NP Nithin Palal

Chapter 6: What AI Will Replace

Part II: The Machine at the Gate

· 21 min read · Discuss · Plain text

AI-Proof: What Machines Will Never Be — Volume I: Awakening, Part II: The Machine at the Gate, Chapter 6: What AI Will Replace

The AI read the chest CT in four seconds. Dr. Sanjay Rao had spent fourteen minutes on the same scan that morning and missed what it found.

A six-millimetre nodule in the right upper lobe, partially obscured by a rib. The kind of finding that, in a busy shift — hour nine, three cases deep into a backlog, the canteen coffee cold on the desk beside the monitors — a good radiologist might overlook. Not a bad radiologist. A good one. Having a bad moment. The AI does not have moments. It processes. It found the nodule the way a calculator finds a sum: without effort, without fatigue, without the distraction of a daughter’s school fees or the ache in a lower back from eight hours in the same chair.

Sanjay did not take this well.

He checked the finding. The nodule was there. He pulled up his earlier read and saw, with the particular clarity of hindsight, exactly where his eyes had passed over the shadow without registering it. He had been thinking about something else — the argument with his wife that morning, or the referral letter he needed to write, or nothing at all, just the low hum of a mind that had been reading images for twenty-two years and had learned to trust its peripheral vision a little too much. The AI had no peripheral vision. It examined everything with equal attention. This was its advantage and, in a way Sanjay could feel but not yet articulate, its limitation.

But the limitation was not the part anyone wanted to talk about. The hospital’s administration circulated a memo noting that the AI diagnostic system had flagged findings missed by human readers in seven per cent of reviewed cases during its first month. The memo was celebratory. Collaborative language. Augmentative framing. Nobody said the word that everyone heard.

Sanjay is fifty-three. He works at a private hospital in Andheri, Mumbai. He chose radiology twenty-five years ago because it let him help people without sitting with them. Images, not faces. Patterns, not emotions. He is precise, deliberate, more comfortable reading a scan than reading a room. He has built his career on this preference, and it has served him well. The preference is about to become a problem.

This chapter is about what AI is replacing — not in some speculative future, but now, in patterns visible across industries and continents. Three people’s stories thread through what follows. Sanjay in Mumbai. Eleanor Marsh, a junior solicitor in London. Rosario dela Cruz, a copy editor in Manila. Their industries share nothing. What they share is the experience of discovering that the ground they built on was not as solid as they believed, and that what the ground’s removal reveals about them is not always what they hoped.

6.1 What “Replace” Actually Means

The word suggests a clean swap. A human walks out. A machine rolls in. The work continues. Almost nothing works like that.

What happens to Sanjay is not substitution. The hospital has not fired any radiologists. It does not plan to. What it plans to do is not replace the three who will retire in the next few years. The positions will simply not be filled. The machine and the remaining radiologists will absorb the work. This is replacement by pre-emption — the quietest form, the one that generates no headlines, because no one loses a job. The job is never offered. The young radiologist who spent a decade training for it discovers, when she enters the market, that the seat she was promised no longer exists.

What happened to Priya’s team is a different form: compression. The team of twelve becomes a team of three. The work remains. The jobs do not. The distinction matters — the work being done is essentially the same, but the human labour required to do it has been compressed by a factor that would have seemed absurd five years ago. Compression is, by most measures, the dominant form of displacement in 2026. It appears in budgets, in headcount reviews, in the reorganisation memos that use the word “efficiency” the way a surgeon uses the word “procedure.”

The third form is redefinition. The job persists, but its content changes so fundamentally that the person performing it is doing different work under the same title. Sanjay is still called a radiologist. What he does each day no longer resembles what he spent two decades learning to do.

The fourth is direct substitution — the dramatic version, the one that makes headlines. A human task is fully replaced by a machine. It is the most visible form and, in aggregate, the least consequential. The other three are doing the heavier work.

All four are happening simultaneously. The public conversation fixates on substitution because it makes for powerful stories. But compression and pre-emption are reshaping the economics of cognitive work far more profoundly than any individual instance of a human being replaced by a machine. The invisible replacements — the positions that were never created, the intake that quietly shrank, the department that restructured without a press release — are where the real numbers live.

6.2 Repetitive Cognitive Work

Eleanor Marsh takes the Central line to Holborn five mornings a week and walks seven minutes to a building that smells, in the stairwell, of old carpet and photocopier toner. She is a junior solicitor at a firm whose name she mentions at parties with a particular kind of careful casualness. She has been there two years. She reviews contracts.

Not interprets them. Reviews them. She reads commercial agreements page by page, checks that the clauses match the firm’s templates, flags deviations, and compiles her findings into a memo that a senior partner will scan for thirty seconds before signing off. The work requires literacy, attention, and a working knowledge of contract law. It is also — and this is the part Eleanor has never said aloud, because saying it would unravel something she is not ready to examine — repetitive. The contracts vary. The variations are finite. After two years, she can feel a deviation before she sees it, the way a proofreader feels a typo in the peripheral vision.

An AI system can perform this review in minutes. Not with Eleanor’s contextual feel. But with sufficient accuracy for the purpose the review serves, which is a first pass. A filter. Eleanor was always the filter, not the judge.

Here is the thing about Eleanor that this chapter needs to say, because she will not say it: the contract review was a hiding place.

She chose it. Not consciously — no one applies to a law firm saying they want the most repetitive work available. But she gravitated toward it, volunteered for it, became known for it. Because the review had clear boundaries. You could tell, at the end of every day, whether you had done the job. The alternative — advocacy, negotiation, the work of sitting across from a client whose business is on the line and having to say something that matters in real time — has no such boundaries. It is open-ended, unscripted, and terrifying in a way that Eleanor has managed to avoid for the entire first phase of her career.

The AI has taken the hiding place. What it has left behind is a junior lawyer who does not yet know whether she is any good at the part that remains.

This is the pattern across repetitive cognitive work. The paralegal who summarises depositions. The analyst who compiles quarterly reports. The compliance officer who checks filings against regulatory requirements. Each of these roles demands genuine expertise. Each follows a logic structured enough that, once the machine can process language with competence, the human advantage narrows to the judgement calls — and judgement calls, by definition, constitute a smaller portion of the total work. What the removal exposes is different for every person. For some, it reveals substantial capacity that was obscured by the routine. For others, the routine was the capacity. The honest answer requires a kind of self-examination that no performance review has ever asked for, and no machine can perform on your behalf.

6.3 Predictable Creativity

The sentence needed one more beat. Rosario dela Cruz could feel it — the way a drummer feels a bar that resolves too soon, the slight wrongness of a phrase that does everything it is supposed to do and does nothing more. She deleted the last three words, wrote four new ones, deleted two of those, and arrived at the version that breathed. The whole adjustment took nine seconds. It was, by any external measure, trivial. No client would notice the difference. The conversion rate would not move. But Rosario noticed, and the noticing was the thing she had spent seven years training.

She writes marketing copy at a content agency in Makati, Manila. Product descriptions. Email campaigns. Social media posts. Landing page text for e-commerce companies across Southeast Asia. The work is skilled. She has an ear for rhythm, a sensitivity to the cultural registers of Filipino, Bahasa, and Singaporean English audiences, and the discipline to cut a sentence in half when half is enough. She is good at this.

She would also admit, if you asked carefully and she trusted you not to judge, that the work follows patterns. A product description for a skincare brand has a structure — benefit, ingredient, social proof, call to action — that varies in execution but not in architecture. An email subject line that converts follows principles so thoroughly documented they could be written as rules. The creativity is real. The range of possible outcomes is not infinite.

AI-generated marketing copy in 2026 is not great. It misses what Rosario catches in those nine-second adjustments — the beat, the breath, the difference between text that functions and text that has someone behind it. But it is adequate. And for the price point at which most of her clients operate — small to mid-size e-commerce companies in competitive markets with thin margins — adequate at speed beats good at cost. The economics are not ambiguous. They are merciless.

Her agency has lost a significant share of its revenue in the past year. Clients who once sent ten briefs a month now send three, handling the rest with AI tools and occasional freelance cleanup. Rosario has been offered a new role: curating and editing AI output instead of writing original copy.

She has not accepted.

Not because the economics are unclear. Because the role is a demotion dressed in new language, and Rosario has too much pride to pretend otherwise. She is a writer. The new job would make her an inspector of machine-generated text, checking for errors the machine usually does not make, smoothing edges the client has already decided are smooth enough. Seven years of developing her ear, her instinct for the phrase that makes a reader pause — and the market’s answer is quality assurance.

A different person might take the curation role and find something valuable in it. Might discover that taste becomes more important, not less, when everything around it is machine-generated. That in a world flooded with adequate text, the person who can feel the difference between adequate and alive holds something genuinely scarce. Rosario cannot see this yet. What she sees is the gap between what she built and what she has been offered, and the gap looks like a cliff. The pride that prevents her from taking the role is the same pride that made her good at the work in the first place. It is her best quality and her biggest obstacle, and the chapter will not pretend to resolve what she herself has not.

Predictable creativity is the category of work where the output, while requiring skill, follows patterns established enough that AI can approximate the result. It includes most copywriting, most template-based design, most routine illustration. It does not include the work that requires genuine originality — the campaign concept no one has seen before, the visual identity that makes people feel something new. But the boundary between predictable creativity and genuine originality is drawn by the person doing the work, not by their job title. The AI does not distinguish between Rosario’s nine-second adjustments and the generic text it generates. The market, eventually, will.

6.4 The Death of the Simple App

Tomas Reyes built the grocery price tracker in eight months. He and three developers in a co-working space in Cebu, fuelled by SanMig and the conviction that budget-conscious shoppers in Metro Manila deserved better data on where to buy their rice and cooking oil. The app was simple — a clean interface, reliable price comparisons across wet markets and supermarkets, a loyal user base that had grown to forty thousand. Tomas’s mother used it. His landlord used it. It was a real business. He had quit his BPO job to build it.

In January, two competitors launched. Not copies — alternatives. One was built by a solo developer in Makati using AI coding tools, in three weeks. Three weeks. Tomas stared at the app store listing and counted the features. The interface was cleaner than his. The coverage was broader. The developer had not quit a job, had not spent eight months debugging price-scraping algorithms at two in the morning, had not maxed out a credit card. She had typed prompts into a code-generation tool and shipped. Tomas’s advantage — the effort, the expertise, the eight months — was no longer an advantage when the cost of building had collapsed.

The simple app economy of the 2010s was built on a barrier to entry: building software required a team, and the team’s cost protected the product once built. That barrier has fallen. When a competent developer working with AI can build in a weekend what once took months, the economics of the simple app stop working. The pattern looks different in different markets — the VC-funded startup ecosystems of San Francisco and London experience it differently from the mobile-first markets of Lagos and Jakarta, where super-apps and platform economies create different competitive dynamics. But the underlying force is the same everywhere: the cost of building has fallen faster than the cost of acquiring users.

The opportunity lives in what the collapse reveals. When building is cheap, the bottleneck shifts from execution to insight — to understanding, with genuine depth, what a specific community needs and why existing solutions fail them. The developer in Nairobi who builds a livestock-health monitoring tool is not competing with generic AI. She is competing on knowledge of pastoralist communities that no language model possesses. The advantage is contextual, relational, human. How to build on this foundation is a question for later. Here, the point is narrower: the death of the simple app is also the birth of the app that requires a conscious human behind it.

6.5 Thin AI Wrappers and Why They Die

Adaeze Okonkwo raised four hundred thousand dollars in Lagos in late 2023 for an AI writing assistant tailored to African English. Igbo idioms. Yoruba business formality. The register shifts that a Silicon Valley language model flattened into mid-Atlantic blandness. She had the insight, the market, the timing. What she did not have was a moat. The product was a user interface on top of a large language model’s API, with prompt engineering that shaped the output. Within six months, three competitors launched identical products. Two of them used the same API. One had copied her prompt structure almost verbatim.

A thin wrapper is a product whose core value is access to a capability that the underlying model already provides. The test is simple: if a well-resourced competitor could replicate your product in a fortnight using the same API, you have a wrapper, not a business. When the model improves — when the provider adds the feature natively — the wrapper’s reason for existing evaporates.

Many have. The lesson is not that AI startups are doomed. It is that the source of durable value has shifted to things the API cannot provide on its own: proprietary data that improves with use, deep workflow integration, trust relationships with specific industries, the kind of domain expertise that requires a human being who has lived inside a problem. That architecture deserves examination in detail. Here, the point is diagnostic: the thin wrapper dies because it confuses access to intelligence with possession of something valuable. Intelligence is a commodity input. What you build with it is not.

6.6 Organisational Middle Layers

Eleanor’s firm restructured in March. She was not on the list — she was too junior to be worth cutting, and too junior to understand, at first, what the list meant. But she watched. Three senior associates and two mid-level partners were moved to “advisory roles” that everyone understood to be the anteroom to departure. They were the people who had occupied the space between the partners who brought in business and the juniors who processed it. They gathered information, formatted it, presented it upward, translated strategy downward. They were the middle layer.

AI compresses this layer. Not by replacing managers — the best of them do far more than translate information. They read a room. They know which associate is struggling before the billable hours show it. They hold institutional memory that no dashboard captures. But the translation function — gathering, formatting, summarising, presenting — is precisely the kind of cognitive work that AI handles competently. The question facing every person in this layer is what fraction of their current work is translation and what fraction is something only a human can do. The honest answer determines whether their role survives.

Eleanor watched the restructuring and felt two things she did not expect to feel simultaneously: sympathy and relief. Sympathy for the people being moved out. Relief that the ceiling above her — the years of grinding through the middle ranks that had always stood between her and the work she might be capable of — had been lowered. The pipeline that used to require a decade of dues-paying before you sat in the room with clients was being compressed along with everything else. The question was whether Eleanor, who had spent two years hiding in the documents, was ready to be in the room.

That transformation deserves exploration in depth — the middle manager’s pivot from controller to cultivator. Here, the point is structural: the layer is thinning, and the people who survive in it will be the ones whose value was never primarily informational.

6.7 The Messy Middle: Why the Next Five Years Will Be Chaos

Previous technology transitions followed a pattern. The automobile displaced the horse-drawn carriage over roughly fifty years. The shift from analogue to digital communication — depending on where you draw the lines — took three decades. Each was disruptive, but the disruption unfolded slowly enough for institutions, labour markets, and individual lives to adapt. The loom displaced weavers, but their grandchildren worked in textile factories. There was suffering. There was also time.

Whether the AI transition will follow this pattern is a genuinely open question, and anyone who tells you they know the answer is selling something.

The case for exceptionalism: AI capabilities improve in months. Enterprise adoption cycles run in years. Regulatory frameworks lag further still. The gap between what the technology can do and what institutions are prepared to absorb is widening, not narrowing, and that gap is where the chaos lives.

The case against: enterprise software transitions have always been slower than they appeared from the outside. Cloud computing took more than fifteen years to move from early adoption to genuine infrastructure. Mobile took nearly a decade. The factors that slow adoption — organisational inertia, regulatory caution, the sheer difficulty of changing how millions of people work — are not eliminated by the speed of the underlying technology.

The honest position is that nobody knows. The transition could unfold over five years or fifteen. What is clear is that the interval between now and the new equilibrium — whatever it looks like, whenever it arrives — will be characterised by simultaneous overshoot and undershoot. Industries that adopt too fast and suffer the consequences alongside industries that adopt too slowly and lose ground. Individuals who retrain for roles that no longer exist by the time they qualify, alongside individuals who cling to roles transforming beneath them.

Sanjay is in the overshoot category. His hospital adopted fast. The AI found the nodule he missed, and the administration drew a conclusion from that data point — a reasonable conclusion, supported by the evidence, but a conclusion drawn without understanding what it costs to turn a radiologist who chose the dark room into a clinician who must sit with patients in the light. Rosario is in the undershoot category. Her agency was slow to adapt, and by the time it began offering AI-assisted services, three competitors had already taken the clients.

In the messy middle, the most valuable quality is not prediction. It is the ability to respond to conditions that change faster than any plan can anticipate. That ability has a proper name, and it will earn it. For now, notice that it requires something a machine does not possess: the awareness that the ground is moving, and the willingness to move with it before the direction is certain.

6.8 Consumer Work vs. Enterprise Work

Not all cognitive work faces the same timeline. The distinction worth drawing is between work that faces consumers and work that faces institutions.

Consumer-facing work operates where the tolerance for imperfection is high and the cost sensitivity is acute. The social media post. The product description. The first-pass customer support response. If the AI’s output is eighty per cent as good at five per cent of the cost, the economics decide. Rosario’s clients did not leave because the AI was better than her. They left because it was cheap enough.

Enterprise-facing work operates under different constraints. When Sanjay’s hospital uses AI for diagnostic imaging, the tolerance for error is appropriately low. When a pharmaceutical company uses AI in drug trial analysis, a mistake is not an unhappy customer but patient harm, regulatory action, legal liability. When a bank uses AI for credit decisions, each error carries costs that dwarf the efficiency gains. The enterprise timeline is measured in years, not months.

But “gradual” is not “never.” The enterprise space is not immune to the same forces. It is delayed. As AI systems grow more reliable, as regulatory frameworks adapt, as organisations accumulate experience and confidence, the tolerance for machine autonomy in high-stakes environments will increase. The time this buys is real. It is best spent preparing, not resting.

6.9 The Uncomfortable List: Jobs You Thought Were Safe

You already know what is on this list. You have known since you started reading this chapter, and possibly since before you opened the book. You may have picked it up because you suspected your name was somewhere on it.

Radiologists. Not disappearing — contracting. Fewer positions, redefined roles, the detection work increasingly handled by machines and the human contribution narrowing to the judgement calls, the patient communication, the clinical context that no imaging algorithm accesses. Sanjay is not losing his job. He is losing the job he chose, and what remains is one he did not.

Lawyers — not the advocates and counsellors, but the associates whose work lives in documents. Research, review, analysis, drafting. The senior partners who bring judgement and relationships are not exposed. The juniors whose value was processing volume are competing with systems that process faster, cheaper, and without the need for sleep. Eleanor is one of them.

Accountants in compliance-oriented roles. Financial analysts producing routine research. Journalists covering predictable beats — earnings summaries, sports results, municipal proceedings — work that automated systems have handled at some news organisations for years, a fact the industry has been reluctant to discuss publicly. Software engineers writing routine code. Copy editors, translators, paralegals, technical writers, entry-level consultants. The list extends into work that requires genuine training and genuine skill.

Teachers, though this one carries a complexity that no list can capture. The teacher whose primary function is content delivery is being supplemented by adaptive learning systems. The teacher whose function is human development — recognising when a student is struggling, modelling how a thoughtful person engages with difficulty, inspiring the curiosity that no algorithm can generate — is irreplaceable. Most teaching jobs are both. The economics push toward automating the first, which threatens to erode the conditions that support the second.

If you felt something while reading that list — not the detached interest of a reader studying a phenomenon, but the specific prickle of recognition that comes when the phenomenon is your life — then you are paying the kind of attention that matters. The feeling is not panic. It is your consciousness registering something your professional identity may not yet be ready to absorb: that the cognitive work you do, however skilled, however respected, may be closer to the machine’s reach than you believed.

Stay with the feeling for a moment before doing anything with it.

The question it raises is not whether AI will affect your work. It is what proportion of your work is pattern — reproducible, structured, amenable to computation — and what proportion is the thing that requires you. Not your knowledge. Not your speed. Not your credentials. You. The person behind the output, the consciousness that decides what matters, the self that takes responsibility for the decision and lives with the consequences. The answer is different for every reader. The honesty required to arrive at it is itself a quality no machine possesses.

There is a counterargument worth taking seriously, because it is not wrong. Some labour economists argue that AI augmentation will increase demand for the professionals on this list rather than reduce it. The logic: cheaper imaging leads to more imaging, which leads to more radiologists needed for supervision and interpretation, not fewer. Faster legal research expands access to legal services, creating more work. ATMs did not reduce the number of bank tellers; they lowered the cost of running branches, which led to more branches. The pattern has precedent.

Whether it holds for AI is unknown. The honest answer is that the complementarity argument may be right about totals and wrong about individuals. Even if the number of radiologists increases, the work they do will be unrecognisable to Sanjay. Even if legal services expand, the junior associate’s role will bear no resemblance to what Eleanor was trained for. The profession survives. The professional must become someone new.


Sanjay is still in Andheri. Four months since the AI arrived. He reads fewer images now. He reviews more. The machine flags; he contextualises, integrates, decides. He has been asked — encouraged, then firmly directed — to spend more time in consultation rooms. Sitting with patients. Explaining findings. Answering the questions that a frightened person asks when the images show something that should not be there.

He is bad at it.

Not incompetent. Not unkind. But the consultation room demands everything the dark room did not: presence, emotional steadiness, the ability to hold another person’s fear without drowning in it or retreating from it. These are qualities Sanjay has never cultivated, because the entire architecture of his career was built to avoid needing them. He chose the speciality that let him help at a distance. The distance has been closed, not by his choice but by a machine that took the work he was comfortable with and left the work he was not.

He does not know yet whether he can learn what the new role requires. He is fifty-three. The skills the consultation room demands are not technical. They are, in the language of this book, qualities of consciousness — presence, empathy, the courage to sit with uncertainty that is not diagnostic but human. He has spent twenty-two years developing a different set of strengths. Whether the old strengths translate, whether they can be redirected, whether a man who chose the dark room can learn to work in the light — these are open questions, and the openness is the most honest thing this chapter can offer him.

Eleanor is in London. Her firm’s restructuring is complete. Half the junior team has been reassigned. She has been moved into a client-facing role that the partners describe as an opportunity and that Eleanor experiences as exposure. Last week, she sat in on a meeting with a commercial property client and was asked, by the partner leading the matter, for her view on a contractual ambiguity. She gave an answer. It was correct. Her voice shook.

She told a friend afterwards that the restructuring was exciting. She was lying. What she feels is vertigo. The safety of the repetitive work was also its ceiling, and removing the ceiling has not liberated her. It has revealed the height, and the height is terrifying. She is not yet sure whether the terror is the kind that fades with practice or the kind that signals she is in the wrong place. She suspects the former. She is not certain.

Rosario is in Makati. She has not accepted the curation role. She has not declined it. She goes to work each morning and does the writing that remains — the assignments too complex or too culturally specific for the AI — and she does them well. In the afternoons, she reads through the AI-generated copy her colleagues are editing, and she sees every place where it is flat and every place where flatness is all it will ever be, and she does not know whether that perception — the ability to feel the difference between text that functions and text that lives — is worth anything to anyone but her.

Three cities. Three professions. Three people in the space between what their work was and what it is becoming, and none of them has arrived at the other side.

The next chapter asks the question this one has been building toward: in all this displacement, in all this compression and redefinition and quiet erasure of positions that will never be filled — what, exactly, cannot be touched?

↑ Back to top

Discuss

Loading the discussion…

More in Artificial Intelligence
  •  ·  26 min read

    Epilogue

    Late in the Gītā, the one who knows most says: ponder it thoroughly, then act as you like. A machine can now set out the reasons on every side. Saying “I answer for this” then is not pride. It is a reply, for your part, to those owed one.

  •  ·  37 min read

    Chapter 40: What This Series Cannot Tell You

    The print edition’s last chapter reduced the book to one sentence, and rested human worth on what machines cannot do. I withdraw both. A book can date its claims and correct them in public. It cannot sit in the room where you decide.

  •  ·  6 min read

    Part IX: The Unfinished Chapter

    The print edition’s last Part said that consciousness never concludes. This one is a single chapter that turns the series on itself: what a series can stand behind, what it owes those who read the print edition, and what it leaves open.