3 February 2020 by Cornell ILR’s Catherwood Library; qz.com/work/1776841): “18% took a generous buyout offer” in 2015; “in the last few years, Zappos has been quietly moving away from holacracy.”
15. Frederic Laloux, *Reinventing Organizations* (Brussels: Nelson Parker, 2014).
16. Susanne Ladopoulos, ZDFheute, 13 June 2026, on Buurtzorg: begun by the nurse Jos de Blok in the Netherlands in 2006, with “selbstorganisierte Teams in der Nachbarschaft” (self-organised neighbourhood teams).
17. Pat Niemann, “Cyber and AI Oversight Disclosures: What Companies Shared in 2025,” Harvard Law School Forum on Corporate Governance, 28 October 2025, based on his report for the EY Center for Board Matters (October 2025), which covers the 80 companies on the 2025 Fortune 100 list that had filed proxy statements and Form 10-Ks by 31 July 2025: “Nearly half (48%) specifically cited AI risk as part of the board’s oversight of risk”, against 16 per cent in 2024; “Around 40% disclosed charging at least one board-level committee (usually the audit committee) with AI oversight responsibilities, almost four times the 11% that did so in 2024.” The figures count what companies disclosed in their filings; they do not measure whether any system was stopped or had a named owner.

*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 organisations are not conscious and yet that “the organisation behaves as if it can see” (“This is not a metaphor”); that “a conscious organisation adapts under unknown ones”, while “In the AI era, unknown conditions are the only kind”; that “The practices were installed. The consciousness was not.” and that Meera might “pour her own consciousness into the space”; that Meera had “built a conscious division”; that the five signals measured “an organisation’s consciousness”; that the night technician “has something the AI does not”; that Friedrich had built “an intelligent company” but “not ... a conscious one”; that Meera’s “consciousness had limits” though she saw her own organisation “with extraordinary clarity”; that “Consciousness does not compress”; that for a drug company using high-risk AI in the EU in 2026 board-level AI governance is “a legal requirement”, that NovaChem’s adverse-event system was “classified as high-risk”, and that NovaChem had got through the EU AI Act before most of its competitors understood it; that “The difference between the two modes is consciousness”; that a “consciousness threshold” decides when to hire a human, resting on “a relationship between conscious beings”; that keeping people is “in the long run, cheaper than the alternative”; that a potter in Seto “doubled his output and held his prices, because the market was paying for his eyes”; that Holacracy and Teal “both largely failed at scale” and that “Most Teal experiments reverted to conventional management within five years”; that organisations which develop consciousness will do better over the next decade, and that consciousness, once developed, “compounds” and is hard for a later competitor “to catch”; and that “consciousness is not installed. It is uncovered.” It said that Meera’s governance team caught the adverse-event bias (“Three people”), that the model underweighted “African and South Asian” sites with “aggregate accuracy above ninety-seven per cent”, and that Hyderabad had a concern log with “one hundred and forty-seven” entries. It explained Lyon by “the professional norms of a French pharmaceutical plant”, quoted a nurse named Astrid at a Copenhagen conference as if reporting, added a Sikh farmer in Punjab, a warehouse manager in Mombasa and a ceramics master in Seto, put a sample-check drill in São Paulo, said that Meera spent “the following year” on a diagnostic, sent Meera home by air on a Sunday evening as print Chapter 37 does again, and ended on “Consciousness without action is philosophy” and a bridge to the next chapters.
- *Section titles in the print edition:* 36.1 “Diagnosing Organisational Consciousness”, 36.3 “Building a Culture of Awareness”, 36.4 “The Board’s Responsibility: Governance as Consciousness” and 36.6 “The Conscious Company in 2035”; 36.2 and 36.5 keep their print titles.
- *What I now hold:* an organisation is not conscious, and nothing in this chapter depends on whether any group could be. It does not see: it is told, or it is not, and people tell where telling has been answered before. It is one by procedure, not by mind. It answers through a representer who speaks in its name and knows only what reaches him (Hobbes), and it hears only through people who speak, which they do where a question is really asked and speaking is made light (the *Pātimokkha*). Neither half works alone, and the staff are not the authors of the representer’s choices. A written reply to every concern, from a named person, is the least evidence that saying something was worth the risk. Swiss law makes a board’s supervision non-transferable; the AI Act imposes nothing on boards, and none of Meera’s systems is high-risk; Boeing’s case shows a board without a channel for the hard reports. Governance is a form of attention: an agenda and a channel. “Mark every point where a person must be present because the cost of their absence is larger than any efficiency gain” is a decision about whose costs count, not a sum, and “present” means a position held. A machine can find half of what a worried person suspected was wrong, and read every report a board receives; what it is not given is the duty to ask. The adverse-event bias was caught by an automated audit, and the threshold was decided by Basel’s physicians in August 2025. Buurtzorg still runs on self-organised teams. A projection about companies in 2035 is mine, with a range, and promises nothing. A reply within a month is the longest wait, never the route for an urgent doubt, and it does not protect the one who spoke. An account is the least owed to people whose work a machine takes. A reply is the start of answering, not the whole: examination, a decision, reasons, repair and change follow, with a route to challenge the reply; where the money for acting on a concern comes from can punish the speaker through his colleagues. A signature can carry only as much as the reading behind it. Delaware’s oversight duty reaches a board with no channel and one that knowingly ignores red flags, not one whose warning came in accurate, bland words. A fallback needs people still able to do the work by hand. A name on a form is not yet someone with the means to act. The EU’s 2019 directive on reporting, in French law since 2022, already requires a channel that acknowledges and replies to suspected breaches; a reply rule reaches the doubts that channel was not built for.
- *What changed my mind:* the print’s own sentence that organisations are not conscious; the monks’ rule that reads silence as purity only after asking three times and promising ease to the one who speaks; Hobbes on representation; Swiss company law and the Delaware court’s account of Boeing’s board; the AI Act as amended in 2026 and this book’s own finding that none of NovaChem’s five systems in Meera’s operations is high-risk; this book’s earlier account of who caught the adverse-event bias; the reports on Zappos and Buurtzorg; the rule that a machine’s ability carries a date and never carries the argument; Delaware’s cases on a board’s duty to oversee; the EU’s directive on reporting breaches, as French law carries it out; and three outside readings of an earlier version of this chapter, in October 2026.
- *What survives:* Lyon and the pilot that failed quietly; the coffee and the carpet; “The questions were specific. They were professional. They were, every one of them, about machinery.”; the handwriting line; “Not because they had nothing to say. Because nothing in their working lives had ever rewarded the act of saying it.”; “What are your people already noticing that no one is asking them about?”; the night shift and the thinner check; Friedrich in Basel and the sentence he asked about; “A memo from the CEO is not permission. A scheduled review is not permission.”; “I cannot do what you are asking. Not the way I did it in Hyderabad.”; “I can give them the questions. I cannot give them the eyes.”; “Culture is not the poster in the lobby. Culture is what happens when the poster’s author has gone home.”; Isabelle’s question at the end of meetings; the reward for the problem that never becomes a crisis; performing governance against practising it; “Someone at the highest level could say stop”, narrowed; the case that a team left only the hardest calls loses its footing; Meera’s map, now with its rule written at the foot.

*Next in the series:* [*Part VIII: The 2035 Human (overview)*](https://nithinpalal.com/ai/ai-proof/part-8-the-2035-human)*.*