AI-Proof: What Machines Can Never Be — Volume I: Awakening, Part III: The First Awakening, Chapter 9: The Architect Mindset

Part III turns from what the machine is to the practice of working beside it. This first chapter asks how to decide what the work is for, and who gets to decide.

← Previous: Part III: The First Awakening (overview) · Series hub · Next: Chapter 10: Building Deep Human Capital →

Priya spent two weeks doing her best work on a part of the product that was already being thrown away. A machine makes you faster at anything, including the wrong thing. Who decides what the work is for?

In Brief: For readers who do good work and wonder whether it is the right work. Chapter 9 opens Part III with a costly mistake: a test lead in Bengaluru builds an excellent design for a module her company has already decided to retire. It separates the operator, who does the work, from the architect, who asks what the work is for, and it asks what is left of that difference once machines can plan as well as do. Aristotle’s “architectonic” arts, the Indian scholars’ four questions before any treatise, Goodhart’s law and Donella Meadows help answer it, with a textile maker in Coimbatore, his floor supervisor, and a brand strategist in São Paulo.

A machine makes you faster at anything, including the wrong thing. Before the work, ask what it is for, who decided, and who will answer if it is the wrong building.

Opening: The Corridor Outside Planning

Anand was apologetic. That was the detail Priya would remember later: not the wasted work, not the embarrassment, but the look on his face when he told her. He assumed she already knew.

She did not know.

It was the first week of January 2025. Priya Chandran leads a team of twelve testers at a software company in Koramangala, in Bengaluru. She had spent the last two weeks of December redesigning the tests for Module 7. Module 7 connected the product to its payment gateway, and it had been rewritten twice in a year. She had brought her team into the work. She had used the machine to map which parts of the code depended on which. It modelled how the module might fail and wrote tests for rare cases the old tests missed. The work was rigorous. By any professional standard it was good. She had not just done tasks. She had stepped back, looked at the whole system and redesigned it from the foundations. Her manager, Rashmi, had praised it. Priya was thinking more strategically now, Rashmi said. More like an architect.

Now Anand, the lead developer on the payments team, was standing in the corridor outside the planning room. The thing he assumed she knew was this: Module 7 was being retired. The company would replace it with a payment service from another firm. The decision had been made three weeks earlier, in a meeting Priya had not been invited to. It was written in the product plan for the year, which she could open at any time and had not opened since September.

Two weeks of careful work, done on a part of the product that was about to be thrown away.

Priya thanked Anand and walked back to her desk. She did not open her laptop. For a long time she looked at the diagram of her new design, pinned to the wall beside her monitor: the clean lines, the order of every check. It was the best work she had done since the morning in October when the machine’s results turned her dashboard green and made her ask what she was for. She had asked. She had found an answer. And the answer was wrong.

Not the method. The method was right. The target was wrong.

A machine that works within the wrong frame (the wrong purpose, limits or measure of success) does not make a small mistake. It makes a large one, quickly, with the polished surface of good work. So this chapter asks a question that the machines have made urgent. When a machine can do the work, and can even plan it, how do you know you are working on the right thing, and who gets to decide?

9.1 Operator vs. Architect: The Fundamental Division

There is a way of working that most people learn early and never unlearn. You receive a task. You do it. You receive the next one. The tasks add up to a career. From the inside it feels like progress, because you are always moving. From the outside it may be a circle, but you cannot tell, because you never lift your head.

Call this operating. Many operators are brilliant; operating has nothing to do with a lack of intelligence. It is a habit of attention. The operator asks how to do the task. The architect also asks whether the task should be done at all, and what it is for. The operator works inside a system to make it run. The architect stands far enough outside it to ask whether the system is aimed at the right problem.

The word is old, and its history helps. In Greek, an architektōn was a master builder: not the person who laid the bricks, but the one who knew what the building was for. Aristotle opens his Nicomachean Ethics with this idea. Some arts serve others, he says: “bridle-making and the other arts concerned with the equipment of horses fall under the art of riding, and this and every military action under strategy”. In every such chain, “the ends of the master arts are to be preferred to all the subordinate ends”.¹ He called the master arts “architectonic”. A bridle-maker can make a perfect bridle for a horse nobody will ride.

Aristotle also said something this book must refuse. In the Metaphysics he wrote that “the masterworkers in each craft are more honourable”, because “they know the causes of the things that are done”. The manual workers, he said, are “like certain lifeless things which act indeed, but act without knowing what they do, as fire burns”.² He was describing a hierarchy of people. This chapter describes something else: two ways of working that any person can move between in the same day. A junior engineer who says “this design will not scale; we should rethink it before adding features” is doing an architect’s work, whatever her title. A vice-president who spends his days checking dashboards against targets someone else set is operating, whatever his. The difference is not rank. It is the question you ask.

But keep two things apart, as Chapter 5 had to. The architect’s question is open to anyone, on any day. The seat where the purpose of other people’s work is decided is not. It is scarce, and it is rarely offered to the person whose task was defined in a room she could not enter. Asking the question does not give you the seat. The ability to frame a problem does not, by itself, give anyone the authority to decide which problem should be solved.

Chapter 5 drew a related line. There, the operator starts a process, and the governor watches a process that is already running and knows when to question it. The architect adds a third kind of attention. The operator runs the system. The governor watches it. The architect asks what it is for. These are not three steps up a ladder. Good work often needs all three at once. A tester’s close attention to one failing check is often where a deeper problem is first seen. A governor who notices a pattern of failures may be the first to see that the goal itself is wrong.

The difference has always mattered. The machine makes it urgent, because machines are now very good at operating. Give one a defined task, such as reviewing documents, writing a report, testing code or drafting a letter, and it works at a speed no person can match. Chapter 4 told how one company said in February 2024 that its assistant did the work of 700 full-time agents. Fifteen months later, it began recruiting human agents again. Replacement is real, uneven and sometimes reversed. But the direction for defined tasks is clear.

And the machines have moved past defined tasks. By 2025 the systems called agents could take a goal, break it into steps, carry out the steps and check the results.³ One research firm predicted in 2024 that by 2028 “at least 15% of day-to-day work decisions will be made autonomously through agentic AI”.⁴ Ask a model in 2026 to plan a project and it will plan it. Ask it to question your brief and it will question it, often well.

The print edition answered this by saying that the architect has “a felt sense of purpose” that machines lack, and that machines “compute; they do not care”. This edition cannot rest on that. Whether a machine can care is an open question, and the Prologue made a rule against building on what machines cannot yet do. So the chapter has to find the difference somewhere else, and Priya’s mistake shows where to look.

9.2 From Doing to Designing

Six weeks before Anand stopped her in the corridor, Priya had her first architect’s thought.

It was a Monday morning in November 2024. She was looking at her list for the week, the same kind of list she had written every Monday for eight years.

Review the results of the machine’s overnight run. Flag anything odd to the developers. Attend sprint planning. Update the test documents for Module 7. Write the summary report for the quarterly review. Follow up with Anand on the payment edge cases. Check that Sneha’s new hire can reach the test environment.

Seven items. All clear. All doable. She could tick them off one by one, and by evening she would feel productive. Rashmi would see output.

But seven weeks of watching the machine do the work she used to do had changed something. Since late October she had kept a page of the decisions she made each week, as Chapter 5 told, and the page had trained her eye. She looked at the list and saw, for the first time, that every item was a task. Not one was a question about whether the task should exist.

Review the results. But why these results? Who decided which modules mattered most? Was the testing plan still right, or was it left over from a time when the product had different users and failed in different ways?

Update the test documents for Module 7. The module had been rewritten twice. The documents described a design that no longer existed. Updating them would not only waste time. It would write down, again, assumptions that were no longer true.

Write the summary report. But what was it for? Who read it? What did it change? She had written it every quarter for three years. Nobody had ever told her what happened to it.

She had the feeling of seeing a familiar room from a new angle. The same furniture, the same walls, but shifted a few degrees, and the layout that had seemed natural for years now looked arbitrary. She was not looking at a plan. She was looking at a habit.

She rewrote the list as questions. Where it had said Review the results, she wrote: Does our testing plan match the risks the product has now? Where it had said Update the documents for Module 7, she wrote: Is Module 7 tested for the product as it is now, or as it was eighteen months ago? Within weeks she had persuaded Rashmi to replace the quarterly summary report with the team’s own version of her page of decisions: not what the team did, but what it decided and why. The machine that had once felt like a threat to her became one part of a testing system she was designing.

She was learning to ask an architect’s question. She was also about to learn that someone who asks it without seeing the whole site makes the same mistake as an operator, only better built.

9.3 The Wrong Building

The Module 7 failure was not a failure of thinking. It was a failure of scope.

Priya had learned to step back from a task and ask whether it served its purpose. She had not yet learned to step back once more and ask whether the system itself still had a purpose. She had asked: How should we test Module 7? She had not asked: Should we be testing Module 7 at all?

There are three questions here, and they come in order.

1. How do I do this task? This is the operator’s question.

2. Is this task designed well? This is where Priya arrived in November.

3. Is this the right task? This is the question she skipped.

The gap between the second and the third is where most architects fail. It is the gap between redesigning a system and asking whether the system should exist. The management writer Peter Drucker put the cost in one sentence in 1963: “There is surely nothing quite so useless as doing with great efficiency what should not be done at all.”⁵

The third question might have taken five minutes. One look at the product plan. One message to Anand. Might, because a line in a plan is not always written so that it reads as a decision. But she did not ask, because the answer seemed obvious: Module 7 existed, so it needed testing. The assumption was so ordinary that it did not feel like an assumption. It felt like the ground. A new way of seeing becomes its own kind of blindness when you step back one level and stop.

The machine did not cause the mistake, but it made it bigger. It mapped in hours what would have taken days. It wrote tests overnight. Before, the work would have taken longer, and someone might have caught the mismatch halfway through. Speed removed the pause in which doubt used to arrive. Chapter 14 will return to that pause and how to build it back.

And the machine could have prevented the mistake, if she had asked. Any tool that could read the company’s documents could have found the line in the product plan in seconds. The information was not hidden. It was simply outside the question she asked. Neither she nor the machine asked whether the work was still needed. And a system could have been built to ask. A testing system that can read the product plan could check, before any redesign begins, whether the part under review is due to be retired. Whether it does is a choice about how a company connects its decisions to its tools, and machines are well able to make the check.

But this is not only Priya’s story, and the print edition told it as if it were. Three different things failed, and they are not the same. There was a document she could have opened and had not. There was a room she was not in: the decision was made in a meeting she was not invited to. And there was a chain above her in which knowing the purpose had already come apart from answering for the work. Rashmi praised two weeks of her team’s work without mentioning the retirement, perhaps because she did not know either. In Chapter 7 a position is a place no skill can fill: the one whose life it is, the one who answers, the one with the right to decide. That chapter said a position is full only when the person holding it can see the evidence. Here that needs sharpening. The plan was the evidence, and Priya could see it. What she lacked was the decision itself, made plain to her in time to stop the work. Some companies ask their people to think like architects, then decide the purpose of their work in rooms those people cannot enter. They have given them the title and kept the site plan. Do not ask people to question the purpose of their work while keeping from them the decisions that set it. Priya should have asked. Someone should also have told her. So the two weeks have more than one author. Each answers for a part: Priya for the plan she did not open, and the people who made the decision for the silence that followed it.

There is one more danger, after the discovery. Work already done pulls on the next decision. Having spent two weeks, a team wants the two weeks to count, and looks for a reason to finish. Economists call this the sunk-cost error. Priya did not open her laptop that afternoon, and that was right. Many teams keep building.

9.4 Systems Thinking as Survival

An architect needs a particular kind of attention: the ability to see how a decision in one part of a system produces a result in another, often months later, often where nobody is looking.

In April 2025 a production-scheduling system arrived at Rajan Subramanian’s factory in Coimbatore. It came with a larger set of changes that Chapter 18 will tell. The system works. It decides which looms run which cloth, cuts waste, and predicts rush orders well enough that Rajan has not been caught short since it arrived. On every measure it was built to improve, it improves.

By September Rajan had noticed something. The bolts coming off the speciality looms felt different, not in quality but in number. The hand-finished fabrics that carry his family’s name, the ones his oldest buyers ask for by name, were thinning. At the far end of the floor Kamala, who is sixty-eight and has finished those fabrics by hand for most of her working life, had less to do each week. Nobody had decided to reduce the speciality work. The scheduler aimed at output per loom-hour (the cloth a loom makes in an hour of running) and less waste. So it had been moving loom time slowly towards standard weaves, where the gains were largest. It moved time away from speciality work, where set-up takes longer, runs are shorter and waste is higher.

The shift was rational. If output per loom-hour is the only measure, the standard weaves win. The machine was doing exactly what it had been asked to do. The problem was what it had not been asked.

Rajan’s first impulse was to override it. Muthu stopped him.

Muthu has supervised Rajan’s floor for nearly twenty years. Chapter 4 noticed that nobody had asked him what the machines should decide. Now he spoke without being asked, one morning over the production reports. “The machine is not wrong about the numbers, sir. The Germans are ordering fewer speciality pieces. It is we who have not noticed.”

Rajan checked. Muthu was right. Over eighteen months, the speciality orders from his buyer in Hamburg, Jürgen Weiss, had fallen by a third. Weiss had not said so. The relationship was too old and too polite for that. The fall was in the order book, and the scheduler had been following it. On demand, the machine had read the records better than Rajan’s feeling about his own customer. Chapter 3 gave the test for when an expert’s quick sense deserves trust: clear signals and fast feedback. Rajan’s fingers pass it. His sense of what a buyer still wants, after years without a direct question, does not.

The environmental scientist Donella Meadows once listed the places where you can change a system, from the weakest to the strongest. At the bottom she put “constants, parameters, numbers”: the settings people argue over most, though changing them changes the system least. She called adjusting them “diddling with the details, arranging the deck chairs on the Titanic”. Near the top she put “the goals of the system”.⁶ The scheduler was very good at numbers. Overriding it would only have been arguing about numbers too. The question Rajan had stopped asking was the goal: what was the speciality line for?

So he called Weiss, who has bought from him since 2013, and said something he had never said to a buyer in nearly thirty years of business. “I want to understand what you need from us that you cannot get from anyone else.”

The call was long. Weiss had been buying standard weaves through a sourcing platform for a year, as Chapter 7 said some of Rajan’s buyers do for their cheaper lines. He had not mentioned it because he did not want to damage the relationship. What he said he still wanted from Rajan was smaller and more particular: a few hand-finished pieces each season, with someone who knew the cloth standing behind every bolt. Neither of them mentioned the new contract, and the clause about whose inspection would be final was still unsettled between them.

Rajan did not obey the machine, and he did not override it. He changed the goal. He reduced the speciality line, raised its price and made it a small seasonal collection, with Kamala’s name on the finishing tag for the first time. The scheduler now gives the speciality looms a fixed share of time each week, set by Rajan, not by the measure.

Changing the goal is not always wiser than changing the numbers. A manager can change a goal to protect his pride, or to move a cost onto people who have no say. Rajan’s change is believable because it followed the evidence: Muthu’s reading, the order book, Weiss’s answer. It is still a bet. And the scheduler is not the enemy of the craft. Given a goal that includes the speciality line, it plans the short runs more cheaply than the old floor ever did. A machine can help pursue a purpose. It cannot, by itself, show that the purpose is worth what it costs. This is not a limit of today’s machines. Whether a purpose is worth its cost depends on whose cost it is and who agreed to bear it. Those are questions of right, and a better machine does not settle them.

Notice, too, who could see and who could change. Muthu could see the evidence. He read the order book and spoke without being asked, which Chapter 4 said nobody had invited him to do. He was not on the call. The right to change what the scheduler was for stayed with the man whose name is on the factory. Noticing is not the right to decide.

Whether this was right is not yet known. By the autumn of 2026 the collection had run for two seasons and covered its costs, narrowly. Weiss was buying it; two other buyers had not decided. And the collection now rests on one woman’s hands. Kamala is sixty-eight. Nobody else on the floor can finish the cloth the way she does. Her name on the tag recognises her skill. It does not pass it on. Chapter 6 asked what is lost when the work that forms a person disappears. Here the work has not disappeared. It has gathered in one person’s hands. Will anyone learn it? Does Kamala want to teach it? Who would pay for the learning while demand is uncertain, and how will the value of her name be shared? Rajan has not decided any of this. That is not only a risk to next season. It is a debt.

9.5 The Discipline of Framing

Before machines, the cost of doing the work acted as a brake on bad framing. If a prototype took three months to build, someone usually asked in the second month whether it was solving the right problem. The slowness made room for the doubt that saves projects. The machine removes that slowness. A team now builds prototypes in days and tests twenty versions of an advertisement in an afternoon. Speed feels like progress, and the feeling makes it harder to ask whether the fast thing is the right thing.

So far the chapter has shown two ways a frame can fail. Priya’s task was no longer needed. Rajan’s goal left out something he valued. There is a third: the measure stops standing for what it was meant to show. In São Paulo, a brand strategist named Luísa Ferreira met it at a larger scale than Priya’s. Luísa, who is thirty-six, runs content strategy for a consumer goods company. Three years ago, when she was promoted, she built the company’s measures for its social media herself. They counted how many people engaged, how fast posts were shared, how long people stayed on a page and how many clicked through to a product. The measures were good. They measured what the company needed at the time, which was attention.

In early 2025 the company brought in an AI system to write its content. Output exploded: a hundred pieces a week, all on-brand, all shaped for the platforms. Every measure went up. Six months later the company’s annual study of how people felt about the brand came back. More people recognised it. Fewer people liked it. The study could not say why. Prices, competitors and a dozen other things had changed in the same months. But the content was scoring well on the measures while the thing the measures stood for was falling. Engagement had been a stand-in for what the company wanted: that people trust and like the brand. The stand-in had come loose.

Economists know this as Goodhart’s law, after the British economist Charles Goodhart, who noticed it in the 1970s in the measures governments used to control money. The anthropologist Marilyn Strathern later gave it the form most people quote: “When a measure becomes a target, it ceases to be a good measure.”⁷ People have always gamed measures. Machines game them harder, because they pursue a target with no distraction and at great speed. In my view, a measure that took years to wear out under human pressure can wear out much faster under a machine’s. Chapter 8 showed the same thing in the scoring of the machines themselves: reward guessing, and you get confident guesses.

This part cost Luísa something. She had built those measures. The dashboard that now showed false success was her work, and the machine was aiming at targets she had set. The machine had not failed. Her frame had. She proposed a change: fewer pieces, a slower and more distinctive voice, success measured by what people felt about the brand rather than how often they saw it. She could not prove it would work. The data supported the old approach, because the old approach had been built to score well on the old measures. The company’s leadership approved it, reluctantly. Her standing to propose it came partly from having built the original measures: she would answer for both the old and the new. The decision to adopt it was theirs. Whether the slower voice works is not yet known.

So what does the discipline of framing require? Indian scholars had an answer long before dashboards. Before beginning any serious treatise, they set out four preliminary matters, the anubandha-catuṣṭaya, “the four connections”. A summary of Vedānta from about 1500, the Vedāntasāra, lists them in one line: adhikāriviṣayasambandhaprayojanāni.⁸ In plain words:

• adhikārin: who is this for, and who is ready for it?

• viṣaya: what is it about?

• sambandha: how does the work connect to its subject, and so to its purpose?

• prayojana: what is it for?

One of these travels less cleanly than the others. In the Vedāntasāra, the adhikārin is the qualified student, the person fit to receive the teaching. In an office I use the word for a different question: whom does the work serve? That is my extension, not the text’s.

The habit is older than that text. The Mīmāṃsā philosopher Kumārila, whom Chapter 8 met, wrote in the seventh century: sarvasyaiva hi śāstrasya karmaṇo vāpi kasyacit | yāvat prayojanaṃ noktaṃ tāvat tat kena gṛhyate || In my rendering: “Of any treatise, or of any action at all: until its purpose has been stated, who will take it up?”⁹

I am applying a scholar’s rule for reading texts to the work of an office and a factory, and the fit is close. The four preliminaries ask what a treatise is for before anyone spends years on it. In an office the same habit becomes four questions, and a person who frames well keeps returning to them. She asks them regularly, the way a pilot runs checks before take-off. She does not expect something to be wrong. The one time it is wrong pays for every check. Like a pilot’s checks, they are short, and they end. Asked every hour, they would stop the work they are meant to protect. Once they have answers, the work begins.

What problem are we solving, and for whom? Priya answered it for the wrong module.

What assumptions have we stopped noticing? Priya assumed Module 7 was permanent. Luísa assumed her measures still measured what mattered. Rajan assumed his buyers still wanted what they used to.

Whose view are we missing? Priya was missing the product team’s. Rajan was missing Weiss’s, and Muthu’s. In each case the missing view was not hidden. It was available, and outside the frame.

What would make this work pointless? This is the hardest, because it asks you to imagine your own effort wasted. It would have saved Priya two weeks and Luísa six months.

In all three stories the question of purpose did the most work. Two warnings go with it. First, a frame is not set once. Module 7 was worth testing until a meeting in December. Rajan’s speciality orders held until they did not. A good frame is one that can be revised when the evidence changes, which means deciding in advance what evidence would change it. Second, a clear frame is not the same as a good one. A company can answer all four questions crisply and still choose a purpose that pushes its costs onto people with no say. Every frame leaves something out, and choosing what to leave out is a use of power. The questions make a purpose visible. They do not make it right.

What would prove this wrong?

Hypothesis. Earlier in this section I wrote that a measure which took years to wear out under human pressure can wear out much faster under a machine’s. That is my view, not a finding. Luísa’s story fits it but cannot test it, because her company’s study could not say why liking for the brand fell.

The claim would be wrong if measures held up as long when a machine pursues them as when people do. A fair test would set people and a machine to chase the same fast measure, such as clicks, in the same kind of work. Beside it, the test would track a slow measure of what the fast one was meant to show, such as trust, complaints or customers who return. If the slow measure fell no faster under the machine, my claim fails.

The place to look is firms that kept both kinds of measure through the years in which machines took over the work. The slow measure has to have been taken before the change as well as after it, or nothing can be compared. One firm’s story, however vivid, is a reason to look, not a result.

9.6 What Architects Do with AI

There is a way of using a machine that makes you more productive, and a way that makes your work more useful. They are not the same.

Take two people writing an analysis of India’s textile exports. The first asks: Write an analysis of the Indian textile export market. She gets a competent, general document that gathers public information and reaches the usual conclusions. It is useful as a start, and no more than a start. The second says what she is analysing: the hand-finished premium segment. She says what worries her: buyers moving from long relationships to sourcing platforms. She asks the machine to attack her own assumption that heritage commands a higher price in a market that is becoming a commodity. And she says what success means: keeping the margin while buyers change. The machine’s answer inside that frame is far more useful. The machine is the same. The frame is different.

The print edition said the architect’s output “cannot be replicated”, because it comes from a particular person’s consciousness. That is too strong. Frames can be copied, taught and, increasingly, proposed by the machine itself. Ask a model in 2026 what would make your project pointless and it will give you a list, sometimes a good one. If Priya had asked that in December, it might have asked her, in return, whether Module 7 was staying.

So the difference is not that only a person can think of the frame. It is that the frame needs someone with the standing to choose it and the obligation to answer for it. A machine can propose ten frames for Rajan’s speciality line. Which one his factory lives by is decided by Rajan, with Weiss, Muthu and Kamala, because it is their work, their money and their names on the tags. And if the frame is wrong, it is Rajan who will be asked why. None of this promises that anyone will pay for the frame. Chapter 5 said that what is scarce is not always what is paid. And the person who asks the better question may still not be the person who chooses.

9.7 Who Decides What the Work Is For

The print edition closed this chapter by arguing that “the architect mindset is an expression of consciousness”, and that this “is also why AI cannot architect”. That section is withdrawn. As a skill, framing is like the skills Chapter 7 rated: it carries a date. Machines already question assumptions when asked, propose plans and test a brief against its own aims. They will get better at it. If the architect’s place depended on being the only one who could think of the right question, it would be shrinking ground.

What does not shrink is the thing Priya lacked in December and Luísa partly had. It has three parts. Chapter 7 named the first two as positions. The third is one of the conditions that keep a position full.

The first is the right to decide what the work is for. That right sits with people whose work, money and lives are at stake, and they are often several, with different kinds of standing. Rajan owns the factory and has the say. Kamala’s skill gives her a stake in what happens to the craft. Weiss has a stake in what he buys. Muthu could see what the others missed. Having a stake is not the same as having the say, and having the say is not the same as being right. What matters is that the arrangement is clear: who decides, whose interests must be weighed, who can object, and who answers. Clear is not the same as fair. Kamala has a large stake and little say, and a clear arrangement can keep it that way. But an arrangement nobody can see cannot even be challenged. The Prologue’s third never adds one thing to that arrangement. Being intelligent will never give a machine the right to decide for people. A machine that proposes the best frame has made a proposal. Someone still has to choose it. A company can let a machine set a schedule, and often should. Even if it lets a machine choose what the work is for, someone chose to let it, and that person answers for the choice. Better framing by machines would not change this. What would reopen the question is a machine with its own stake in the work: aims formed over years, failures it carries, and changes its makers did not decide. Then the argument would have to be made again.

The second is the duty to answer for the frame. Luísa’s standing came partly from the fact that she would answer for both the old measures and the new. When a frame is wrong, it is not enough that someone chose it. Someone must explain, repair and stand behind it. Often the duty is shared among the people who set the goal, chose the measure and put the system to work, in proportion to what each knew and could change.

The third is access to the evidence, in the sharper form 9.3 found: seeing the decision that made the work, in time to refuse it. Priya had the habit and the duty. She did not have the decision, and she did not have a seat in the room where it was made. Machines are taking over more of the operating, and some of the deciding, so organisations will ask more people to act as architects. The test is whether they also give those people the site plan. A second test is what happens to a person who asks. Where a question about purpose counts against the person who asks, people stop asking. Then architect is a title and nothing more.

And the architect can be wrong. Stepping back changes which facts you can see. Some come into view; others, often the details that make a plan work, drop out of it. Priya stepped back and still missed the retirement. Rajan’s sense of his buyer was wrong until Muthu and the order book corrected it. The architect is not the person who sees everything from above. The architect gathers what others can see, the floor, the order book, the buyer, and keeps the frame open to correction.

Run this through two futures, as Chapters 7 and 8 did. If machines grow quickly, they will soon propose frames as well as most people can, and the skill of framing will matter less as a skill. The right to choose and the duty to answer will matter more, because there will be more proposals to choose among and larger consequences to answer for. If machines grow slowly, framing stays a skill worth learning for longer. In both futures, the question at the end of the work is the same: who decided what this was for, and will they answer if it was the wrong building?

Practice

Ask the four questions of one piece of work. Choose one piece of work you will spend the most time on this week. Before you start, write one line for each: What problem is this solving, and for whom? What am I assuming that I have stopped noticing? Whose view am I missing? What would make this work pointless? Then add a fifth: Where was the purpose of this work decided, and could I see that decision? And a sixth: What would make me change this frame? If you cannot answer the fifth, find out before you begin. Ask the person who would know. You may also ask a machine to attack your answers; it will often find a gap you missed. Finding out is not the same as being answered. If no one replies, decide how much work you will risk before you hear back, and tell the person who should have told you. Most weeks the answers will change nothing, and you will begin with more confidence. Some weeks they will save you two weeks.

Ending: The Card on Monday

Three weeks after the corridor, on a Monday morning, Priya took a card from her jacket pocket before she opened her laptop. She had written the same line on a new card every Monday since the corridor: What would make this week’s work pointless?

Most weeks the answer was nothing. This week it was a question about a change the platform team was planning, and before her first meeting she sent a message to Deepak, who ran that team. That evening she walked home through Koramangala. She was not yet comfortable calling herself an architect. The word felt too large for someone who had just spent two weeks on the wrong building. She still felt the pull of the old habits: the clear task, the ticked box, the familiar self who finds the bugs. The pull did not go away. She had begun to notice it. Deepak had not replied. The answer might be nothing. The question had cost her two minutes.

Words to keep

• anubandha-catuṣṭaya: “the four connections”; in Indian scholarly tradition, the four matters settled before beginning a treatise: for whom, about what, how the work connects to its subject, and for what purpose.

• architect: in this book, a person who asks what the work is for, not only how to do it; a way of working, not a rank. Asking the architect’s question does not give the right to decide what the work is for.

• architectonic: Aristotle’s word for a “master” art whose purpose the arts beneath it serve, as riding is served by bridle-making.

• frame: the purpose, limits and measure of success within which a piece of work is done.

• Goodhart’s law: “When a measure becomes a target, it ceases to be a good measure” (Strathern’s wording).

• governor: a person who watches a process already run by a machine and knows when to question it (Chapter 5).

• operator: a person who does the work as defined, or starts a machine’s process (Chapter 5), without asking whether it should be done.

• prayojana: purpose; what a piece of work or a treatise is for.

Sources

1. Aristotle, Nicomachean Ethics I.1, 1094a9–16, trans. W. D. Ross (Oxford, 1925; Internet Classics Archive). The Greek word Ross renders “master arts” is architektonikai (1094a14); politics is called “most architectonic” at 1094a27.

2. Aristotle, Metaphysics I.1, 981a30–b5, trans. W. D. Ross (Oxford, 1908; Internet Classics Archive).

3. OpenAI, “Introducing ChatGPT agent,” 17 July 2025: “handling complex tasks from start to finish” on its own virtual computer.

4. Gartner, “Gartner Identifies the Top 10 Strategic Technology Trends for 2025,” press release, Orlando, 21 October 2024: “by 2028, at least 15% of day-to-day work decisions will be made autonomously through agentic AI, up from 0% in 2024”.

5. Peter F. Drucker, “Managing for Business Effectiveness,” Harvard Business Review, May 1963; wording confirmed by Quote Investigator, 9 February 2017.

6. Donella H. Meadows, Leverage Points: Places to Intervene in a System (Hartland, VT: The Sustainability Institute, 1999); the Donella Meadows Project edition.

7. Marilyn Strathern, “‘Improving ratings’: audit in the British University system,” European Review 5, no. 3 (1997): 305–321, at 308. Goodhart’s original observation: C. A. E. Goodhart, “Problems of Monetary Management: The U.K. Experience” (1975).

8. Sadānanda, Vedāntasāra 1.5 (about 1500; some date it to the early sixteenth century): tatra anubandho nāma adhikāriviṣayasambandhaprayojanāni; see Swami Nikhilananda, trans., Vedāntasāra of Sadānanda (Mayavati: Advaita Ashrama, 1931). The glosses in the text are mine.

9. Kumārila Bhaṭṭa, Ślokavārttika 1.12 (the opening verses), text as quoted by Balabhadra; identified as Ślokavārttika 1.12 by Madhav Deshpande and Dragomir Dimitrov on the INDOLOGY list, May 2015. The last word follows Martin Gansten’s transcription of Balabhadra’s quotation (INDOLOGY, 20 May 2015); the printed edition itself not seen. The English is my rendering.

Named people in this chapter are composites, and their conversations and circumstances are constructed; the studies, laws and figures cited are real and sourced.

Revised from the print edition (2026)

• What the print edition claimed: that the architect needs “a felt sense of purpose” that AI “does not have”; that machines “compute; they do not care”, and that “the difference is not capability. It is consciousness”; that “the architect mindset is an expression of consciousness” and “It is also why AI cannot architect”; that the architect’s output “cannot be replicated”; that Klarna’s direction “is not reversible”; that Rajan heard what “only… a conscious human being” could hear in his buyer’s silence; that his Hamburg buyer, “Frau Winkelmann”, had been waiting for him to ask and said “This is what I have been wanting from you”; that his floor supervisor was “Murugan”, whose grief the closing then called “Senthil’s”; that Priya took the 500D bus; that she invented a “monthly decision log” after November; that the Module 7 failure was hers alone; and that Priya’s willingness to sit with not knowing “is what makes her irreplaceable”.

• What I now hold: the operator runs the work, the governor watches it (Chapter 5), and the architect asks what it is for. These are kinds of attention, not ranks, and good work often needs all three at once; Aristotle’s hierarchy of persons is refused, and the refusal travels with the word. The architect’s question is open to anyone; the seat where the purpose of other people’s work is decided is scarce. The ability to frame a problem does not, by itself, give the authority to decide which problem should be solved. As a skill, framing carries a date: machines plan, propose frames and question a brief when asked. What does not carry a date is who has the right to decide what the work is for (with stakes and say kept apart), who answers for the frame, and whether the people asked to frame can see the decision that made their work, in time to refuse it. A frame can fail in three ways: the task is no longer needed, the goal leaves out what matters, or the measure stops standing for the outcome (Goodhart). Changing the goal is stronger than changing the numbers (Meadows), but not always wiser. A frame is revisable, and a clear frame is not necessarily a good one. Frame before you work: for whom, about what, how connected, and for what (the anubandha, extended and marked). Each link in a chain answers for its part of a wasted frame. Whether a purpose is worth its cost is a question of right, not a dated limit, and letting a machine choose a purpose is itself a choice someone answers for. A clear arrangement is not yet a fair one, and a question about purpose must not count against the one who asks. A machine with a stake of its own would reopen the argument.

• What changed my mind: the Prologue’s rule against defining people by what machines cannot yet do; agents that plan and break goals into steps; Klarna’s reversal; Aristotle’s own hierarchy, read to the end of the passage; the Indian scholars’ four preliminaries; Goodhart and Strathern; Meadows’s places to intervene; Chapter 7’s account of positions that can be emptied.

• What survives: Anand’s apology and the wrong building; “Not the method. The target.”; the three questions; Muthu’s line, “The machine is not wrong about the numbers, sir”; Rajan’s question to his buyer; Luísa’s measures and her standing to say they were wrong; the four framing questions; Priya’s Monday card.

Next in the series: Chapter 10: Building Deep Human Capital.