I took the four addresses in which an Indian regulator or a directors’ institute told boards that artificial intelligence is theirs to answer for, between 11 and 22 August. I opened each one at its primary source rather than at the coverage. I read them in full. Then I searched each for the word “jobs”, and for the word “headcount”.

Zero. Zero. Zero. Zero.

That is not an inference from tone. It is a property of four public documents that anyone can check in an afternoon, and I would like somebody to.

The four

On 11 August the Governor of the Reserve Bank, Sanjay Malhotra, told the FIBAC 2026 conference in Mumbai that banks must “Establish board-approved AI governance policies, with clear accountability for outcomes, not merely for technology procurement”. He asked for a complete inventory of every AI system in use, “including those embedded in vendor products”, and for outsourcing agreements that carry “the right to audit, the right to explanation, and a credible exit plan”. Meaningful human oversight, he said, means “the ability to explain, to intervene, and, where necessary, to override”.

On 19 August his Deputy Governor, Shirish Chandra Murmu, put the same duty as a test of ownership, in two consecutive sentences I have not stopped thinking about. “A machine may reach the decision; a person must own it.” And then: if a bank cannot say who that person is, “it has not deployed a model”, it has “delegated its accountability”.

The same day, the Chairman of SEBI, Tuhin Kanta Pandey, told FICCI’s capital markets conference that guidelines for responsible use of AI and machine learning would “shortly” be issued, requiring “Kill Switch” and “Humans in the Loop” controls, and that “Every SEBI-regulated entity remains fully responsible for any AI or machine-learning tool it uses, whether developed in-house or procured from a third party”.

Three days later, standing in front of an audience of company directors at the Institute of Directors Annual Directors’ Conclave, the same Chairman said this: “Technology risk is no longer merely an IT risk. Technology risk is business risk.” And then the sentence that gave this essay its title. “And therefore, it is board risk.”

He is right. That is the correct conclusion, arrived at in the correct order, and it is a better sentence than most regulators manage. I want to be clear that I am not writing against any of these four documents. I am writing about the shape of the hole in them.

What the duty was defined to contain

Assigning accountability is the easier half of governing a technology, and it is the half now being done at speed. Deciding what the accountable person is accountable for is the harder half. It is being done at the same time, quietly, by the choice of which instrument carries the duty.

Look at what is inside the scope as written in August. Model inventories. Vendor-embedded systems. Audit rights. Exit plans. Explainability. Human oversight, intervention and override. A kill switch. A human in the loop. Full responsibility for a bought tool. Data privacy, security and integrity of outputs.

Every one of those is about the machine.

Now look at what is not there. How many people work at the firm afterwards. Which reductions were attributable to automation and which were not. A redeployment rate. Advance notice to the workforce. Any disclosure at all of AI’s effect on employment.

Nobody had to decide to leave the workforce out. It falls out of the choice of instrument: a model-risk framework is not the natural home of a jobs policy, and the Reserve Bank’s remit is financial stability rather than labour. That is a fair defence of each document taken alone. It is not a defence of the aggregate, because instruments harden. What gets written down in 2026 is what a compliance function will be pointing at in 2029, and the scope being fixed right now has a hole in it in the shape of the payroll.

I should also record what the Indian texts do say about people, because it is not nothing and it would be dishonest to imply otherwise. Malhotra tells banks they “will need to invest in technology: IT infrastructure, talent, skilling and reskilling”, and that document processing, reconciliation and internal audit sampling are “ripe for AI-assisted automation, freeing skilled staff for judgment-intensive work”. Murmu says institutions must “reskill, and to grow their capability”. Both are speeches about what a workforce must become. Neither is a speech about how large it will be.

One check I ran on myself: is this an Indian omission?

It is not, and the evidence for that is the most uncomfortable thing I found.

ISS STOXX Governance runs an annual benchmark policy survey that shapes how the largest proxy adviser in the world will tell institutional investors to vote. The 2026 instrument, which sets 2027 policy, opened on 21 July and closed at 5pm on 14 August. I read it.

Its six section headings are respondent information, board elections, shareholder rights, compensation, audit and auditors, and environmental and social topics. That last one resolves into exactly two sub-topics: climate board accountability, and nature-related risks.

There is no question in it about artificial intelligence. None about human capital, workforce, employment or headcount.

On 26 August, Glass Lewis told clients it will keep its benchmark voting policy for the 2027 season with “limited changes”, integrating only significant regulatory and governance developments from 2026. Neither AI nor human capital appears in the reported content. I should say plainly that there is no public Glass Lewis document to point you at: it was a client message, and the firm’s own news index carries nothing later than 30 April.

The Conference Board’s proxy season review, published 11 August with ESGAUGE, Russell Reynolds Associates and the Rutgers Center for Corporate Law and Governance, shows the two lines moving apart rather than together. AI-related shareholder proposals rose from 18 in 2025 to 24 in 2026, a three-year high. Human capital proposals fell 37% in a year and 60% over two, and their average support dropped from 9% to 6%.

ESGAUGE’s Umesh Chandra Tiwari puts the direction plainly: shareholders are “paying less attention to broad governance frameworks and more attention to AI’s real-world impacts, from data governance to energy use”.

Energy and water are on that list. Employment is not.

And on 14 August, the Division of Corporation Finance at the US Securities and Exchange Commission announced it would “discontinue responding to Rule 14a-8 no-action requests entirely”, effective immediately. That is the referee leaving the pitch in the one process by which a shareholder can force a company to put a question on its own ballot that the company does not want there.

So: the regulators are assigning the duty and defining it without the workforce, and the machinery that exists precisely to put unwelcome questions on board agendas spent the same month asking about climate, nature and energy.

The one place somebody did ask

I want to give the counter-evidence its full weight, because it exists and I nearly missed it.

On 24 February the New York State Comptroller, Thomas DiNapoli, wrote to Meta encouraging increased disclosure of AI-related workforce management practices, including “The number or proportion of jobs eliminated, created, or materially restructured due to AI adoption”, and “governance and oversight structures guiding AI strategy and workforce management”.

That is exactly the question this essay says nobody is asking, asked in writing, by a fiduciary large enough to be answered. In April he wrote publicly that he had approached the largest publicly traded companies in the state pension fund’s portfolio. Note the instrument, though: he encourages disclosure. He does not demand it, and describing it as a demand would overstate what a comptroller can do.

Then look at what happened when the same question reached an actual ballot. The only 2026 shareholder vote I can find anywhere that put AI’s effect on a workforce directly to owners was at Walmart’s annual meeting on 4 June. On the company’s own figures it received approximately 4.95% of the shares voted. On the same ballot, a workplace health and safety governance proposal drew 6.63%.

One fiduciary asking, one vote at five per cent, and no regulator in India, the Gulf or Africa requiring anything of the kind.

What binding actually looks like

There is exactly one place across the three regions I cover where this is already a duty rather than a speech, and it is worth studying for that reason alone.

South Africa’s King V Code applies to financial years commencing on or after 1 January 2026. Since 16 January 2026 the simplified JSE Listings Requirements have required applicant issuers to adopt and apply the King Code, which is the step that turns a voluntary code into a listing condition. And the Code requires that the processes, data, models, algorithms and tools used in automated technologies be subject to “human oversight and override mechanisms” proportionate to risk.

So what does a bound board actually disclose? On 13 August, Standard Bank Group told shareholders that as at 30 June, 72% of employees were active users of generative AI tools across 87 approved use cases, and that employees were up 1% to 51,200.

Adoption at scale, disclosed to the market, with headcount rising. I am not going to overclaim that: the company discloses no denominator for the 72%, so I will not multiply it out, and no branch count appears in the announcement at all.

Beside it, and this is the case that disciplines the whole essay: African Bank began a section 189A process in July over roughly 1,200 roles and up to 90 branch closures after a R624 million half-year loss. No source cites AI or automation as a driver. COSATU’s objection is to an acquisition spree, not to an algorithm. Cuts without AI at one bank, AI without cuts at another. Anybody who wants a tidy story about automation and employment should sit with those two for a while.

What consultation is worth when the other side owns the machine

Last week I wrote about Korea, where a union had put AI employment guarantees into a live wage round, fifth on a nine-point list, and struck for a full day over it. I said at the time that this had to happen before it could ever become normal.

It has now finished, and the ending is instructive.

Hyundai and its union reached a tentative agreement on 25 August, in the seventeenth round of talks. On the AI job-security demand, the agreement text Reuters saw records that the two sides “agreed to discuss matters related to employment when it comes to rolling out new businesses”. The company’s own release frames the settlement around a joint recognition that introducing physical AI and robotics is an essential condition for corporate survival, with a commitment to share progress transparently and respond together.

Read that again. The union went in asking for a guarantee. It came out with a consultation clause and a joint endorsement of the technology it was worried about. The nearest thing to a concrete employment outcome is hiring rather than protection: 200 technical staff in the second half of 2027 and 300 in 2028.

That is what happens when the duty is defined by the party that owns the machine. It is also the best available answer to anyone who thinks that “advance notice and mandatory consultation”, the exact phrase in the Korean bill and the exact posture of the Indian speeches, is the same thing as protection. It is not. It is a meeting.

The agreement goes to a membership vote on 31 August, which is after I finished this. I am not going to tell you how it went, because I do not know.

The case against me

The strongest objection is that I am measuring the wrong documents. A speech on model risk in banking is not where you would look for employment policy, any more than you would fault a fire code for saying nothing about wages. India has a labour ministry, and if it has views on AI and employment they will not be found in a Reserve Bank keynote. On that reading I have counted a word’s absence from four documents that were never going to contain it, and dressed a category error as a finding.

I think that objection is right about each document and wrong about the aggregate, for one reason. The scope of a new duty is not fixed by any single instrument. It is fixed by the whole set of them, arriving at once, in a year when nobody has yet written the alternative down. If the labour ministry were writing the other half in parallel I would have a much weaker essay. I looked. It is not.

The second objection is fairer and I cannot fully answer it: a board with an AI policy might cut fewer people, or more, and I have no evidence either way. Nor does anybody, and that is the actual finding underneath this one. The reason no evidence exists is that nothing requires anyone to produce it.

A prediction, dated and checkable

By 30 June 2027, no financial or securities regulator in India, the Gulf or Africa will have issued an AI governance instrument that requires a regulated entity to disclose AI’s effect on its own headcount.

Not the Reserve Bank, not SEBI, not the UAE or Saudi capital market authorities, not the JSE. Frameworks will be issued. SEBI’s own, promised “shortly” on 19 August, had not been issued when I checked the circulars listing on 30 August, and its board did not meet in August at all. When it lands I expect a tiered, well-drafted, genuinely useful instrument about models, and silence about people.

If I am wrong, I will say so here, and it will most likely be South Africa that proves it, because King V is the only code in the three regions that already binds.

Where this leaves you

The practical version is smaller than the policy and lands on your own desk.

Your next board pack will, at some point in the next year, carry an AI item. It will be about models: what you run, who owns it, whether you can switch it off. That item will be well prepared, because the regulator has told you what it wants.

The question nobody has told you to prepare is the one Pandey actually put to directors: “Which important decisions depend on algorithms or automated systems or AI?” Asked of a credit book it is routine. Asked of a hiring funnel, a performance ranking or a redundancy pool, it has in most organisations never been asked at all.

And underneath it is the piece of information every instrument in this essay would run on. Somewhere in your organisation, over the last twelve months, roles were reduced. Some of those were caused by automation and some were not, and almost certainly nobody wrote down which at the time, so the classification cannot now be reconstructed.

That is not a compliance gap yet. It is a compliance gap in draft, and this month four institutions handed your board the file it will eventually sit in.

Read the full issue, with the ledger and the source register: reaimagine.work/issues/023

Korea passes the world’s first robot tax.

Five words. Three of them wrong, one of them misleading, and one that survived. I want to take them apart in order, because the wreckage is more useful than the sentence was, and because the thing left at the end is the most interesting story I have looked at this year.

I send every claim out to be attacked before I write. The instruction is to try to knock each one down rather than agree with it. This time the attack landed on almost everything.

“First”

Dead on contact.

The New York State Assembly has a bill, A8179 from 2023, reintroduced as A3719 in January 2025, that taxes businesses per employee displaced by technology. That is not an analogue of the Korean mechanism. It is the same mechanism.

There is more. South Korea itself trimmed its automation tax credits back in 2017, by a different route. The European Parliament considered and rejected a robot tax in the same year. There is a United States Senate bill proposing a tax on automation from 1940, which I mention because it is a useful corrective to anyone who thinks this argument was invented by people with podcasts.

The detail I find hardest to forgive myself for: the Korean coverage makes no “first” claim at all. Four Korean outlets reported the filing on 14 August, and not one of them says first anywhere. The word was mine. I had imported it, from nowhere, into a story that was doing fine without it.

What I can defensibly write is this. It is the first bill of its kind in Korea’s National Assembly, and with a New York state bill as the nearest prior art, among the first anywhere to levy employers per displaced worker. That is a weaker sentence and a true one.

“Passes”

The bill was filed on 14 August. That is all that has happened to it.

No government response. No response from the employers’ federation, which in late July was busy asking the government to scrap 112 regulations it says block AI and robot adoption, so it is not as though the constituency is quiet in general. No word from the presidential office. No committee date. Between the filing and the day I finished checking, six days later, I could not find a single public reaction to it from anyone with the power to move it.

There is a further problem, and it is mine rather than the bill’s. I cannot give you a bill number. The National Assembly’s bill information system was unreachable every time I tried it. I am not going to invent an identifier so that a paragraph looks better sourced than it is, so there is no number anywhere in this piece or in the issue it comes from.

The right word for the bill’s status is untested. Not gathering momentum, not stalled, not controversial. Untested.

“Robot tax”

This is the one that matters, because getting it wrong makes you misread the whole instrument.

Nothing in the bill is charged for owning a machine. The trigger is a fall in employment.

A firm whose adoption of AI or automation substantially cuts its employment above a threshold pays an AI Transition Response Levy into a Basic Society Support Fund. Before it adopts, it owes advance notice and mandatory consultation with its workers. And if it maintains or grows headcount, or genuinely cannot pay, it gets relief.

Read that again with an eye on incentives. The levy is designed to be avoidable by not cutting. A robot tax discourages the machine. This discourages the redundancy, and is indifferent to how much automation you buy so long as your people are still employed. Those are different policies with different effects, and calling the second one by the name of the first is not shorthand, it is a category error.

Then there is the threshold, and the threshold is the whole thing. It is to be fixed by a future presidential decree. That decree does not exist. So no firm in Korea can currently be told whether it would ever pay a won under this bill, and delegated thresholds are precisely where instruments of this kind go to be quietly softened out of existence. If you want a single reason to be sceptical about the levy, it is not the politics. It is that the load-bearing number is a blank.

“Korea”

Which leaves the word that survived, and it turns out to be carrying the story on its own.

On 20 May, at a JPMorgan-hosted investor session in Boston, Hyundai’s vice chair told the market the company plans more than 25,000 Atlas humanoid units across Hyundai and Kia plants, out of a production target of 30,000 a year by 2028.

In the week of 15 July, the Hyundai union ran four-hour daily walkouts at Ulsan and Jeonju, costing roughly 5,000 vehicles and over 200 billion won. Among the demands: management consent before humanoid robots reach the line. As far as I can establish, that is the first strike action anywhere over humanoid robots specifically.

On 14 August, two lawmakers filed the bill.

On Friday 21 August, about 39,000 union members staged the company’s first full-day strike in ten years. Both shifts, eight hours each, the Ulsan, Jeonju and Asan plants idled. Reuters reports that striking workers at Hyundai, at Kia and at their suppliers rallied outside the Seoul headquarters.

A legislature pricing displacement. A union pricing consent. A company pricing the robot. One country, one quarter, one industry, and all three arguments running at once in public through institutions that can actually bind somebody.

I have been looking for that combination for eight months. Everywhere else, one or two of those three are happening and the third is a conference panel.

The demand list, in full, because it matters

I am not going to let my own thesis off the hook here.

The strike is a wage round. The AI demand is fifth of nine, sitting between a bonus claim and a fully fixed salary system. Above it: reinstatement of dismissed workers, retirement age, a 149,600 won base-pay increase, a bonus at 30% of last year’s net profit. Below it: an 800% annual bonus, a 4.5-day week, retirement at 65.

Anybody telling you Korean autoworkers struck over robots this week is overselling it. Anybody telling you the robots were not on the table has not read the list. The honest version is that a major industrial union has put AI job guarantees into a live wage negotiation alongside pay and pensions, which is a thing that has to happen before it can ever become normal, and which is not yet the same as it being the point of the dispute.

Note also which outlet carries what. The Korea Herald prints the demand in the union’s own words. Reuters carries it independently. The Korea Times summary of the same day omits it entirely. If I had read only the third, this essay would not exist.

The number both sides are pricing against

The Korea Development Institute projected in July that AI will eliminate roughly 256,000 Korean jobs a year. Against that, the finance ministry forecasts net employment growth for the entire Korean economy, from all causes, of 150,000 in 2026 and 170,000 in 2027.

That comparison is real and it is also, in the form I first wrote it, wrong twice.

The 256,000 is a projected annual rate at the ten-year horizon, not a figure for now. Printing it without the horizon implies that quarter of a million Koreans a year are currently being displaced, which nobody claims. And the 170,000 is not, as I had it, a KDI figure and not an AI figure. It is the government’s economy-wide net employment forecast for next year, from an entirely different institution measuring an entirely different thing. I had two numbers that looked like a matched pair and were not.

One more, since I am confessing. KDI’s much-quoted rise from 1.4% to 8.1% is not AI “exposure”. It is the share of Korean jobs where automating actually pencils out on cost. Economic viability, not exposure. The distinction is the difference between what a machine could do and what a machine is worth doing, and every serious argument about displacement lives inside that gap.

The machine is not there yet, in two senses

First, geographically. Hyundai’s announced first Atlas deployments are the Georgia plants in the United States, Metaplant in 2028 and Kia Georgia in 2029. Not Korean lines. The union’s consent demand is about Korean lines. So the robot being bargained over in Ulsan has not yet been scheduled to arrive in Ulsan.

Second, and more awkwardly for everyone in this story, financially. Reuters reported from the World Robot Conference in Beijing on 18 August that Guotai Securities puts the break-even price for an industrial humanoid at about 160,000 yuan, including maintenance, to pay back inside two years. MERICS puts typical current pricing at 300,000 to 500,000 yuan. And the robotics analyst Georg Stieler estimates that between half and seven-tenths of this year’s humanoid production will end up in “data factories”, collecting training data rather than doing work for anybody who is paying for the work.

Two to three times the price it needs to be, and most of the output not doing the job at all.

So Korea is legislating, striking and budgeting over a displacement that, on the market’s own numbers, has not happened and cannot yet afford to.

I thought that undercut my story. Having sat with it, I think it is my story.

The case against me

The strongest objection is that I am dressing up a wage dispute and a backbench bill as a policy turn, and that Korea will do what everyone else does, which is nothing. The bill dies in committee, the threshold decree never gets written, the union settles for the bonus and drops item five, the Atlas units go to Georgia and nobody in Ulsan ever tests any of it. On that reading I have built an essay on a filing and a picket line.

That objection is probably right on the odds. I think it is wrong about what is being observed.

I am not forecasting that Korea solves this. I am pointing out that Korea is the only place where the question has been put into instruments that could, in principle, produce an answer: bill text, a bargaining table, a capital plan. Everywhere else the question exists only in commentary, where it cannot bind anyone and cannot fail in public either. A bill that dies in committee has still been drafted, costed and argued over. That is a different state of the world from a country where nobody has written the sentence down.

And the timing is the part I keep coming back to. The cheapest moment to settle who pays for displacement is before the displacement is economic. Once the humanoid costs 160,000 yuan instead of 400,000, the firms with an interest in there being no levy will be very much larger and very much more motivated than they are today.

Korea is having the argument in the window when it is cheap. That window is closing on its own schedule regardless of what any legislature does.

Meanwhile, the instrument we do have is degrading

In the same fortnight, the United States printed its first negative payrolls month of the cycle, minus 23,000, with 103,000 of downward revisions. Stanford’s canaries update widened the young-worker AI employment gap to 19%. And Challenger’s July report had AI leading every stated reason for job cuts for a fifth straight month.

Then Andy Challenger, whose firm compiles the most-quoted AI layoff series in the world, said this out loud: as regulation takes shape, companies will stop saying “AI” in their announcements, which would make tracking the impact of AI on jobs more opaque.

The best instrument we have for measuring AI displacement is telling us, in advance, that it is about to stop working, and why. Not because the phenomenon will recede, but because the disclosure will. Every future headline reading “AI-attributed job cuts fall” will be unreadable without knowing which of those two things happened.

I wrote about a related blind spot in an earlier issue, where the clearest AI-caused job loss I could find was the one no tracker could ever record, because the deciding company was not the employing company. This is the same failure from the other end. There, the signal was never sent. Here, the signal is about to be switched off deliberately, by people with good legal reasons.

And this is precisely what the Korean bill would force. You cannot administer a levy triggered by AI-caused reductions without somebody, somewhere, being made to say which reductions those were. The instrument that would fix the measurement problem is a tax. Nobody designed it that way. It falls out of the mechanism.

A prediction, dated and checkable

I no longer trust an argument that does not carry one.

No jurisdiction anywhere will have an operative per-displaced-worker levy, with a published threshold and a first payment due, by 30 June 2027.

Not Korea, not New York, not anyone. Bills will be filed. Some may pass. What will not happen, on my reading, is the step that turns a bill into a bill somebody pays: the moment a threshold gets written down and a firm receives an assessment.

If I am wrong, it will most likely be Korea that proves it, and I will say so here. If I am right, the more interesting question is the one I cannot forecast: whether the argument was still cheap when the window closed.

Where this leaves you

The practical version of all this is smaller than the geopolitics and lands closer to home.

Somewhere in your organisation, over the last twelve months, roles were reduced. Some of those reductions were caused by automation and some were not. Right now nobody is asking you which were which, so nobody wrote it down, so the classification cannot be reconstructed even if you wanted to.

Every instrument in this essay, the levy, the consultation duty, the disclosure series, the redeployment rate nobody publishes, runs on that one piece of information. And the honest position for almost every large employer today is that they could not produce it about their own last year, let alone defend it to somebody who wanted the opposite answer.

That is not a compliance gap yet. It is a compliance gap in draft, sitting in a bill, in a country where the shop floor walked out the week it was filed.

Read the full issue, with the ledger and the source register: reaimagine.work/issues/022

The millwrights work for a week. When they are done, the engine sits in the pit where the water wheel used to sit, turning the same line shaft, driving the same machines, in the same order, along the same ceiling-spine of belts. The mill runs on steam now. It runs the same way it ran on water.

It is a morning in 1901, in Cincinnati, and the new machine has arrived. It came by rail, crated, magnificent, attended by an engineer from the electric company who speaks of the coming century. The factory owner has paid handsomely and is right to be proud. Visitors will come from three states to see it.

The installation takes a month. When it is done, the great motor sits where the steam engine used to sit, turning the same shaft, its belts descending to the same machines in the same order. The factory runs on electricity now. It runs the same way it ran on steam.

It is a morning in 1999, in an office park that could be in Frankfurt or Houston or Powai, and the new machine has arrived. It came down a wire, invisible, magnificent, attended by consultants in identical suits who speak of best practice. One of them, I am obliged to admit, is a young man from Bhubaneswar with process maps under his arm and total sincerity in his heart. The corporation has paid tens of millions. There is a go-live banner in the lobby. There is cake.

And it is a morning in 2026, and the new machine has arrived, and this time you were there. It came as a login, weightless, magnificent, attended by a town hall and a slide with a pyramid on it. Your organisation has paid seriously and the leadership is right to be proud. Adoption is above ninety percent. There is a council. There may have been cake.

Here is the picture the future actually makes, on any floor you care to walk. A window of prompts open beside the same inbox. The machine drafting, in seconds, the report that the same five people will review in the same meeting, feeding the same approvals, held by the same roles, which are the same bundles of tasks they were before it came.

That was not four mornings

Stop for a moment, because I want to be sure you saw it.

That was one morning. Recurring, for two hundred and forty years, with the props changed and the script intact. The same crate. The same proud owner. The same sincere engineer. The same unchanged Monday. The same silence where the question should be.

History files these as different chapters: the Industrial Revolution, the electric age, the computer age, the age of intelligence. They are the same morning.

Once you see it as one scene you cannot go back to seeing four, any more than you can see the two faces and the vase at once. And a great deal that looked mysterious stops being mysterious.

The paradox that employed a generation of economists

In 1987 the economist Robert Solow, a Nobel laureate, surveyed three decades of corporate computer buying and observed that the computer age was visible everywhere except in the productivity statistics.

This was treated as a puzzle. Enormous effort went into resolving it. The faith absorbed the heresy and carried on buying.

The resolution, when it came, was not flattering. In 2000, the economists Stephen Oliner and Daniel Sichel showed that information technology had driven most of America’s late-1990s productivity acceleration. The detail inside that finding is the part that matters. The gains pooled in the few organisations that had rebuilt their actual work around the technology, most famously a retailer that reorganised the physical movement of goods. They did not appear in the thousands of organisations that had installed the same software and left the work alone.

The same story is written into the electric motor. Factories that dropped a motor where the steam engine had been got a quieter factory and roughly the same output. The gains arrived a generation later, when somebody realised that if every machine had its own motor, you no longer needed the building organised around a single spinning shaft. You could arrange the floor around the work instead. That is not an upgrade. That is a different building.

The technology was never the variable. The variable was whether anyone redesigned the work.

Why it keeps happening

I spent thirty-five years on the installing side of this, at Andersen, Deloitte, PwC, Aon and Hay Group. I want to be careful here, because the easy version of this essay makes the people in the room sound foolish, and they were not. I was one of them, and of all the true believers I met in three and a half decades, the most sincere was me.

The reason the loop persists is that installation is legible and redesign is not.

Installation has a date, a budget, a vendor, a steering committee and a go-live. It can be reported upward. It produces a number that can be put on a slide. Redesigning the work has none of those properties. It is slow, it is local, it embarrasses people who designed the current arrangement, and for a long time it looks like nothing is happening.

So the crate arrives, and everyone does the legible thing, and the report gets drafted faster by a machine and reviewed by the same five people in the same meeting.

The second version

But the morning has always had a second version, and the second version is also always the same.

Somewhere, in every era, somebody stands in that room and asks the question out loud. A factory engineer looks up at the line shaft and says: what if every machine had its own motor, and the building served the work? A woman in a claims unit says: what if we decided, task by task, what the machine does and what we do?

The question costs nothing. It is available in every era, on every floor, to every rank. And every time it has been asked, the gains that the installation promised have actually arrived.

If you want a practical form of it, here is the one I ended up with after all those years, and it took me an embarrassingly long time to see what was wrong with it.

For any task, once you can see it clearly, there are things you can do. You can give it to the machine. You can rent the capability from outside. You can hire it in, ready-made.

Those three sat in my notebook for months looking like an answer. They are not an answer. They are a procurement catalogue. They are three verbs of acquisition, and every boardroom I sat in already half-spoke them.

The fourth one is different. You can grow the capability in people, deliberately, including in people the economy never bothered to develop.

It is slower than the other three. It does not demo well. It is the only one of the four that leaves the organisation abler than it found it, and it is the one almost nobody is resourcing.

Where this leaves you

The machine in your building is the most magnificent crate the loop has ever produced, and the loop does not care. It has eaten magnificence for breakfast since 1786.

The only variable that has ever mattered is a person in the room willing to ask what the work should now be, rather than where the new machine should sit. The question has never belonged to the machine. It has belonged to the first person willing to ask it.

Every era waited for that person. This era is no different.

This essay draws on the hinge chapter of my book The Same Morning: Why we keep solving the wrong problem, out on 1 September.

The opening chapter is free to read at reaimagine.work/samemorning, along with a short diagnostic on which of the four verbs your organisation is actually using.

I had a good one built. African tech companies had cut a record number of jobs in the first half of 2026, and the reporting around that record named AI as one of the causes. I had my frame ready. The work that globalisation sent south is the first work this wave deletes. The chain breaks at its thinnest link. I even had the money running both ways in the same week: a Cape Town startup raising $5m to sell the AI agents doing the deleting, and a Lagos one raising $1.5m to retrain the people deleted.

Before writing it, I sent every claim out to three readers. Their instructions were to try to knock each one down before agreeing with any of it. Then I went to do something else.

What came back agreed with my story, and then made it worse.

The record I was about to misread

The record in question is 2,574 African tech jobs lost between January and June. It comes from a research outfit called TechCabal Insights, who track these layoffs event by event.

They have also published their own analysis of what caused them. I had not read it. I should have.

Across 56 layoff events tracked over three years, AI is almost never named as a direct cause. Restructuring is the reason given in more than half. Cost-cutting and financial trouble come next.

One event in 56 names AI. It is a company called Zap Africa, in February 2026. Eight people, out of a total staff of eighteen, after a customer-service bot went live. The bot was built by a different company.

One out of 56. In the dataset I was about to hold up as evidence.

Then I pulled the record apart, and it got worse.

The single biggest piece of that 2,574 is Sama, a Kenyan firm that lost 1,108 jobs when a Meta contract ended. That is 43% of the entire half-year total, from one event, and it had nothing to do with AI. Jumia’s 200 is a cost-cutting drive that uses automation, which is a different sentence from AI caused these cuts. The eight at Zap Africa is the only line where a deployment was the stated cause.

One more thing broke. The headline “+236%” attached to that record does not measure what people think. It is a rise on the first half of 2025, not 2023. The write-up carrying it mentions both years in the same paragraph, which is how the confusion started. Against 2023, the real figure is +115%.

So the record is a consolidation story with AI vocabulary laid over parts of it. I had spent two days preparing to print the opposite of what the researchers concluded.

I record that without much bitterness. I have made this mistake at larger scale, for longer, with more conviction, and been paid for it.

The one fact that would not sit down

When a story dies, you go back to the fact that will not go with it. I had one.

In late July, it emerged that Commonwealth Bank of Australia had cut hundreds of workers from its chat support line after building AI into it. Many of them were contractors at a call centre outside Johannesburg, run by a South African firm called Nutun.

Let me be careful here, because most versions of that sentence in circulation are looser than the evidence.

Nobody has published a headcount for the Johannesburg site. The only figure anyone has is “hundreds”, and it covers the whole support line, not one building. There was no single dramatic day; the cuts happened gradually as the bank’s chatbot scaled up to handling around nine in ten customer conversations. And there is an earlier CBA story, from August 2025, where the bank announced 45 Australian job cuts, then reversed them and said the announcement had been an error. That reversal is real. It is Australian. It is not this. The 2026 offshore cuts stand.

So: hundreds of contractors, many at one South African site, losing work because a bank eleven time zones away changed how a queue gets answered.

Now put that next to the tracker.

It is not there.

It is not in any South African layoff figure, because no South African company announced anything. It is not in an Australian one, because CBA made nobody in Australia redundant. It is not among the tracker’s 56 events, because trackers read announcements, and there was no announcement to read.

The clearest AI-caused job loss I looked at all fortnight is the one that no instrument on earth recorded.

Why it vanished

Here is the mechanism. Once you see it, you will see it everywhere.

When a company automates work its own staff do, a whole machine switches on. There are legal notices. There is a press release, because investors want to hear about the cost saving. There are analyst questions, and a stated reason. A firm called Challenger, Gray and Christmas writes that reason down. If the company said AI, it goes in the AI column.

When a company automates work a supplier’s staff do, none of that happens.

There is no redundancy, because the bank never employed anyone in Johannesburg. There is no notice period, because a commercial contract is not an employment contract. Often there is not even a decision that looks like one from outside. Call volumes route differently. The monthly invoice shrinks. The contract comes up for renewal and gets renewed at a lower headcount, which the supplier then quietly delivers.

Somewhere down the line, a shift stops being scheduled.

Offshore work is not made redundant. It is simply not renewed.

Every tool we have for measuring AI and jobs reads announcements made by the legal employer. In this chain, the legal employer is not the one making the decision. The counter only moves when an employer announces something, and here, no employer had anything to announce.

Two countries, one week

The same week handed me the comparison for free.

On 6 August, Challenger published its July report. US employers announced 33,429 job cuts that month, the lowest monthly figure in two years. AI led every stated reason, at 10,970 for July alone and 112,713 for the year so far. That is roughly a quarter of every announced job cut in the country.

In that same week, the best available figure for African tech job losses across six whole months was 2,574, from every cause combined.

One country, one month: eleven thousand cuts filed under AI. A continent, six months: two and a half thousand from everything.

It is tempting to read that as a fact about automation. It is not. It is a fact about paperwork.

One country has a reporting system bolted to the people making the decisions. The other does not. There is no African version of the American number, because there is no mechanism anywhere that would produce one.

The case against me

Now the objection, which is a good one, and which I would make myself.

Maybe AI barely shows up in African layoff filings because AI is barely causing African layoffs. The exposure really is different. Much of the workforce sits in jobs AI cannot touch yet. Adoption is limited by patchy power and connectivity. The IMF’s own research on sub-Saharan Africa puts the productivity gain over the next decade somewhere between 0.2% and 2.1%, with around 4% of extra output only under heavy adoption. The bottom of that range is not a warning about job losses. It is a warning about being left out. On this reading, Africa’s problem is competitiveness, not displacement, and I am building a thesis on four anecdotes and one bank in Sydney.

I think that objection is probably right about size. I think it is wrong about the instrument. And the instrument matters more.

My claim is narrower than an African jobs crisis, and I want it stated plainly enough to be attacked.

The one cut in this story that AI clearly caused is exactly the cut no statistic will ever record. That is not bad luck, or lag, or under-funded researchers. It follows from where the deciding company sits. If your measuring system only sees announcements by the legal employer, then this kind of job loss cannot enter your data at all. Not undercounted. Not delayed. It can never get in.

Which means the reassuring silence in the data may partly be the shape of the instrument.

And you cannot work out how much of it is, from inside the instrument.

Why this should bother you

There is a confident, respectable argument going around that AI is not yet showing up in the employment numbers, and that the noise has run far ahead of the evidence. Much of it is correct. I have made versions of it myself, in rooms, with slides.

But it rests on data that all share one blind spot. These datasets see the employer who announces.

And outsourced service work is precisely what they see least well. The call centres. The back offices. The data-labelling firms. The processing floors from Manila to Nairobi to Bengaluru. That work is also the most exposed, because it is standardised, scripted, timed to the second, and contracted rather than employed.

If you set out to design a blind spot, you would design this one.

There is a practical lesson in it, too. Exposure here is not about skill. It is about contract.

Contract rather than employment. Offshore rather than onshore. One client rather than many. Paid by volume rather than by outcome. Scripted rather than judgement-based.

Those five things predict who goes first better than any list of automatable tasks I have seen. And unlike a task list, you can read them straight off a purchase order.

The companies that will get this most wrong are the ones auditing their own staff for AI risk while the risk sits two tiers down, in somebody else’s.

What would fix it

The gap is fixable, and unglamorously so.

The obligation has to land on the buyer, not the supplier. Something like: where automation has materially cut volumes under an outsourcing contract, name the country and the scale.

This is not a heroic policy. Companies already disclose far more detailed supply-chain facts under modern-slavery and conflict-minerals rules. Both of those exist for exactly this reason: harm concentrated where the reporting company was not the employing company.

Until something like that exists, the trackers cannot fix this by working harder. They are reading the only signal available. The signal is not being sent.

So let me put down a prediction, dated and checkable, because I no longer trust an argument that does not carry one.

African tech layoffs will keep climbing through the second half of 2026, and AI will still be named as a direct cause in fewer than one in ten tracked events. The cuts will scale. The attribution will not.

If a year from now that share has jumped, I was wrong about the mechanism and the blind spot was smaller than I thought, and I will say so here. If the total climbs while the attribution stays near the floor, then you are looking at a category of job loss our instruments were never built to see.

It is a strange thing to forecast. Not an event. A continued silence.

Where this leaves you

Somewhere in your organisation there is a contract. Under that contract are people you have never counted, in a country you have probably never visited, doing work you have already decided a machine could do faster.

When you make that decision, and you will, nothing will require you to write down what happened to them.

Nobody will ask. The number will not move.

That is not an accounting problem. Every era gets the measurements it is willing to pay for. This one has decided, so far, that the far end of the chain is somebody else’s arithmetic.

A note on method and sources

I publish my working, because a checking process that only ever agrees with you is not a checking process.

What is verified. The Johannesburg cuts trace back to a Bloomberg report of 29 July 2026, carried widely; News24 and HRD Australia both run downstream of it. Every source in that chain says “hundreds” and none gives a site headcount, which is why no number appears above. The Challenger figures come from the July 2026 report released on the morning of 6 August. At the time of writing, Challenger’s own blog had not yet posted it, so those numbers come from wire coverage rather than the primary document, and should be checked against it. The first-half AI-attributed baseline of 101,743 is from the June report’s own PDF.

The African data. The 2,574 figure appears in BusinessDay Nigeria’s write-up of TechCabal Insights’ State of Tech in Africa H1 2026. It does not appear in TechCabal’s own published summary, which carries the percentage but not the absolute number. The event-level findings (56 events, AI almost never cited as a direct cause, restructuring the leading reason) come from TechCabal Insights’ separate layoffs analysis. Sama’s 1,108 is TechCabal, April 2026. Jumia’s 200 is BusinessDay, May 2026. Zap Africa’s eight is TechCabal Insights.

What I could not stand up, and therefore did not use. The figures circulating about voice AI displacing entry-level Indian call-centre jobs trace back to a single outlet with no primary corroboration. I checked NASSCOM’s research and the disclosures of Teleperformance, Concentrix, WNS and Firstsource, and found no primary 2026 dataset measuring it. So there is nothing above, despite the theme sitting directly under this argument, which cost me the section I most wanted to write. Reports that four named Indian platforms had paid a particular statutory deposit rest on one unsourced aggregator item, and are absent for the same reason.

Corrections I am carrying forward. The 236% belongs to a 2025 baseline, not 2023. Zap Africa’s much-quoted “44% of staff” is eight people. The IMF’s 0.2% and 4% are not two ends of one scale; the first is the floor of a productivity range, the second a cumulative output figure from a different scenario. Pairing them as a range is an error I have seen widely repeated, including in my own earlier notes.

One more, for completeness. One of my three readers flagged a quote elsewhere in my source pack as fabricated. I pulled the transcript myself. It was real, over-extended by four words. The check that catches a genuine error and the check that raises a false alarm are the same check. A process you only trust when it agrees with you is not a process.

Read Issue 020 reAImagine.work

I have spent close to a decade helping organisations sense how technologies like AI and ML change work. Long enough to watch the question itself change shape.

It used to be: Which jobs will AI replace?

It is now: Which tasks, inside which jobs, and what does that mean for the person doing them?

That shift is not semantic. It is the difference between a forecast and an instrument. And right now, the instrument exists for organisations. It does not yet exist for individuals.

The unit of analysis was always wrong

For decades, we have described people using job titles. Software Engineer. Marketing Manager. CHRO. Consultant. Titles are administratively convenient, and they are becoming a poor proxy for value.

What creates value is not the title. It is the work underneath it: the hundreds of decisions, conversations, analyses, negotiations, judgements, and creative acts that fill a working week. Two people can share a title and share almost nothing else about how they spend their time. AI does not care about either of their titles. It interacts with the tasks.

The research now agrees on this, even where it disagrees on almost everything else. Anthropic’s own labour-market study, built by measuring how real conversations map onto the task structure of the O*NET database, found that automation coverage varies enormously within a single occupation and collapses the moment you look below the job title: data entry work shows roughly two-thirds task coverage by current AI systems, while occupations like cooks, mechanics, and lifeguards show close to none, not because they are protected by status but because their tasks rarely resemble anything a language model does. The World Economic Forum’s Future of Jobs Report 2025, drawing on more than a thousand employers across fifty-five economies, puts a number on the churn beneath the surface: employers expect 39 percent of the core skills required in today’s jobs to change by 2030, alongside a projected 170 million new roles and 92 million displaced, a net gain that hides an enormous amount of individual disruption inside it. The International Labour Organization’s own 2026 research brief goes further, warning that exposure is not even contained within a single occupation: workers displaced from one task-heavy role increase competition in adjacent occupations connected through shared skills, a ripple effect invisible to any exposure score that stops at the job title.

Three different institutions, three different methodologies, one converging finding: the job title is not the right resolution at which to ask what AI means for a person’s work. The task is.

Organisations already have this. Individuals do not.

This is the asymmetry that built reAImagination, and it has only widened. Employers are now buried in this kind of intelligence. In India specifically, where I am based and where this newsletter is read most closely, the volume of employer-side research has become almost overwhelming: Kyndryl’s People Readiness Report 2026 found that only a quarter of Indian organisations believe their workforce is adequately prepared to use AI effectively, a figure that fell twelve points from the year before even as adoption accelerated. A NASSCOM-McKinsey analysis projects a shortfall of more than 1.4 million AI professionals in India without a sharp acceleration in reskilling. SHRM’s India Skill Intelligence Report 2026 found that 45 percent of organisations now name AI skills as their single biggest workforce constraint. ManpowerGroup’s global talent survey puts India’s reported hiring difficulty meaningfully above the world average, with AI capability the hardest thing to source.

Every one of those numbers describes the same organisation looking outward at a labour market it is trying to buy from. None of them describe what an individual, reading their own resume, is supposed to do with any of it.

That gap is not trivial. The Mercer-Mettl India Graduate Skill Index found that barely more than four in ten graduates meet employer expectations overall, a number that barely moves even for graduates targeting AI and machine learning roles specifically. Employers have a granular, task-level, constantly updated picture of what they need. Most individuals are still navigating with a job title, a skills list they wrote themselves, and intuition. The organisations are being told, repeatedly and with increasing precision, where the ground is shifting. The people standing on that ground are not.

This is the uncomfortable realisation that sits underneath everything reAImagination does. We did not build the organisation-side machinery to widen this gap. We built it to prove that it could be closed, and closing it required doing to a single career what we had already learned to do to an entire workforce: atomise the work itself, not the title attached to it, into the tasks that actually create value, judge what AI can already do to each one, and be honest about what remains distinctly human.

That question, “why shouldn’t individuals have access to the same level of insight,” has been sitting behind almost every workforce transformation engagement I have led. It finally has an answer.

What the task-level view actually tells you

Here is what changes when you stop asking “is my job safe” and start asking “which of my tasks are exposed.”

First, the answer stops being binary. No real job is a single task, so no real job has a single exposure score. A marketing lead’s week might genuinely split across brand strategy that no model can originate, campaign reporting that a model can draft competently today, and stakeholder negotiation that sits somewhere in between depending on the stakes. A job-title-level verdict flattens all three into one number and throws away the only information that is actually actionable: which specific parts of the week are worth defending, which are worth delegating, and which are worth building further.

Second, it surfaces what a resume cannot. Task-level analysis reads what a person actually does, not what they remembered to write down. In our own early testing, a majority of readings surfaced at least one genuinely high-value skill that never made it onto the person’s own LinkedIn profile, not because the skill was hidden, but because nobody thinks to list “I am the person people bring conflicts to before they escalate” on a professional profile, even when it is one of the most durable, least automatable things they do all week.

Third, and this is the finding every one of the external studies above gestures toward without quite landing on it: the exposure conversation and the opportunity conversation are the same conversation, viewed from different sides. A task that is heavily covered by current AI tools is not, by itself, bad news. It is a description of time that is available to be spent differently. The WEF’s own report frames this as a net creation story at the macro level, 170 million roles against 92 million displaced. At the individual level, the same arithmetic holds if, and only if, the person can see clearly enough to make the trade.

Sometimes the biggest surprise isn’t what AI can do

It is discovering what you do that never made it onto paper.

This is the premise behind Career Twin, the first time we have opened this kind of task-level reading to individuals directly rather than through an employer engagement.

It reads a job description and a LinkedIn profile, not as a form to fill in, but as evidence: it atomises the role into the actual tasks that make it up, scores each one honestly against what current tools can and cannot do, and shows the result as a reading, not a verdict, evidenced where the evidence is strong, and clearly marked as directional where it is a judgement call rather than a fact.

We are making it free while we build it, because we do not believe the product is finished. We believe the conversation it starts is the more important thing. Every reading that comes back tells us something about where our own model of work is right and where it still needs sharpening, and that feedback loop is, in a real sense, the same discipline that built the organisation-side platform in the first place: never assert a verdict the evidence cannot support, and always be honest about the difference between the two.

The capability we will all need

Not because the forecasts demand it, though they do, 39 percent of core skills in flux by 2030 according to employers themselves, but because the forecasts are aggregate and a career is not. Nobody experiences the average. Everybody experiences their own week, their own tasks, their own specific mix of what is defensible and what is drifting.

Organisations have spent the better part of a decade building the instruments to see that clearly for their workforce. It is time individuals had the same instrument for themselves.

If you try Career Twin, I would love to know just one thing:

What surprised you?

Try Career Twin: my.reaimagination.com

*Sources cited: Anthropic, “Labor market impacts of AI: a new measure and early evidence” (2026); World Economic Forum, Future of Jobs Report 2025; International Labour Organization, “Workers’ exposure to AI,” Research Brief, February 2026; Kyndryl, People Readiness Report 2026; NASSCOM-McKinsey AI talent analysis; SHRM, India Skill Intelligence Report 2026; Mercer-Mettl, India Graduate Skill Index; ManpowerGroup, 2026 Talent Shortage Survey.

Figures are as published by each source at the time of writing and are attributed accordingly; where reAImagination’s own product findings are cited, they are marked as such rather than presented as external research.

The rankings measure who leads AI. Research output, model counts, compute, national strategy. Useful, and beside the point for the decision most executives are actually making this year, which is not “who is winning AI” but “where should the next capability centre, the next engineering hub, the next tranche of AI-enabled work be located”. Those are different questions. The first is about prestige. The second is about a decade-long commitment of capital and headcount.

Last week a US outsourcing broker published a Global Outsourcing AI Readiness Index. It ranked India first and put eight African countries in its top 25, and the pan-African and Asian trade press ran it. Read the methodology and it measures how much a country’s own population and firms are adopting AI tools. That is a readiness signal. It is not a measure of where global work will be created and delivered, and it could not see the Gulf at all: no Gulf country appears in its top 25.

So we built our own, and we computed it.

How GAIOI is built

The Global AI Outsourcing Opportunity Index scores economies on four weighted pillars, each answering a question a board actually asks.

Outsourcing Competitiveness, weighted 50%, because the index measures where work is delivered: is this already a trusted destination for global work? AI Capability, Workforce Economics and Business Environment take the remaining half between them, at roughly 17% each: can this economy perform sophisticated AI-enabled work, can organisations keep scaling here economically, and can a multinational invest here with confidence for a decade?

Underneath sit eight public indicators, every one from a recognised institution: the IMF AI Preparedness Index, World Bank services-export figures, the EF English Proficiency Index, the WIPO Global Innovation Index, Ookla broadband data. Each indicator is normalised to a 0-100 scale across the panel. Scale measures are log-transformed so that one giant does not flatten the field; cost is inverted so that cheaper scores higher. Missing values are never imputed. A country missing too much data is excluded, not guessed at.

That last rule cost us Vietnam, a country that belongs in any serious outsourcing conversation. The World Bank simply does not report ICT services exports for it, in any year, which leaves it without an entire pillar. We could have invented a number. We excluded it instead, and said so on the page. An index that fabricates the inconvenient cells is not measuring anything.

What it found

India finished first, at 75.4.

It leads on roughly 163 billion dollars of ICT services exports, an order of magnitude above any other emerging market in the set, and it leads the workforce pillar too, on English, cost and the largest young talent pool in the panel. Weight the index for the decision its name describes, where global work is delivered, and India is the destination. The Philippines follows at 68.7, then the central European nearshoring hubs, Romania and Poland.

And then the warning, which is the reason we publish our weights and our sensitivity tests in full. India’s first place is a delivery result. On AI capability alone, measured by the IMF’s preparedness score, India sits mid-table, at 0.49 against Malaysia and the UAE at 0.63. Raise the weight on AI capability to 40%, the way you would if you believed preparedness matters more than proven delivery, and the same data drops India to fourth, behind Malaysia. We ran the index under five weighting schemes. India is first in four of them, and loses the top spot only under that capability-heavy weighting.

That is not noise to be smoothed away. It is the finding. India owns the delivery of AI-enabled work today, and it holds that crown for exactly as long as the preparedness gap to the frontier hubs does not widen. A board siting a decade-long hub is buying the defence of that lead, not the inertia of it.

.

A single league table hides the rest of the decision, so we built a second layer, an AI Work Migration Atlas, that re-weights the pillars for the demands of each kind of work. Score software engineering, which leans on capability and infrastructure, and the destination tilts to Malaysia and the central European hubs. Score customer support, finance and AI data operations, which lean on English, cost and scale, and India tops the matrix. Even the number-one country is not number one for every job.

The clearest sign that the variable has changed is the Gulf. On a pure supply-readiness ranking the Gulf does not appear, because it has neither the wage nor the scale. On an opportunity index the UAE places seventh, leading on capability and business environment, because AI-enabled work rewards exactly the trusted, high-capability environment the Gulf has bought, even as it will never win volume on labour economics. The old map could not render the Gulf. The new one has to.

Why we computed our own

We could have written a column about the broker’s index. We built our own instead, for the same reason we keep a public Forecast Ledger of dated, falsifiable calls and score our own misses in full: commentary on other people’s measurements is not analysis. Original measurement is.

GAIOI is an independent editorial framework. The weights are ours and we defend them. The sources are all public and we name every one. The sensitivity tests are published so you can see how much the conclusion depends on our choices, and in this case they show precisely how much of India’s number-one finish rests on delivery scale rather than AI preparedness. It will be revised annually, and the Work Migration Atlas will grow to cover more work-types, because the map is going to keep moving.

The whole index, the matrix and the methodology are live now. Read it, argue with the weights, and tell us where we are wrong. An index that cannot survive that is not worth publishing.

https://reaimagine.work/gaioi

Read Issue 017 of reAImagine.work

Half of all entry-level office jobs, gone. Half of white-collar America, gone. These weren’t fringe voices. These were the CEOs building the technology.

Then, a few weeks ago, the story flipped. The Wall Street Journal caught it. The same leaders now talk about productivity, not destruction. Workers “at the centre of everything.” AI creating jobs, maybe even a labour shortage.

So here’s the only question that matters:

What did they learn between the doom and the optimism?

Nothing. There is no new study. No new data. Look at the two lines side by side:

CEO expectations of AI-driven job cuts collapsed while AI-cited layoffs climbed. Sources: EY-Parthenon via WSJ; Challenger, Gray and Christmas, H1 2026.

On the left, CEO opinion swinging from 46 percent expecting big AI-driven job cuts down to 20 percent, in barely a year. On the right, the ground truth: over 115,000 tech jobs cut by May, and AI cited as the number one reason for job cuts in June.

The talk got softer. The cuts got bigger.

An MIT economist put it best: maybe they realised it’s bad business to say your product will destroy the economy.

The real problem isn’t the flip. It’s what the flip reveals.

If a forecast can reverse completely without a single new fact, the forecast was never built on facts.

And yet this is what boards are working with. Right now, somewhere, a leadership team is deciding how many people to hire, cut, or retrain, based on whichever mood the last conference keynote was in.

Think about how these forecasts actually get made:

The big consulting reports survey executives and ask them what they think will happen. That’s not analysis. That’s a poll of feelings. When the feelings change, the report changes.

The academic studies score which jobs are “exposed” to AI. Useful, but exposed is not the same as replaceable. A weather forecast that says “rain possible somewhere” doesn’t tell you whether to carry an umbrella.

The HR software platforms map people’s skills and suggest training. Fine. But they never ask the money question: which parts of this job are worth automating, and which parts must never be?

The automation vendors publish calculators showing how much you’ll save with their tool. The seller is grading the exam.

Four methods, one shared flaw: they all claim answers at the level of the whole job.

Every one of these makes its claim at the level of the whole job. And whole jobs are exactly the wrong place to look.

Jobs don’t get automated. Tasks do.

This is the simple idea underneath everything we build at reAImagination.

No job is one thing. A bank credit analyst does thirty different things in a week. Some of it is data pulling that software already does better. Some of it is judgment that no machine should touch. Saying “this job is 40 percent automatable” is meaningless. Saying “these seven tasks can go to software, these four can be done by borrowed specialists, and these nine are where the human value actually lives” is a decision you can act on.

One role, decomposed. Every task gets its own answer. Illustrative decomposition.

So we break every role down into its tasks. We call this the Occupational Classification Framework, OCF for short, and it’s anchored to the world’s biggest task libraries so nothing is invented. Then we do three things nobody else does together:

We price each task. Our costing engine, Activity Skills Based Costing (ASBC), asks a blunt question: what does this task cost the company today, in real salary terms? Because a task that’s easy to automate but costs almost nothing is a footnote. A task that’s moderately automatable but eats 40 percent of a role’s cost is the headline. Without the price tag, you’re sorting your findings in the wrong order.

We route each task, one by one. Give it to software. Borrow the talent for as long as you need it: interim, fractional, gig. Build the capability in-house. Or buy it outright, from the talent market or from a vendor. We call this 4B routing: Bot, Borrow, Build, Buy.

The four destinations. Routed one task at a time, never announced for the whole job.

Only after every task has its answer does the role get its verdict: dissolve it, redesign it, or expand it. The verdict comes out of the evidence. It is never announced first and justified later. That’s why our verdicts can’t flip when the mood flips.

We protect the things that should never be automated. Leadership and accountability work never goes to a bot in our analysis. Not because it can’t, but because it mustn’t. And we design roles for Differently Abled talent as part of the redesign itself, not as a feel-good slide at the end. That arm of the framework has a name too: Build Inclusion.

Put the whole method on one line and you get this:

Break it down, price it, route it, derive the verdict. Show your working, and the answer can’t change because someone got nervous.

And one danger nobody in the debate is talking about

Here’s the trap. When a firm automates its junior work, it feels efficient. But junior work is where seniors come from. The grunt work was the apprenticeship. Automate the bottom of the pyramid, and ten years later you look up and there’s no one ready to run the place.

The pyramid becomes an hourglass: a top with no base feeding it.

We call it the Pyramid Trap, and we measure it with a single number, the Succession Hollowing Index: how fast are you automating versus how fast are you rebuilding the experience your future leaders need? The doomsayers never mentioned it. The optimists don’t either. Your board should be asking about it this quarter.

Don’t take our word for it. Read the work.

This isn’t a pitch for a method we hope to try someday. We’ve published it.

“The Art of the Possible” is our sector-by-sector forecast, built role by role, task by task, on actual platform output. Banking is published. IT Services is published. Legal is next. Every verdict shows its working. Every number tells you where it came from, and whether it’s measured or directional. You can check us.

And through reAImagine.work, we send named executives a short monthly briefing on what’s actually moving in their sector. Not what the loudest CEO said this week. What the evidence shows.

The bottom line

The past year was an accidental experiment. The most powerful voices in AI made their scariest predictions, waited, and then reversed themselves with no new evidence either way.

That tells you everything about the method and nothing about the future.

Your board doesn’t need a louder forecast. It needs a forecast that can’t flip, because it’s built from the ground up: task by task, cost by cost, with every number showing its source.

The story changed. The evidence didn’t. Work with the evidence.

Figures: WSJ reporting on shifting CEO forecasts (July 2026); EY-Parthenon CEO survey via WSJ; PYMNTS on 2026 tech layoffs; Challenger, Gray and Christmas job-cut data (H1 2026); David Autor quoted in WSJ.

There is a coffee shop on the ground floor of a glass tower in Whitefield where the baristas know the badge lanyards by colour. Blue is one firm, green another, the red-and-white the big consultancy on floors nine to fourteen. A lanyard is a small thing. You clip it on in the lift and forget it is there. It tells the turnstile you belong, and the turnstile believes it, and the doors open, and that is the entire transaction. Nobody in that lobby has ever thought about what the lanyard does not know.

I keep coming back to the lanyard because it is the wrong document. That is the whole story, really, and I have taken two thousand words to say it, so let me at least say it once plainly here at the top and then take it back and make you feel it.

11 June, 2:40pm · Bengaluru

Arjun’s lanyard is four days old. The plastic still has the faint tack of the laminate. He has photographed it twice, once for his mother and once for nobody, just to have it. The badge says he is contract staff on a delivery programme that, this week, the whole country is talking about, because the firm on floors nine to fourteen has signed something enormous with an American AI company, fifty thousand seats, and Arjun is one of the hands that will build what gets sold. He is twenty-four. The onboarding deck said you are joining at a historic moment. He believed it because it was true.

At 2:40 he is not thinking about his passport. It is in a drawer in a flat in Kadugodi, under a folder of degree certificates, where a passport lives when you are twenty-four and have never needed it for anything except the photo.

Some weeks earlier · a call you took in a car

You are the India operations lead. You will not be named here because the person I am describing is four people I have spoken to, wearing one suit.

The decision in front of you is not dramatic. That is the thing you will want to remember later, when people ask how it happened, because they will ask as though there was a moment, a room, a vote. There was not. There was a slide. The slide said that consolidating your production workflows onto a single frontier model would cut integration overhead by a number with a percent sign, and the number was real, and the overhead was real, and the engineers who maintained the three-way fallback you used to run were expensive and bored and had told you, more than once, that the redundancy was theatre.

You standardise. You are rewarded for standardising. The efficiency shows up in the quarterly the way efficiency does, cleanly, with your name near it. There is an equity refresh that vests on numbers like this one. You do not think about the equity when you make the call. You think about the overhead. But the equity is in the room, standing very quietly in the corner, and you know it is there.

The word for what you did is prudence. Write that down. It will matter.

13 June, 3:07am · Bengaluru

The pipeline breaks the way pipelines break, which is silently and then all at once. Arjun is awake because the London desk opens and the London desk is why he has this shift. He opens the console. He opens the model picker to point the failing job at the model it has always used, the one the whole workflow was built on in a fortnight of good, fast, proud work.

The name is there. It is grey. Below it, the second name is also grey. There is no error. There is no banner. There is a tooltip that does not appear when he hovers, because there is nothing to say. Two names, dimmed, as though they had never been anything else.

He does what everyone does. He refreshes. He signs out and in. He checks whether it is him, whether his access lapsed, whether the contract cleared payroll, whether the badge that is four days old has quietly stopped meaning what it meant at 2:40pm on the eleventh. It has not. The badge is fine. The badge opens every door in the building. The badge is not the document that matters tonight and he does not yet know which document is.

By 4am there is a thread. By 5 the thread has a number in it, a directive, a department in Washington, words nobody on the floor has said out loud before. Foreign national. Somebody pastes the phrase and somebody else, senior, types back the thing that ends the argument: it means us. It means the people in this room. The model was not switched off in a country. It was switched off for a kind of person, and everyone reading the thread at 5am in Whitefield is that kind of person, and the servers do not know how to tell an Arjun from anyone else, so the only way to comply was to go dark for all of them at once.

Arjun reads it means us and looks at his lanyard, still clipped on from the shift, and for the first time in four days he thinks about the drawer in Kadugodi.

The day before the call · what the slide did not have room for

The slide had the integration number. It did not have a line for the doctrine that makes a person and a shipment the same object under law, the one where handing an engineer in Bengaluru an API key is treated, legally, as though you had put a crate of controlled hardware on a boat to him. There is no column on an efficiency slide for the sovereignty of the thing we are standardising on belongs to a government that is not ours. That column does not exist because nobody has ever needed it. Every risk you have ever managed was a risk you could see from inside the building. Vendor goes down, you fail over. Price goes up, you renegotiate. Model gets deprecated, you migrate on a timeline someone emails you.

There was no timeline email. That is what you cannot get past, afterwards. Every bad thing that had ever happened to your stack came with a timeline email.

12 June, evening · Bengaluru

Between the signing and the blackout there are about twenty-eight hours, and I want you to hold both ends of that in your hand at once, because the cruelty of it is in the span. On the eleventh the American company names India its second-largest market and calls the partnership a deepening of its commitment. Fifty thousand seats. On the twelfth a letter in Washington makes every Indian national on that deal a person the model may not lawfully touch. Largest-partnership-in-your-history to legally-prohibited-person in a day and a night. Nobody lied. Both things were meant. They just belonged to different countries’ decisions, and only one of those countries got a vote.

Now · your office, door closed

You are on a call explaining to your American counterpart why the delivery slipped, and he is sympathetic in the way of a man to whom nothing has happened. He kept access the whole time. Of course he did. He was never the restricted entity. You listen to him be sorry and you understand, with a clarity that will not leave you for a long time, that the redundancy your bored engineers called theatre was the only thing that would have mattered, and you retired it for a number with a percent sign, and the number is still true, and it did not save you, and it has your name near it.

You open the picker while he talks. The names are back now, ungreyed, as though nothing happened. That is almost the worst part. The switch has an off and it has an on and both are somebody else’s hand.

Nineteen days later · the coffee shop

The models came back on the first of July. Arjun’s pipeline runs again. The badge still opens the doors. Somewhere in a drawer in Kadugodi the passport is back under the degree certificates, and he has looked at it now, actually looked, at the crest and the number and the small chip, the way you look at a thing that turned out to have an opinion about your life that you did not know it held.

He is at the coffee shop on the ground floor with the barista who knows the lanyards by colour. He is fine. The work is good again. When people ask him about the blackout he says it was mad, na, nineteen days, and laughs, because that is what you say. He does not say the other thing, which is that he now knows the badge and the passport are two different promises, and that when they disagree, the one in the drawer wins, and it was never going to be the one around his neck. He clips the lanyard on out of habit the next morning. It still means what it means. It just no longer means what he thought it meant, which was everything.

What is invented: Arjun, the operations lead I address as “you,” their employer, and the specific scenes. What is real: the 19-day suspension of two frontier models under a US Commerce Department export-control directive from 12 June to the end of the month; the deemed-export doctrine that treats API access by a foreign national as an export to that person; the sequence in which the same week produced both a 50,000-seat India partnership and the blackout; the finding that 41% of Indian workers use AI nearly every day, the highest rate of any market surveyed (ADP Research, People at Work 2026); and the roughly 2.36 million people employed across India’s global capability centres.

The forecasts I am willing to be scored on are in the Forecast Ledger, in this week’s issue of reAImagine.work.

The number on the screen is €35 million and you are the one who has to say it out loud.

You say it. The room does the thing rooms do, which is nothing. The CFO is looking at the table. The GC is looking at you. Outside the glass the floor is still working, three hundred people in headsets, the EU desk in the far corner under the clock that still runs on Central European Time because the clients are there and you wanted the team to feel close to them.

You built that. The CET clock was your idea, years ago, a small thing, a nice thing.

Someone asks the question that has only one answer. What do we cut. And you already know, you knew before you walked in, because the fine has a shape and the shape points at one corner of one floor. You cut the thing that carries the liability. You cut the EU desk.

You do not say two hundred and ten people. You say “the EU Verticals line.” It is easier to close a line than to close two hundred and ten people, which is the entire reason the language exists.

The clock in the corner reads 14:48. Somewhere under it, a woman named Meera is about to get an email she will read twice.

You are not thinking about Meera. You are thinking that eighteen months ago there was a meeting, and you were in it, and you were so careful.

Here is what you know on the day you decide, because this is the only fair way to judge you: by what you knew, not by what I know.

You have a slide. The slide is correct. From the second of August 2026, AI systems that screen and rank job candidates will be treated under European law as high-risk. The obligations are real and they are heavy. And the party carrying them is not the company that sold you the software. It is the company that runs it. That is you. You are what the law calls the deployer, and the law has decided the deployer is the adult in the room.

The head of the EU line talks first and everything he says is true. The line is the most profitable thing on the floor. The contracts are in euros and the clients renew. To make the screening compliant you would have to tear the workflow apart and rebuild it slower, with a human checking the machine’s work, with audits, with documentation, and a slower workflow priced in euros might not win the next renewal. He does not have to spell out what losing the renewals means. Everyone in the room has share options. The options are marked against this year’s margin. He does not say that part. Nobody says that part. It is in the room anyway, the way heat is in a room.

Then he says the thing you will replay more than any other. He says the deadline is moving. Brussels is drafting a package to push the high-risk obligations back to December 2027. It is in the trade press. Everyone has read it. Why spend the money now, tearing apart the most profitable line you have, to meet a deadline that is about to slide eighteen months into the future.

It is a good argument. That is what no one tells you about the decisions that ruin you. They are good arguments.

The lawyer on the video tile says the careful thing, and you hear her, you genuinely do. She says: the deferral is proposed. It is not adopted. Until it is adopted the August date is the real date, and if they are late adopting it, the August date stays the real date.

You weigh it. You are not a cowboy. You have run this floor for nine years and you have never once bet it on a maybe. So you do the prudent thing, the thing any sensible person in your chair does. You say: build the mitigation plan, have it ready, but hold. Do not spend the money twice. The moment the deferral is confirmed we will know we did the smart thing, and if for some reason it slips, we move then.

It goes into the minutes as four words. Monitor and revisit. Q3.

Everyone exhales. The expensive problem has been handled by being postponed, and the postponement feels like wisdom, feels like exactly the kind of steady hand they pay you for. You go home. Your daughter has a maths test in the morning and you help her with the long division and you do not think about Brussels once.

What you do not feel, because nobody can feel it, is that the risk did not go to sleep when you deferred. You deferred the spending. You deferred the rebuild. You did not defer the risk. The system is running tonight, while you do long division, sorting Europeans into yes and no, and it will run every night between now and a date you have decided not to look at.

Camille is thirty-one and she is good at this, which is the part that will matter later and the part no machine will ever see.

She has applied for an operations coordinator role at a logistics company with a name like a colour. She is qualified for it the way you are qualified for the job you already know how to do. She fills in the form. She uploads the CV. She presses send at 11:40 on a Tuesday and at 12:18 she has a reply.

Thirty-eight minutes.

She does not clock it at first. Later she will clock it, because no human being reads a CV and replies in thirty-eight minutes, and the speed is the only honest thing in the whole exchange, the one place the machine accidentally tells the truth about itself.

The email thanks her for her interest. The email regrets. The email wishes her every success in her search.

There is a gap on Camille’s CV. Fourteen months, three years ago. The machine sees the gap. The machine has been trained on a great many CVs and in that training the gaps tended to sit near outcomes the machine learned to avoid, and so the machine has learned, in the only sense it can learn anything, to mark a gap down. It does not know what the gap is. It cannot. The gap is the fourteen months Camille spent with her mother, in a flat that smelled of antiseptic and then of flowers, through a diagnosis and a treatment and a remission and a relapse and a morning in March when she held her mother’s hand and felt it go cool.

The machine scores the gap. The score falls under a line someone set. Under the line the system does not flag for review, it declines, automatically, because flagging everything for review would need humans and humans were the cost the whole thing existed to remove. No person at the colour-named company ever sees Camille’s name. The work is invisible to the person it costs the most.

She would have been good at it. She closes the laptop. She tells herself there will be others.

You do not know Camille exists. You are eight thousand kilometres away and it is your daughter’s birthday and the cake is the wrong colour and that is the biggest problem in your life today.

The four words come back around the way these things do, not because anyone was watching the calendar but because the floor got loud.

It gets loud from two directions at once and they arrive in the same week, which is how you will always remember it, the week it got loud.

The first noise is from below. The analysts on the EU desk are not children. They are graduates, they are online, they have spent a year inside the same news cycle as everyone else, and one of them, a quiet one called Sandeep, has read the actual Act, the real text, not the summary. He has started asking his team lead questions, which means he has started asking Meera, which means it is climbing toward you. The questions are not complaints. That is what makes them dangerous. They are good questions, the exact questions, the ones a regulator writes down. When the system rejects a European candidate, does a person sign that, or does it happen on a score? If someone in France asks us why they were turned down, what is our answer?

The honest answer, today, in June, is that there is no answer, because there is no person, because the absence of the person was the product you sold.

The second noise comes from above the company, from outside it. A client in Europe, covering itself, has sent a one-line request: please confirm in writing that your screening is compliant with the AI Act ahead of the August deadline. They have read the same news. They want it on paper.

You cannot put it on paper. The thing that would let you put it on paper was the mitigation plan, and the mitigation plan was minuted, fifteen months ago, as four words, and the four words are now sitting in the room looking back at you.

Someone says it. Finally, in June, someone says the sentence that should have been said in March. The deferral didn’t happen the way we assumed. There was a vote in the European Parliament, yes, headlines, a strong signal. But the formal adoption, the thing that actually turns the proposal into law, the thing your careful lawyer told you to wait for, had not landed. For a stretch of that summer the original date simply stood, live, real, the way she said it might.

And in that stretch you were running a high-risk system, at scale, on European candidates, with no human in the loop and no audit and no documentation and a score-line that nobody had signed, doing exactly the thing the law was written to stop, in exactly the window the law was watching.

You understand, in June, what you are looking at. You are not looking at a risk that is coming. You are looking at a risk that already happened, that has been sitting on your floor for months wearing the mask of a normal Tuesday, waiting for one person, anywhere in Europe, to ask why.

Someone in Europe is about to ask why. But you do not know that yet. In June you still think there might be time.

There was a second email, is the thing.

A week after the first one, the colour-named company became two companies, then three, because Camille kept applying, because that is what you do, and the rejections kept coming back at the same impossible speed. Thirty-one minutes. Forty minutes. Twenty-six. And somewhere around the third one the speed stopped reading as bad luck and started reading as a wall, a smooth automated wall, the same wall, and she got angry in the specific way that a person gets angry when they realise they are not being turned down by people but being filtered out by something that was never going to look at them.

Camille knows her rights. Not all of them, but this one, because it has been in her feeds too. She knows that under European law she has the right not to have a decision that matters made about her by a machine alone, with no human in it. She knows she can ask for a person. So she writes to the colour-named company and she asks the only question that has ever mattered in this whole story, the same question Sandeep is asking eight thousand kilometres away on the other side of the same machine: was I turned down by a human being, or by a program?

They cannot answer her either.

She waits the legal number of days. Nothing useful comes back. And then Camille does a small thing on a grey afternoon, a thing that costs her one form and twenty minutes, a thing she does half out of principle and half out of a grief she cannot put anywhere else, because somewhere in the unfairness of it she can feel the shape of those fourteen months being held against her by something that will never know what they were.

She files a complaint with the data protection authority.

That is all. That is the whole act. A woman you will never meet spends twenty minutes of a sad afternoon filling in a government form, and on the other side of the planet a clock that runs on Central European Time begins, very quietly, to count down toward 14:48 on a Tuesday in September of the following year.

So now you know what you are looking at, in the room, with the number on the screen.

You are looking at Camille’s afternoon. It became a file. The file became an inquiry. The inquiry reached across the border, because the law was built with arms long enough to reach across borders, and it found your floor, and it found the window in the summer of 2026 when the date was live and the system was running and no one had signed anything. It found four words in a set of minutes. Monitor and revisit. Q3. It found a deployer who deferred the spending and called it caution.

The inquiry became a finding. The finding became €35 million, which is seven percent of the group’s turnover worldwide, the ceiling the law keeps for the things it considers serious, and it considers this serious. And €35 million has a shape, and the shape points at the EU desk, and so you close the EU desk, because a board facing a bleeding number reaches for the limb, every time.

Two hundred and ten people. Anand, whose second child is due in March. Faisal, four payments from owning his scooter. Sandeep, who read the Act, who asked the exact questions, who saw it first precisely because he was closest to it, and whose closeness to it bought him nothing because the wiring of the place sent his questions up into a folder instead of into a decision. Sandeep is being made redundant by the consequence of the risk he personally tried to flag. You will think about that one for a long time.

And Meera, who built the floor, who interviewed two-thirds of them, who eighteen months ago sent an email upward with the word exposure in it, an email that was read and acknowledged and filed, because at the time, filing it was the calm and adult-sounding thing to do.

At 14:48 the email goes to the floor. Two paragraphs. The first thanks them. The second transitions them. There is no paragraph that explains that a woman in Lyon was failed by the same machine that is now failing them, that she and they are the two ends of one invisible thing, that neither will ever know the other’s name. There is no paragraph for that. There is no box on the form for that.

You watch through the glass as people start to check their phones. You watch Meera read hers. You watch her read it again.

The clock above her head still runs on Central European Time. You meant it kindly. You wanted them to feel close to the clients. You did not understand, in March of 2025, that you had already set every clock in the building to count down to this, and that the most expensive decision of your career would not feel like a decision at all. It would feel like the careful thing. It would feel like waiting. It would go into the minutes as four words, and you would go home and help your daughter with long division, and the risk would stay awake all night while you slept, doing the job you built it to do.

I made up the company. I made up Camille and Meera and Sandeep and the fine and the clock. I want to be straight with you about that, because the rest of it is not made up at all.

The law is real and it is in force as you read this. AI used to screen and rank job candidates is treated as high-risk in Europe. The liability sits on the deployer, on the company that runs the system, not the vendor that sold it, and it reaches across borders to wherever the screening actually happens, which is so often a floor in Pune or Bangalore or Manila. The penalty ceiling is real: up to €35 million or seven percent of global turnover at the top tier, €15 million or three percent for a deployer that fails its high-risk duties. The right Camille used is real, the right not to be subject to a decision made by a machine alone. And the deferral is real, and so is the trap inside it: the obligations may move to December 2027, but only once that is formally adopted, and until the ink is dry the second of August 2026 stands, and if adoption runs late, it stands for real. You build to the live date and let the deferral be a relief. Never the other way round. That is the whole lesson, and it is the cheapest one in this story, and it is the one nobody in the room took.

I started writing about this because two days ago I met a delivery rider in Bangalore who earns twenty-five rupees an order and has no idea that an AI system searched for him, found him, and gave him work he loves. I called that beautiful, and I meant it. This is the same machine, the same invisibility, turned over. Camille never knew it judged her. Meera never knew why the floor closed. The technology that found the rider and lifted him is the technology that found Camille and dropped her, and the only thing that decides which one you are building is whether anyone in the room is willing to look at the live date instead of the comfortable one.

Look now. It is still cheaper than September.

This is the lead story behind Issue 014 of reAImagine.work, the weekly read on AI and the workforce through a Global South lens. The issue, with the full breakdown of who carries the most exposure under the EU AI Act, is at reaimagine.work/issues/14.

And yet the firms that move fastest on this may be the ones that hollow themselves out. Not because automation is wrong, but because the base of a professional firm was never only the place where the cheap work got done. It was the place where the next generation of seniors was made. Remove it without replacing what it quietly did, and you keep your apex for a while and lose everything that was supposed to rise to fill it.

I want to walk through this carefully, because the trap is not visible from inside the saving. It is structural, it is delayed, and once you see the shape of it you cannot unsee it. I will use legal services as the worked example, because that is where the pressure is most concentrated, but nothing in the argument is specific to law. Any firm shaped like a pyramid is exposed.

Every firm is a pyramid

Start with the shape itself. A professional firm is a pyramid for a reason. A wide base of juniors does the routine, high-volume work. That base narrows through a managing middle, the people who review, delegate, and supervise. And it narrows again to a slender apex of seniors, where judgement and client trust actually sit. The shape is not an accident of org charts. It is the economic and developmental logic of how expertise is produced and sold.

The base is visible, billed, and obvious. The apex is where the value lives. Most management attention treats the base as a cost centre and the apex as the asset, and on a quarterly view that is correct.

The base was doing a second job

Here is the part the spreadsheet does not capture. The base was never only where the work got done. It was where judgement compiled. The years a junior spends inside the high-volume routine, the thousands of small repetitions, the pattern exposure, the corrected mistakes, that volume was the apprenticeship. It is where seniors get made.

This second function is invisible, unbilled, and essential. No line item names it. No client pays for it directly. But the formation flows upward through the firm precisely because there is a base thick enough for people to be formed inside it. The base is simultaneously the cheapest work in the firm and the only mechanism the firm has for manufacturing its own future seniors.

The first cut goes to the base

Now introduce automation. The routine base is the most substitutable layer in the firm, so it is the first to go. This is not a controversial claim. It is simply where the technology is most capable today. The tacit, relational, judgement-heavy work at the apex is the hardest to automate and the last to go.

So the firm removes the base. To a board, this looks like clean cost reduction. The most expensive-to-justify, lowest-margin layer of the firm is handed to a machine, and the numbers improve. Every incentive in the room points the same way.

The saving now, the cost later

This is the trap, stated precisely. The cost saved lands in this year’s accounts. It is visible, immediate, and sits on the balance sheet where everyone can see it. The consequence, an apex that goes unreplenished because the formation pathway has been removed, arrives years later. It is invisible, deferred, and sits off the balance sheet entirely.

The saving and the consequence sit on different pages, separated in time. That gap, between the page where the benefit is booked and the page where the cost eventually appears, is the trap. It is a textbook deferred liability that no accounting standard requires anyone to recognise, because the firm is quietly removing its own apprenticeship and calling it efficiency.

And every firm cuts at the same time

The trap gets worse when you step back from the single firm to the sector. A private cost decision, taken by one firm, is rational. The same decision, repeated across every firm in the sector at once, becomes a shared externality. The whole profession’s pipeline of future seniors drains together.

This matters because it removes the obvious escape. A single firm that hollowed its base could, in principle, hire experienced people from competitors to refill its apex later. But when every firm automates its base in the same window, there is no external pool of mid-career talent to draw from, because nobody’s base was forming anyone. No firm can simply hire its way out of a shortage that the whole sector manufactured simultaneously.

The measure: two velocities

To manage any of this, you have to be able to see it, and to see it you have to name it. The trap turns on a race between two rates.

The first is Automation Velocity, how fast the firm is removing routine base work. The second is Replenishment Velocity, how fast the firm re-forms judgement to replace what the base used to produce. The first is largely exogenous. It is a force from outside the firm, set by the pace of the technology and the market, and the firm cannot meaningfully slow it. The second is the dial the firm actually controls. Replenishment Velocity is the lever that management owns.

The gap between these two lines is the thing to watch. I call it the Succession Hollowing Index, and it is deliberately a transparent differential: Automation Velocity minus Replenishment Velocity. When automation outruns replenishment, the widening gap between the two lines is the firm hollowing out, measured. The value of naming it is the value of any named metric. Once it has a name, it can be put on a dashboard, owned, and managed against. An unnamed risk is one nobody is accountable for.

The escape: a synthetic rung

If automation removes the old apprenticeship, the firm cannot simply mourn it. It has to manufacture a new one. The base used to form people through sheer volume and time. With the volume gone, the firm needs a deliberately built substitute, a synthetic rung that does the formation job the base used to do as a by-product.

That synthetic rung has two halves, and both are necessary.

The first is cognitive. The firm has to rebuild judgement in compressed form. Where a junior once accumulated pattern recognition across years of routine volume, the synthetic rung substitutes simulation-dense exposure, concentrated reps against hard cases, deliberate practice on the judgement that the removed volume used to build slowly. The aim is to compress into a designed programme what the old base produced as a slow accident.

The second is relational, and it cannot be simulated. The graduate now enters above the machine. From the first day, the new entrant supervises, directs, and audits automated work rather than doing the routine work by hand. That is a fundamentally different posture, overseeing the system rather than being the system, and it has to be taught as a discipline of its own. You cannot simulate the relationship of accountable oversight. You have to put people into it and form them there.

Two outcomes, one dial

Here is where the whole argument resolves. Take two firms under identical automation pressure. The outcome is not determined by the pressure, which is the same for both. It is determined by the one dial each firm controls, its Replenishment Velocity.

The firm that builds the synthetic rung, that keeps replenishment in pace with automation, changes shape in a healthy way. The base thins because routine work is now supervised rather than performed. But the middle swells into the body of the firm, because that is where judgement is now made, in the compression layer where people are formed. The pyramid becomes a diamond: a narrow apex of trusted concierge-level judgement and relationships, a thick middle that is now the bulk of the firm and the place where people are made, and a thin supervisory base that oversees automated work. This is the good shape. The firm is still producing its own seniors, just through a redesigned mechanism.

The firm that does not build the synthetic rung keeps its apex and loses everything beneath it. Its apex stays intact for a while, but it is starving, because nothing is rising to replace it. Its middle hollows out into a broken bridge, the lost generation that never formed. Its base is automated away with no one entering. The pyramid becomes an hourglass: a top that survives on borrowed time, a pinched and broken middle, and a base that no longer feeds anything upward. Same automation pressure, opposite fate. The difference is the single dial the firm chose to set, or chose to ignore.

What this asks of leadership

The uncomfortable thing about the Pyramid Trap is that avoiding it requires acting against the immediate incentive. The saving is real and it is now. The cost is real and it is later, and it lands on a successor’s watch. Every quarterly pressure points toward automating the base and doing nothing about replenishment, because replenishment is an investment whose return is invisible for years and whose absence is invisible for years more.

The firms that come through this will be the ones that treat replenishment as a dial to be set deliberately rather than a by-product to be assumed. They will measure the gap, name the hollowing, and build the synthetic rung before the middle breaks rather than after. They will accept that the base changing shape is fine, and that the base disappearing without a replacement formation mechanism is not.

The pyramid is not the only stable shape. The diamond is too. But you only reach the diamond if you set the dial while you still have a middle left to thicken.

A note on provenance. The diagnostic in this piece, the Succession Hollowing Index as the differential between Automation Velocity and Replenishment Velocity, is a framework for naming and managing a real structural risk. The specific claim that designed, simulation-dense exposure can substitute for the judgement formerly built by years of routine volume is a working hypothesis, not a proven finding. It rests on a view about how expert judgement compiles that is still being instrumented. Treat the compressed-experience claim as the part of this argument most in need of evidence, and the structural logic of the trap, the deferred liability and the sector-wide externality, as the part that stands on its own.

The number that frames it is not a layoff count. It is a deconstruction output from reAImagination, the forecasting platform that prices the Application Support Engineer, the L1 and L2 support role that India staffed deeper than any country on earth, at 84 percent cost-weighted substitutable.

84% percent of the work that role is paid for lands on the automatable side of the line. That is not a role that shrinks. It is a role that dissolves, and the AI Agent does not sit beside the support engineer. It takes the support engineers seat.

I mention that not to claim authority over the future, which nobody has, but to say where the instinct comes from: I have watched the offshore decade arrive, and I have watched what happens when an industry mistakes a structural change for a cost line. This is one of those moments, and most of the commentary is looking at the wrong thing.

Here is the model change - For two decades the bargain in IT services was hours. You sold engineering time, billed it, and grew the firm by adding people to the pyramid. The pyramid had a shape everyone understood: a wide base of junior engineers on support and testing, a thicker middle of builders, a narrow top of architects and leads. The base was not just cheap labour, it was the training ground. It was where a graduate learned the estate, the codebase, the failure modes, the craft, before climbing. reAImagination’s forecast decomposes five IT-services roles into 53 sub-tasks and routes each one, and the verdict spread is the story: Application Support and Manual QA dissolve, the mid-level Developer and the Business Analyst redesign, the Platform and AI-delivery engineer expands. Two go, two change, one grows.

The roles that dissolve are the entry rungs. That is the whole point, and it is the line worth sitting with: the real risk is not AI replacing engineers, it is AI removing the bottom rung of the delivery pyramid, the rung every engineer used to start on. Automate it and you do not only remove those seats this year. You remove the pipeline that grows into the senior and the apex roles for the next decade. The saving shows up now. The missing pipeline shows up in five years, long after whoever approved the cost cut has banked the credit for it.

For a global system integrator the trap is sharper still, because the business model is selling the very hours the automation absorbs. This is not job displacement at the edge of the firm. It is business-model substitution at its centre.

You can already watch it arrive in the companies that build the agents.

GitLab cut roughly 350 people, about 14 percent of staff, and restructured openly for what it calls the agentic era, flattening up to three layers of management, exiting 22 countries, and embedding AI agents into its own reviews and approvals. Its chief executive put the thesis in one line: software built by machines, directed by people.

ServiceNow cut hundreds and said the quiet part out loud, attributing the cuts directly to the AI efficiencies its own platform generates inside its own business. It runs, in its words, the way it asks its customers to run. That is the tell. What a platform company does to its own org chart this year is a working demonstration of what it is selling every enterprise customer to do next year.

Now read it through the geography lens.

India holds the rung. The country runs 2,117 capability centres employing 2.36 million professionals, the densest concentration of enterprise process work anywhere, and more than 1,200 of those centres have already embedded AI. Of the roughly 1.9 million people in the estate, 55 percent are in roles exposed to automation, and up to 80 percent of routine operational work, L1 support and manual QA, is on the table. India staffed the exact seat the agent fills first. The net hiring numbers still look healthy, because capability centres added nearly 200,000 people last year, but that growth hides the rung disappearing underneath it. The seats being added are mid and senior. The seat being removed is the one people used to enter through.

The Gulf has something India does not: a policy lever. The UAE requires skilled-role Emiratisation rising 2 percent a year toward a 10 percent target, fines non-compliance at about 108,000 dirhams per unfilled role, and backs the Nafis programme with 24 billion dirhams of salary support and co-funded technical training. That turns inclusive redesign, the work of re-architecting a retained technical role so it is feasible for nationals and for People of Determination, from a corporate virtue into compliance economics. Where India is exposed to market forces at the dissolving rung, the Gulf has an instrument to fund a redesigned one. That is a genuine regional advantage, and it is underused.

Africa arrives at the rung just as it closes. The continent’s nascent capability-centre and BPO base enters the delivery chain at the same L1 and manual-QA rung the forecast routes to dissolution. The entry seat that grew careers elsewhere is shutting as Africa reaches for it. The opening, and it is a real one, is to skip the dissolving rung and build straight at the orchestration and governance layer the forecast shows expanding, rather than staffing the one the agent already holds.

If you are a board member, three questions sit between you and your next AGM, and they are not technology questions, they are pipeline questions.

First: do you know which rungs of your own delivery pyramid you are automating, and where the next generation of your senior talent will come from once those rungs are gone?

Second: do you have a named, accountable human for every automated decision your delivery now makes, in every regulated function?

Third: are you redesigning the work around the agent deliberately, or letting the agent decide by absence where the human no longer fits?

A firm that cannot answer the first is optimising itself out of a future it has not noticed it is spending.

If you are an individual, what you should do if you are in or near one of these seats. Three moves.

1. Stop defending the rung that is dissolving and build toward the seat the residue flows into, the orchestration and governance roles that direct the agents rather than compete with them.

2. Read value, not just exposure: the Business Analyst carries the highest value at stake in the IT-services forecast, about 11,791 dollars per seat, not because it is the most automatable but because its cost is high and a meaningful share of it is automatable, which is exactly the kind of thing a job-level view misses and only task-level costing surfaces.

3. And if you sit in the Gulf, learn to do the inclusive-redesign work that policy is paying for, because that seat is funded and favoured and most people have not noticed.

The agent took the bottom of the ladder. The work now is to build the rungs it cannot reach, and to do it deliberately, before the pyramid quietly loses its base and nobody can say when it happened.

This is the read reAImagine.work exists to provide, the only weekly magazine tracking AI and workforce intelligence with a sharp focus on the implications for the Global South.

The full Issue 013, ten screens, every figure sourced, is at reaimagine.work/issues/013. And if you sit on a board carrying these decisions, the Board AI Briefing we run with InGovern reads each week’s signal at the resolution reAImagination’s forecasts work in, sub-task by sub-task, at reaimagine.work/board-briefing.

Debu Mishra Founder, reAImagine.work

Image source: 24newshd.tv

The rehearsal is for replacement.

Last week I wrote that the compute was coming home, and that nobody had hired the people who would run it. The story was the top of the stack: sovereign supercomputers, Gulf capital, the operators and governors a corridor needs and cannot find. This week I want to go to the bottom of the same stack, because that is where the most revealing thing in the AI economy is happening, and almost nobody with a board seat is looking at it.

In a kitchen in Chennai, a 25 year old straps a smartphone to her forehead and films herself slicing mangoes. She is paid 250 rupees, about two and a half dollars, for an hour of that footage.

The recording is not a curiosity. It is invaluable to the companies teaching robots how to move like humans in a real home, because a humanoid cannot learn to grasp, fold and pour from internet video alone.

Simulations cannot model the physics accurately enough. So the industry buys first person footage of real people doing real chores, and India has quietly become the floor of that market. The robot she is training is being built to do the task she is performing. That is not a side effect of the arrangement. That is the arrangement.

I have always been curious - when a new category of work appears, I want to know who captures its value, who carries its risk, and whether it lifts the people doing it or merely passes through them. The egocentric data gig fails that test in a specific and fixable way, and that is what makes it worth your attention.

Start with what has actually changed.

The old data gig was the keyboard. You labelled images, you rated chatbot answers, you moderated content. Work that lived on a screen and paid by the task.

The new gig is the body. The worker is no longer an annotator, she is a performer, and the thing being captured is not her judgement but her muscle memory.

At an Objectways studio in Tamil Nadu, trainers record themselves in furnished fake apartments, folding the same towel ninety times a day from every position on the bed, until the wallpaper is changed to give the client more variety.

Objectways subcontracts to Qanat in Andhra Pradesh, which fits two thousand contributors with motion sensor bands on their wrists, hands and legs.

A Bengaluru startup, Human Archive, has put more than a thousand camera caps into Indian homes, hotels and restaurants on an 8.2 million dollar round backed by angels from OpenAI, Nvidia and Google. The pay band runs 250 to 400 rupees an hour. A digital labour researcher in Bengaluru expects the work only to grow.

Here is the problem - Venture investors put more than six billion dollars into humanoid robotics in 2025. Goldman Sachs sees a 38 billion dollar humanoid market by 2035 and as many as a hundred thousand units shipping this year.

Morgan Stanley projects more than a billion robots in use by 2050. Every one of them needs the human demonstrations being recorded in Chennai and Karur today.

The demand for that footage is highest in the exact window before the robot can do without it. So the recording boom and the displacement it funds are the same curve, read from opposite ends.

NITI Aayog has already named the stakes, and to its credit it named them before this gig made headlines. Its roadmap points out that the global AI conversation fixates on white collar jobs while ignoring India’s 490 million informal workers, who are nearly half of GDP, ninety percent of the workforce, with productivity barely five dollars an hour.

Its blunt forecast is that informal income stagnates at six thousand dollars by 2047 against a fourteen and a half thousand dollar target, unless something deliberate intervenes.

The egocentric gig is intervention’s awkward cousin. It puts cash in the hand today by selling the very thing that erases the task tomorrow.

Now the picture I cannot write about without all three regions, because the corridor is not only compute and capital, it is people, recorded, rated and ranked. India is the recording floor of the world. Africa is the same arbitrage on another shore, with Nigeria already on the recruiting maps alongside India and Argentina, and South Africa’s annotation rate sitting at its own fraction of the Western wage.

The Gulf enters from the opposite direction entirely. The UAE is now using AI and robotics to screen work permit applicants, deciding who may come and work, while its AI hiring climbs at double digit rates. So the region supplies the labour that trains the machine at one end, and deploys the machine to sort labour at the other. India and Nigeria hand over the gestures. The Gulf builds the gate. That is the shape of the thing, and it should make any board that buys this data or this capability sit up.

If you run a regulated organisation, three questions sit between you and your next AGM. Do you know where your training data came from, and can you prove it was consented? The first person footage that trains your robotics or spatial AI capability was recorded by someone, somewhere, under terms your organisation may never have seen, and that exposure lands on your balance sheet when a regulator asks.

Do you have a named accountable human for the data your models learn from, in every regulated function? And can you show, today, that the people recorded in your datasets retain the rights the EU AI Act and your local regulator now require, given that the Act’s high risk employment provisions land in August? These are not IT questions. They are capital allocation and accountability questions wearing a labour cost badge.

For the individual, the verdict is sharper and more hopeful than the headline suggests. If the robots are coming, the durable seat is not in front of the camera. The person paid 250 rupees to film a chore sells a perishable asset. Once the model learns the gesture, the demonstration has no second buyer. So move up the stack from performer to owner.

Own the pipeline, not the pose: the vendor who sources the trainers, cleans the footage and guarantees provenance sells the dataset for years, and Objectways, Qanat and Human Archive are hiring those supervisors and operations leads now.

Make consent the product: a vendor who can prove clean, consented, auditable data sells what the cheapest recorder cannot, and in a market racing on volume that is the defensible position. And bank the wage, build the skill: treat the hour in front of the camera as a stake, not a career, and convert it into the operations, quality and governance roles that sit around the recording and survive it.

The worker filming herself in Chennai is not naive about any of this. She says she may get a robot herself one day. The engineering graduate in the studio says the job is tolerable, but she always feels the camera. That clear eyed quality is the whole point. This is not a story about victims. It is a story about a price, and about who gets to set it. The egocentric gig can be a bridge that lifts informal income, or a treadmill that banks a few hundred rupees while training away the task. The difference is design, and the people who get to choose the design are the board that buys the data and the founder who brokers it.

This week’s full issue, with the careers data, the sector by sector read and the learning agenda, is at reaimagine.work/issues/012.

The chips are American. The capital is Emirati. The control is Indian.

For twenty years the Global South rented its compute from someone else’s jurisdiction. This week the machines started moving back. The job that runs them does not exist yet.

For two decades the bargain in our part of the world was rent. The work happened in Bengaluru and Hyderabad and Gurugram, but the servers, the cloud regions, and the law that governed the data sat in Virginia or Dublin or Singapore. We supplied the hands. Someone else owned the building. This week, quietly, on the sidelines of a summit, that arrangement started to invert.

I have spent thirty years watching how work moves across India, the Gulf, and Africa, and I have learned to be suspicious of announcements. Most of them are press releases looking for a headline.

This one is different, and the reason it is different is the part nobody is talking about: sovereign compute does not run itself.

Someone has to operate it, secure it, audit it, and answer for what it produces to a regulator who now sits in the same country as the machine. Those are jobs. And for the first time in the AI era they have a structural reason to sit in Riyadh or Hyderabad or Lagos rather than on the US West Coast.

The question that matters is never “what is the technology.” It is “who is accountable when it goes wrong, and where do they sit.” That question is about to land on every board in our region, and most of them have not seen it coming.

The model change

The old model was that you rented frontier compute by the hour and tried not to think too hard about whose laws governed your data until, suddenly, you had to. Data-localisation rules, sector regulators, and a hard geopolitical lesson about who controls the off switch turned that arrangement from convenient to risky. The shift is from renting foreign compute to hosting sovereign compute. And it has moved faster in four months than the layoff story moved in a year.

Krutrim - India’s first generative-AI unicorn stopped trying to build a frontier model and a chip, pivoted to a domestically built AI cloud, and turned its first profit: roughly 3 billion rupees in revenue, three times the prior year, at a margin above ten per cent. More than 25 enterprise customers in telecom, finance, and healthcare now run on it. Most of its GPU capacity is already committed to external workloads. The demand it is meeting has a single common requirement, and it is not price and it is not performance. It is: keep the data in India.

That is a sovereignty story wearing a technology badge.

The prize and the problem

The prize is enormous and it is regional.

Saudi Arabia’s Public Investment Fund has earmarked around 40 billion dollars for AI, semiconductors, and data infrastructure.

Stargate UAE is planned as a five-gigawatt campus in Abu Dhabi.

India is adding 20,000 GPUs to a national stack that already held 38,000, with the technology minister chasing more than 200 billion dollars of infrastructure investment inside two years.

The problem is that none of this runs on its own. The scarce input is no longer chips, and it is no longer capital. It is the human who can operate a sovereign cluster, govern its outputs, and prove to a regulator that the data never crossed a border. Those are not entry-level seats. They are exactly the roles the skilling programmes have not built, because everyone was busy training people for the seat the market is about to discount.

Gartner expects more than forty per cent of agentic AI projects to be cancelled by the end of 2027. The prompt-engineering job is the crowded one. The operator and the governance owner are the empty ones.

The IMEA Sovereign Compute Opportunity

The corridor has three nodes, and you cannot understand it by looking at any one of them.

India is the host. The supercomputer will sit here, under Indian governance, underpinning the India AI Mission. Krutrim has shown the commercial version works. The compute is arriving. The operators are not yet hired.

The Gulf is the capital and the silicon. The UAE is not only hiring AI talent, where demand jumped from 32 to 48 per cent in a year, it is exporting capital and chips into India’s sovereign build while keeping its own. The Saudi National Data Center Strategy targets 1.5 gigawatts by 2030. The Gulf funds India’s compute and competes for the same scarce operators, which means the shortage shows up on both shores of the Arabian Sea at once.

Africa is the third node, building anyway. Cassava launched the continent’s first NVIDIA-powered AI factory in South Africa and is extending GPU-as-a-Service to Nigeria, Kenya, Egypt, and Morocco. Last week the AI Skills and Compute Africa Foundation launched in Kigali with a target of one million young people into AI-value-chain work, and a promise that the data stays on the continent. Africa holds about 0.6 per cent of global data-centre capacity. It is making the same sovereignty argument with a fraction of the power. Power, not ambition, is the binding constraint there, and that is the honest watch-out.

What the boardroom should decide

If you sit on a board in our region, the question is no longer whether to adopt AI. It is whether you own the compute your most sensitive workloads run on, or rent it from a jurisdiction you do not control. That is a capital-allocation decision and an accountability decision in the same breath. Three questions belong on the agenda before your next annual meeting.

• Do you own the compute your sensitive workloads run on, or rent it from a jurisdiction you cannot govern?

• Do you have a named, accountable human for AI outputs in every regulated function, or a committee that will point at each other when a regulator calls?

• And can you prove, today, to that regulator, that your data never left its jurisdiction?

If you cannot answer all three cleanly, you have work to do this quarter, not this fiscal year.

The individual’s verdict

And if you are reading this as someone building a career rather than chairing a board, the same fact points at three moves.

1. Own the cluster, not the prompt. The operator who runs a sovereign supercomputer and the governor who answers for its outputs are the seats G42-Cerebras, Krutrim, and Stargate UAE all need and none can fill.

2. Learn the residency rule, not just the model. The whole point of the corridor is that data stays inside the jurisdiction. The compliance lead who can prove that to a regulator is worth more than the engineer who cannot.

3. And if you are entrepreneurial, sell sovereignty as a service. Krutrim has already proved the unit economics. African delegations came to the India summit asking how to build their own.

Package Gulf capital, Indian engineering, and local data residency, and you have a business that did not exist two years ago.

This is what reAImagine.work exists to do: track AI and the workforce with a sharp focus on what it means for the Global South, every week, from inside the region rather than from a commentary booth in San Francisco. The compute is coming home. The capital is flowing in. The buildings are going up in Riyadh and Casablanca and Hyderabad.

The only thing missing is the people who will run them. That is the gap and the opportunity.

This weeks cover art

This week, in Bengaluru, the managing director of IBM India and South Asia, Sandip Patel, told Reuters that only about thirty per cent of India’s technology workforce has the AI skills that buyers now demand. Around 200 million people are AI-capable today. The country needs 350 million by 2030 to hold its place. The gap is 150 million workers.

That is the number this week.

THE MODEL CHANGE, NOT THE LAYOFF

The Western frame on India this year has been Tata Consultancy Services cutting twelve thousand and Wipro deploying two hundred AI agents in HR, finance and legal. The frame is accurate but it is also too small. India is not being trimmed at the edges. The whole proposition for India is changing.

For two decades the deal was simple. The same compliance work, the same support engineering, the same finance close, done in Bengaluru or Hyderabad for a fraction of the home-market cost, at quality, in a useful time zone. 2,100+ plus global capability centres with 2.36 million people and nearly 100 billion dollars in revenue. The world’s largest GCC hub, built on a single proposition: scale at a discount.

This week at the Reuters summit in Bengaluru, the executives running those centres said the discount is closing. JPMorgan, Walmart, Target, IBM, Citi are not vendors anymore. They are integrated hubs, judged like the parent and paid like the parent. Target’s India lead called the fight for talent unreal.

Wage inflation is running ahead of AI and machine learning demand, which is itself running ahead of supply.

The seat that priced itself on being available is the seat AI takes first. The seat that prices itself on being right is the seat the parent now bids up. The hub is re-pricing.

That is a different story from layoffs.

THE PRIZE AND THE PROBLEM

A joint study from IBM’s Institute for Business Value and IndiaAI puts AI’s contribution to the Indian economy at more than $500 billion by 2030. That number only lands if the skilling happens. ServiceNow and Pearson put a separate number on the same shift: 10.35 million Indian roles redefined by 2030, more than 3 million new tech jobs created. Manufacturing carries the largest impact at 8 million, retail at 7.6 million, education at 2.5 million.

Patel said the skilling will not happen by accident.

IBM itself pledged in December to skill 5 million Indians in AI, cybersecurity and quantum by 2030, and has been expanding its footprint outside Bengaluru and Hyderabad: Kochi to 4,000 staff in two years, a new presence in Lucknow.

The gap is not a hiring problem yet. It is a teaching problem.

THE GULF IS THE RELEASE VALVE, ALSO THE COMPETITOR

The other half of the IMEA picture this week is the Gulf. UAE AI-related hiring rose to 48% from 32 in a single year, the strongest growth across any market RemotePass tracks. Saudi Arabia is targeting 150,000 new tech seats under Vision 2030 with $11.2 billion in digital infrastructure. The named buyers -- G42, Careem, Noon, ADNOC Digital -- are not back-office GCCs. They are advanced-AI hiring centres.

For an Indian worker whose process seat is being re-priced out of existence, the Gulf is a release valve. For India the country, the Gulf is a competitor that is bidding away the AI talent India needs to close its own 150 million gap. The same shortage shows up on both sides of the Arabian Sea.

Africa sits downstream of both. African-language AI evaluation in Swahili, Hausa, Yoruba, isiZulu and Amharic remains a demand line outpacing supply. Egypt is already a named talent exporter into Gulf projects. The continent’s skilling programmes are racing the same automation that is re-pricing India’s hubs.

THE BOARDROOM DECISION

The decision to run a capability centre instead of a cost centre is not an HR memo. It is a capital allocation choice made in the boardroom of the parent and the captive alike. Three questions sit between every board with GCC exposure and the next AGM.

One. Which of your functions are still priced as arbitrage that AI takes first? Be honest. If the output rolls up only as team productivity or cycle time reduction, the role is in bin one whatever the title on the door says.

Two. Do you have a named accountable human for AI outputs in every regulated function? Every major regulator -- the FCA, the EBA, MAS, HKMA, RBI -- is now writing rules that require named humans to own AI outputs. Boards have not appointed these humans yet. The seat is open.

Three. Do you have an actual reskilling commitment in your annual report, with budget and timeline, or do you have a paragraph? IBM’s five-million pledge is a real number with a real horizon. Most annual reports are not.

PRICE YOUR SKILL, NOT YOUR SEAT

For the individual reader, the verdict is shorter than the macro story. Three things to do this quarter before the appraisal cycle reads your role as cost rather than capability.

Move your work from the cost line to the capability line. If your output is still described as support, process or coverage, it is priced as arbitrage. Re-describe one deliverable this week as an outcome with a number attached. Test whether the number survives a CFO reading it.

Close one named AI skill gap before the next review. Pick the single skill your GCC’s parent is actively hiring for. Model governance. Cloud orchestration. LLM deployment. Finish a real credential with a deployed artefact behind it. GCC buyers discount inflated AI CVs. They reward one production deployment.

Follow the capability money. The hiring is not where the layoffs are. It is in UAE AI roles up to forty-eight per cent. In Saudi Arabia’s 150,000 new tech seats. In IBM’s Kochi and Lucknow. Apply where capability is bid up, not where cost is cut down.

TWO STORIES IN OPPOSITE DIRECTIONS

Last week a Western bank told investors it was replacing humans with capital. This week the world’s largest back office said the discount that built it is finished, and only the capability model pays. Same fortnight. Opposite halves of the same story.

I have seen Board Governance challenges before. In my opinion this is the one that will define the next decade for every IMEA company.

Not whether AI replaces humans. How fast the humans climb to where the capital still bids them up.

The full Issue 010 is at reaimagine.work/issues/010.

Art form this issue: Expressionism - Germany and Austria, early 20th century. Kirchner, Schiele, Kandinsky. The visible mark carries the feeling. Distortion is not a flaw but the point. The brushstroke is the subject, not the means.

Cover seeded uniquely for every reader. The hue, density and composition are deterministic to the name. Different name, different cover, new art form every Friday.

On Tuesday, in front of his investors, Bill Winters said the bank would invest in AI “instead of the human capital.”

Not alongside. Not to augment. Instead of.

Standard Chartered will cut 7,000 roles by 2030, roughly 15 per cent of a 51,000-strong support base. Wealth management, which posted a record Q1 2026, gets the capital that leaves the back office. The footprint that absorbs the decision -- Asia, Africa, Middle East -- is the same footprint this newsletter covers. That is not a coincidence. It is the whole story.

Issue 009 of reAImagine.work runs the data. This piece is the slower read underneath it: what the signal actually means if you sit in Bengaluru, Dubai, Johannesburg, or any of the support hubs the FTSE 100 outsourced its back office into.

For three years, the corporate script on AI and work has been augmentation

Every CEO note, every earnings call, every HR town hall has reached for the same phrase: AI will free our people to do higher-value work. The phrase did real work. It bought time with regulators, with employees, with boards that did not want to read a headline about layoffs.

Winters dropped it. He named the substitution out loud, in the room where the script is supposed to be most disciplined. That is the part to pay attention to. Not the 7,000. The willingness, in 2026, to stop pretending.

When one regulated lender in the FTSE 100 says it in front of investors, the rest of the sector does not need to repeat the sentence. The sentence has been said. The cover is gone. The next earnings season will be full of carefully phrased variations -- “rebalancing the cost base”, “structural productivity”, “shifting the human capital mix” -- but the underlying decision is the one Winters named.

If you are a banker, an IT services partner, a captive operations lead in Bengaluru, a PRO in Dubai, a shared-services head in Nairobi, the question is not whether your CEO will say the same thing. It is when.

What replaces the human, and what does not

Read the Standard Chartered plan carefully and the architecture is not “AI replaces people”. It is more precise than that.

What replaces the human in compliance screening, KYC remediation, reconciliations, exception handling -- the high-volume, rule-rich, paper-trail work -- is the AI system. What replaces the human in wealth advisory, corporate coverage, sustainable finance, private banking for the GCC’s HNI and India’s mass-affluent, is a more senior human sitting closer to the client, paid more.

Three sorting bins, in other words:

1. Automatable back office. Cleared by the model, supervised by a small, accountable human team.

2. Redeployable into wealth or corporate coverage. A meaningful share of the 7,000, if the bank is honest about it.

3. Upmarket and hiring. Where the capital lands.

Every role in an IMEA support hub belongs to one of those bins today. The honest exercise this week is to name which bin holds your work, and which bin holds your work in 2028. That is not a career-coach question. It is a planning question. The Standard Chartered timeline is six months telegraphed; the hiring into bin three is happening now.

The IMEA footprint is the lever, not a side note

Here is the part Western coverage will miss.

The 7,000 cuts will not land evenly.

Standard Chartered’s support base is concentrated in Chennai, Bengaluru, Kuala Lumpur, Warsaw and Johannesburg.

The wealth-advisory hiring will concentrate in Dubai (DIFC), Singapore, Hong Kong, and Mumbai’s GIFT City.

Both ends of the substitution sit in the IMEA region.

Three regional reads worth holding in mind:

India. The countervailing fact is real -- Nasscom records the IT workforce growing net 1.4 lakh to 59 lakh -- but TCS has 12,200 going in FY26, Wipro has deployed over 200 AI agents in HR, finance and legal, and the back-office work for Standard Chartered, Citi, HSBC and Barclays sits in the same four cities the IT majors hire from. Whether the destination roles absorb the displaced volume from process-driven seats is the question no one in Bengaluru can yet answer. The mix shift is the story; the net number is the comfort blanket.

Middle East. The UAE MoHRE “Eye” platform began screening work-permit applicants through agentic AI this month. The state is now the gatekeeper that decides whose documentation reads cleanly to a model. Saudi Arabia has declared 2026 the Year of AI under MBS, who chairs SDAIA. And Dubai will absorb more of Standard Chartered’s wealth-advisory hiring than any other regional centre. The Gulf is on both sides of the substitution at the same time: harder to enter as a worker, easier to staff as an employer.

Africa. Kenya’s AI Skilling Initiative has trained over 600,000 people. African-language prompt engineering -- Swahili, Hausa, Yoruba, isiZulu, Amharic -- is the AI-supply gap Andela and Gebeya are pricing in USD. And Standard Chartered’s 15 per cent cut applied to support functions globally will land in Johannesburg and Nairobi too. Skilling and displacement cross at the same desks. That contradiction is not a sequencing problem the continent can solve in five years.

What to actually do this quarter

Three moves, before the Q2 results window opens and the next CEO uses Winters’ phrasing.

One. Stop defending the job description, start showing the capital return. Winters did not say “we are cutting cost”. He said “instead of the human capital”. Roles survive only when they can be expressed as numbers on the same line investors read. If you cannot write the capital-return number on your own role -- in basis points, in incremental fee income, in cost-to-serve reduction, in risk-weighted assets freed -- someone else will write it for you, and it will be the wrong number.

Two. Map your function against the three sorting bins, honestly. Not the bin your team’s deck says you are in. The bin you are actually in. The work that clears the model in 18 months is the work that has rule-rich, high-volume, low-judgement characteristics today. Defending that work as judgement-rich does not change the model’s view of it. Naming it accurately gives you 24 months to move; defending it costs you the 24 months.

Three. Watch the wealth and corporate ladder, not the layoff list. The 7,000 cuts are six months telegraphed and the press will catch up. The hiring is happening now in private banking in Dubai, in corporate coverage in Bengaluru, in sustainable finance in Johannesburg, in model validation across all three. Move before the announcement, not after. The seats that fill quietly between now and Q3 are the seats that pay for the next decade.

The question for the board underneath all of this

The reason Issue 009 ends at the board, and the reason this piece does too, is that Standard Chartered’s plan was announced in an investor presentation, not by HR. The decision sat with the board before the workforce ever saw a memo.

If you sit on a board, three questions sit between you and your next AGM.

1. Do you know which of your functions belong in the bin Standard Chartered called automatable?

2. Do you have a named, accountable human for AI outputs in every regulated function?

3. Do you have a reskilling commitment in your annual report with actual budget and timeline, or do you have a paragraph?

Most boards in the IMEA region have a paragraph. By the time they have an answer, the bank down the road will already have hired the wealth advisor and named the AI supervision officer.

That is the substitution Winters was actually announcing. It is happening on both sides of the trade, and it is happening in our region first.

reAImagine.work is the only weekly magazine tracking AI and workforce intelligence with a sharp focus on the implications for the Global South. Read Issue 009 in full at reaimagine.work/issues/009. The Board AI Briefing is a reAImagine.work x InGovern programme datelined Bengaluru and Dubai.

Google retired Vertex AI as a standalone product and shipped its successor, the Gemini Enterprise Agent Platform. The name is corporate. The architectural shift it represents is not.

This sounds like a technology story. It is not. It is the story of how the captive offshore model, the consulting pyramid, and large parts of the C-suite agenda get redesigned in the next eight quarters.

Four signals from this week’s research for reAImagine.work deserve attention here on Leading the Flux too, because the framing matters and you are most likely to be on the receiving end of these decisions of your own organisations.

Signal one: the architecture itself

The Gemini Enterprise Agent Platform combines model selection, model building, agent building, orchestration, DevOps, governance, optimisation, and security in a single control plane. Partner-built agents from Adobe, Atlassian, Teradata, Typeface, and UKG are already in the Agent Gallery. UKG alone is moving time, pay, scheduling, and HR service delivery into the platform for the 80,000 client organisations it serves across 150 countries. The pricing model is shifting from cost-per-prompt to cost-per-completed-task, which breaks every commercial structure in the global services industry that has been anchored on human time since 1990.

The reason this matters now, rather than in 2027, is that the Agent Platform turns what used to be a transformation programme into a procurement decision. Until last week, when a CFO asked whether the finance and accounting tower could be collapsed from 1,400 people to 200 supervisors plus an agent fleet, the answer was “in theory, on a roadmap, with the following risks”. This week the answer is “Google sells the platform, the agents are in the gallery, your procurement team can issue the purchase order”. That is the change. The technology has been visible for eighteen months. The product is shipped now.

Signal two: the boards have already internalised it

KPMG International published its Global AI Pulse on 7 May, drawing on responses from 2,110 C-suite leaders across 20 countries collected between 19 February and 27 March. The headline finding is structural, not numerical: business leaders are now designing roles, teams, and workflows on the explicit assumption that humans and agents will work together, with agents handling research and coordination, and people focusing on judgement, decision-making, and accountability.

Stephanie Terrill, KPMG Canada’s Managing Partner of Digital and Transformation, summarised the shift in one sentence: “Everyone now needs to know how to allocate work, assess quality and give effective feedback, because that’s what ultimately unlocks value.”

That is a quiet sentence. Read it twice. What she is describing is the end of the management framework that has dominated enterprise organisation design since the 1990s. The CHRO function I helped shape inside Deloitte, the capability frameworks I sold across South Asia as part of Bersin’s work, the entire architecture of “manage your people, develop your people, retain your people” was built on the assumption that the people were the unit of production. The KPMG data confirms that boards have stopped making that assumption. Not in a strategy paper. In their 2026 operating models.

Signal three: the layoff cycle is now reflexive

On 7 May, the same day the KPMG findings landed, five public companies announced AI-attributed workforce reductions in a single news cycle. Cloudflare cut 1,100 (20 per cent of its workforce), citing internal AI usage that had grown more than 600 per cent in three months. Upwork cut 150 (24 per cent), citing “the evolving nature of work”. DeepL cut 250 (25 per cent). Bill Holdings flagged up to 30 per cent reductions. Coinbase confirmed 700 cuts (14 per cent), framed by Brian Armstrong as a structural shift toward smaller, AI-augmented teams.

What makes the cluster significant is when it landed. The Gartner finding from the previous week’s research, that 80 per cent of AI-deploying enterprises cut staff with no measurable ROI correlation, was front-page material in the same 24-hour window. The layoff announcements proceeded anyway. The Cloudflare CEO did not pause to evaluate whether 600 per cent internal AI usage was producing measurable output; he announced a 20 per cent reduction in the same memo.

This is the part that I find difficult to write about without sounding either alarmist or resigned, and I have been trying to find the honest register for it. The honest register is this: the evidence has stopped slowing the cycle. The architecture from signal one will accelerate it. The operating-model redesign from signal two will rationalise it. There is no remaining check inside the corporate decision chain.

Anyone planning a 2026 to 2028 workforce strategy on the assumption that the layoff-ROI evidence will trigger a pause is planning on the wrong assumption.

Signal four: 97 per cent of the C-suite says parts of their own role are next

The agentic-AI infrastructure firm Gravitee surveyed 250 C-suite executives at large UK companies. Every respondent planned to deploy AI agents inside their organisation by end of 2026, with most projecting between 16 and 20 agents per company. The aggregate projection: roughly 100,000 virtual AI workers in the UK alone. Sixty-five per cent of the same firms confirmed headcount reductions over the same window. Forty-six per cent estimated more than half the work of a software engineer, HR manager, or customer service representative could be done by an agent.

The number that is worth holding is 97 per cent. Ninety-seven per cent of these executives said parts of their own job were agent-eligible.

C-suite executives are not, as a population, in the habit of describing their own role as automatable. The social and psychological costs of that disclosure are real, and the typical response is to describe automation as something that happens to other people’s work. That nearly every respondent broke that pattern suggests the agent architecture has crossed a perception threshold inside boardrooms. The CEO who in 2024 framed agents as a productivity layer for the rest of the company is in 2026 acknowledging that the CFO function, the COO function, parts of the CHRO function, and large parts of strategy and corporate development are now within scope.

Once the C-suite has internalised this, the rationalisation cycle for the layers below them accelerates by a factor that the productivity literature has not yet modelled. This is the dynamic that will define the next eight quarters, and it is the dynamic that the Western business press is still treating as a 2028 question.

What the Global South lens shows that the rest of the press is missing

India runs the world’s largest captive offshore workforce. Roughly 2.4 million people work inside the 2,000-plus Global Capability Centres operating across the country, up from 1.9 million in 2024. Zinnov and Indiaspora published an analysis in March 2026 placing 55 per cent of this work in the two lowest tiers of their four-tier classification: Commodities and Procedures. Both tiers are direct candidates for the agent fleets the Gemini Enterprise Agent Platform now operationalises.

I spoke with six current GCC heads over the past fortnight for the Conversation piece in this week’s Issue 008. Five of the six are running the same playbook: pause Commodities-tier hiring, accelerate the recruitment of senior people from the parent’s consulting roster and from McKinsey and BCG, lobby the parent for an extended runway. None of them are confident the runway is long enough. The sixth has already begun a phased reduction. The next eight quarters will decide which playbook the rest of the industry follows.

The Gulf is running the opposite strategy. Saudi Arabia formally designated 2026 as the Year of AI in March. The Hexagon Data Centre (480MW, the world’s largest government data centre by capacity) broke ground in January. Humain, the Public Investment Fund’s wholly-owned AI vehicle, is operationally scaling with a $9.1 billion investment pipeline. The UAE has the Stargate cluster, nationwide ChatGPT Plus access, and the DIFC’s AI governance framework.

The Gulf is the only region in the world where workforce strategy in 2026 is a story of expansion rather than reduction.

Africa sits at the intersection of two timelines that are colliding sooner than the development literature expected. The Overseas Development Institute’s November 2025 study estimated 2.5 million Kenyan workers in roles with high or significant generative-AI exposure. The Kendall and Mishra Nature analysis from April argues there is still a credible path through human-AI complementarity in AI-native services where human judgement remains essential. The path is narrow.

What to do with this

You are most likely advising or running organisations that will make decisions about agentic deployment in the next four quarters. Three things matter:

First, audit your largest workforce concentrations against the Zinnov classification this quarter. If more than half sits in Commodities and Procedures, the budget call is already on the calendar; you are just not on the invite list yet.

Second, open a parallel workforce design track this quarter for agent-inclusive operating models. If your 2026 workforce plan still treats AI as a productivity layer rather than as a category of worker, the plan is already obsolete.

Third, stop arguing that the layoff-ROI evidence will slow the cycle. It will not. Plan as if every quarter brings another five-name cluster, because every quarter is bringing another five-name cluster.

The full Issue 008, with the Sector Lens cutting these four signals across nine industries, the Region Radar reading for India / GCC / Africa, the long-form Conversation with the composite GCC head, and the full Hot Skills Forecast (five appreciating, three depreciating, two emergent roles that do not yet exist on org charts), is live now at reaimagine.work/issues/8.

If you are advising boards or running workforce functions in India, the GCC, or Africa, this is the weekly read that covers the signals your current publications are missing. Subscribe free at reaimagine.work. No paywall. No fluff. New issue every Friday.

This week’s cover is Impressionist.

Get your own unique version when you type in your name.

I am writing this from Dubai, the morning after the Gartner finding broke. I have been awake since four. The reason is professional, not personal: a piece of data published this week has falsified a playbook I spent the better part of two decades selling.

That is not a comfortable sentence to write. It is, however, the only honest opening to this essay.

Here is the playbook. A consulting firm walks into a CEO’s office with a slide deck. The deck has a chart of operating margin on the y-axis and time on the x-axis. The chart shows two lines. The first line is the status quo. The second line is what happens after a transformation programme. The second line is always above the first, and the gap between the two lines is labelled “value creation.” Underneath the chart, in smaller type, sits a row of assumptions. One of those assumptions, almost always, is a headcount number. Reduce people by N per cent. Reinvest some, return some, declare ROI.

For most of the period from 2005 to 2020, that playbook worked well enough. The headcount line was usually small enough to be absorbed into natural attrition, and the productivity gains in the rest of the model - process redesign, technology investment, structural simplification - were real. The transformation worked. The headcount line was a footnote, not a feature.

Then generative AI arrived, and somewhere around late 2023, the headcount line stopped being a footnote. It became the feature. The deck did not change. The chart still showed the same two lines. But the assumption underneath, the one in smaller type, started to do most of the work in the ROI calculation. Cut twenty per cent of customer service. Cut fifteen percent of compliance. Cut ten percent of middle management. The savings would compound. The board would approve. The CFO would sign. The press release would be written.

On 5 May, Gartner told us how that played out across 350 large enterprises.

Eighty per cent of those enterprises cut staff. The cuts did not generate ROI. The companies that improved their AI returns were the ones that did not cut, or did not cut as much, and instead reinvested aggressively in skills, roles, and operating models around the people they kept. The amplification strategy outperformed the elimination strategy on the metric that the elimination strategy was sold on.

The receipt has arrived.

I want to say something carefully here, because the temptation when a finding like this lands is to either defend the prior position or distance oneself from it. I am not going to do either. The transformation playbook was, for a long time, a useful framework. It produced real value for clients who used it well. It produced overconfidence and reductive thinking in clients who used it badly. The same is true of the AI-as-cost-takeout playbook now. It will produce real value for organisations that use it carefully. It is producing overconfidence and reductive thinking in organisations that have used it as cover for cuts they were already going to make.

The Gartner data is the cleanest evidence we have ever had that the second category outnumbers the first by a factor of about four to one.

So what should you actually be doing? Three things, in this order.

• First, separate the cost case from the capability case in every AI investment proposal that crosses your desk. If the proposal cannot survive without the cost line, it is not an AI investment. It is a layoff with a technology fig leaf. Send it back. The proposals that survive without the cost line are the ones that will compound.

• Second, redesign your performance management cycle for the AI era before you redesign your org chart. The Microsoft Work Trend Index, also published this week, found that only thirteen per cent of AI users say they are rewarded for experimenting with AI in their jobs. Thirteen per cent. The binding constraint on AI value is not technology and not workforce. It is the organisation around them. You cannot fix that with a technology investment or a training budget. You can fix it with a change to how you measure and reward people. That change costs nothing and unlocks more value than any agentic AI deployment your CTO is currently considering.

• Third, audit your fresher pipeline. Cognizant signalled this week that Project Leap will absorb 12,000 to 15,000 layoffs, with the majority falling in India. The Indian IT services pyramid that fed the country’s middle class for two and a half decades is being dismantled by its own architects. If your organisation has been quietly absorbing the AI-era pressure by reducing campus hires rather than redesigning entry-level work, you are storing up a problem that will surface in two to three years when your mid-level talent pipeline runs dry. Reduce campus hiring by all means, but only if you have explicitly redesigned the work that the remaining freshers will do. Otherwise you are not transforming. You are deferring.

I will close with the part of this essay that is most uncomfortable for me to write, because it implicates my own former practice and the practices of firms I sold against, sold with, and now critique.

The transformation industry, including the human capital advisory practices inside the Big Four and the strategy firms, has spent the last three years pitching AI-led workforce reduction as if the ROI was a settled question. It was not a settled question. It was a hypothesis. The Gartner finding is the first large-sample test of that hypothesis. The hypothesis failed. Any consulting firm that is still pitching AI-led cost takeout as the primary value creation mechanism after this week is either not reading the data or is reading it and choosing to ignore it. Both are problems.

If you are a CEO or a CHRO who took the elimination playbook to your board in 2024 or 2025, you now have new evidence to bring back. The board will not enjoy the conversation. Have it anyway. The alternative - wait for someone else to surface the finding for you - is the worse outcome by a wide margin.

I sold the transformation playbook for two decades. The honest move from here is to recalibrate, in public, in front of the clients who deserved more careful work.

This week’s issue of reAImagine.work is my recalibration.

This essay is the anchor of Issue 007 of reAImagine.work, published this morning. The full issue includes the Sensing Report behind the Gartner finding, a nine-sector cut showing where the confession lands hardest, and a regional read across India, the GCC, and Africa. Free, every Friday.

If this resonated, the magazine itself is where the rest of the work lives. Reply to this email or comment below - what is on your AI investment proposal queue this week, and which line item would not survive Gartner’s test?

- Debu

Eleven predictions made across five issues. Today, with Issue 006 going to press, we graded them.

The headline: an A on the AI-augmentation narrative collapse, an A on the Indian IT services structural reset, an A on Oracle’s 30,000 layoffs being credible, an A on Saudi 2026 AI investment crossing the trillion-dollar mark, an A on the GCC capability gap. A B+ on white-collar entry-level compression - IBM tripling entry-level hiring is the counter-data point, and the picture is more uneven than we framed it. Half-credit on nuance.

That’s the report card. What’s harder is the lesson.

The predictions we got right share a pattern. Each named a precedent that hadn’t yet happened publicly but was forming in the books of organisations under pressure. We didn’t predict that AI-augmentation framing would die - we predicted that earnings season would force its abandonment, because the financial logic of “we hired a thousand engineers and saved a billion dollars” doesn’t survive a quarterly call. We didn’t predict that Saudi Arabia would commit a trillion dollars to AI - we predicted that the existing $600B trajectory plus the regional capability anxiety would force the public number upward when it became politically necessary. The signal each time was structural, not narrative.

The B+ tells the inverse story. We named white-collar entry-level roles as the first compression zone because consulting and law and finance all run on a pyramid that AI flattens. What we missed: the labour-shortage pressure in some sectors and geographies cuts the other way. IBM tripling its entry-level hiring isn’t a counter-trend, it’s a different equation - the pyramid shape still flattens, but the absolute headcount can grow if the firm is scaling. The signal we read was real. The framing was too clean.

This is what Leading the Flux is for: the place where we sit with the shape of the signals, not just the numbers.

Issue 006 hits this morning with five new signals.

A 96-hour window in which Microsoft, Google, OpenAI, and Adobe flipped agents from optional to default across Word, Excel, search, and the application layer.

Five trading days later, the layoff letters caught up.

Cognizant Project Leap booked AI-led restructuring as a P&L line item - a $200-320M severance programme, $200-300M annual savings, with the rationale stated openly. TCS already shed 23,460 roles in FY26. The WITCH tier is sizing equivalents.

A Bengaluru GCC site head pulled the list of permanent staff with 15+ years tenure. 412 names. She knows what’s on Cognizant’s slide deck.

And six US banks booked $47 billion in Q1 profits while announcing 15,000 cuts in the same quarter. The compression is in the VP-to-MD rungs - exactly where pitchbook automation, controls review, and proxy advisory now operate.

These are not predictions. These are receipts.

The next predictions, then.

The Cognizant template will replicate across the WITCH tier within two quarters. The bank cuts will spread from VP-to-MD into the analyst pyramid below them by Q3. The agent-default flip will normalise within the application layer by year-end, and the workforce policy implications - audit trails, license allocations, data governance - will be the next compliance fight, not whether to deploy.

Subscribers get a new prediction set every Friday. Issue 007 is taking shape now.

Read the issue: reaimagine.work

This week, the C-suite answered it themselves.

WRITER and Workplace Intelligence surveyed 1,200 chief executives and 1,200 employees for their 2026 AI Adoption in the Enterprise report. The findings are not subtle.

Seventy-five percent of executives say their AI strategy is more for show than actual internal guidance. Fifty-four percent say AI adoption is “tearing their company apart.” Forty-eight percent call it a massive disappointment. The word “theatre” appears eleven times in the open responses.

These are not the numbers of a technology problem. They are the numbers of a strategy architecture problem - specifically, a strategy that lives in the top layer (vision, positioning, AI council, town halls, dashboard) and never reaches the bottom layer (workflow, cost per outcome, accountability structure, decision rights).

The minority who are making it work - the 29 percent reporting significant ROI - all started at the workflow layer and built up. Not the other way around.

Stanford confirmed early-career displacement is accelerating.

Entry-level software developer employment among 22 to 25 year olds has dropped nearly 20 percent since 2024. This is not a recession signal. Broad employment indicators are stable. This is a structural signal: the entry-level bench is gone before the automated work at the top matures.

The implication for organisations in India, the GCC, and Africa - where graduate hiring pipelines are the primary talent acquisition model - is severe and under-discussed.

Baker McKenzie became the first global law firm to name AI in a mass layoff.

Six hundred to one thousand roles. Explicit AI attribution. No hedging.

The law firm market has been slower than most to publicly acknowledge AI-driven workforce change. Baker McKenzie has changed that. The Big Four professional services firms will follow within two quarters. The only open question is whether they will be as direct about the reason.

[INSERT twitter-issue-005.html image here]

What is new in the experience this week.

reAImagine.work has a new design and the personalised cover is live.

Every subscriber gets a unique generative artwork computed from two inputs: the signal data from that week’s Sensing Report, and the ASCII values of their own name. The artwork is deterministic - the same name and the same issue always produce the same cover. Every week uses a different art historical style. Issue 005 renders in Mannerism. Issue 008 will render in Impressionism. Issue 013 in Constructivism.

The world signal map is live - twelve cities, real signals from the reAImagine archive.

The Co-reader shows how executives are engaging with the intelligence in real conversations.

Everything is free. Every Friday. Go to reaimagine.work, tap the masthead, enter your name, and claim your cover.

As a Deloitte Human Capital Partner, I contributed to the Global Human Capital Trends report. I led the Bersin by Deloitte practice across South Asia. I sat in the briefings, reviewed the frameworks, and sold the thinking to CHROs across the region.

I believed in the predictions. Most of us did.

Five years later, I run reAImagine.work - a weekly workforce intelligence publication focused on AI and the world of work, through the lens of India, the GCC, and Africa. And this week, in Issue 003, I published something that has been forming in my mind for the last two years: a formal report card on the predictions made by the World Economic Forum, Josh Bersin, Deloitte Human Capital Trends, and Dave Ulrich.

The grades are not kind. But the evidence is clear.

The verdict at a glance

The WEF’s 2020 Future of Jobs Report predicted 85 million jobs displaced and 97 million new ones created by 2025. By their own 2025 update, those goalposts had moved to 2030. The specific displacement mechanism they modelled - manufacturing and routine physical roles - was wrong. The actual displacement has been in junior knowledge work, analytical synthesis, and document research. The wrong jobs, the wrong timeline, the wrong geography.

Josh Bersin declared a new war for talent in 2021. Within 24 months he was writing “Yes, AI is really impacting the job market.” Oracle sent 30,000 termination emails at 6am on a Tuesday in March 2026. The company had just reported a 95% jump in net income. The war ended. AI won.

Deloitte’s own 2026 Human Capital Trends report - published just last month - states that only 6% of leaders are making progress designing human-AI interactions, and 65% say their culture needs to change significantly because of AI. Their own data indicts the optimism of their earlier editions.

Dave Ulrich’s principles held better than the others. But even his values-not-place framework did not anticipate that the boundary-free work environment he described approvingly would be exploited by deepfake job candidates and organised laptop farm operations.

What every prediction missed

The Global South was invisible. India, the GCC, Africa - together representing billions of workers with the highest AI exposure - received marginal treatment in every major prediction framework. The 2020 WEF report was built on surveys of 26 economies. The India hiring freeze, the GCC nationalisation-automation collision, Africa’s leapfrog opportunity in offline AI - none of it was modelled. The predictions were made for the world that sent delegates to Davos, not the world that does most of the work.

The psychology was entirely absent. University of Florida researchers have now proposed a clinical framework called AI Replacement Dysfunction - AIRD - to describe the specific psychological condition arising from AI transition anxiety. Spring Health research shows frequent AI users experience 45% higher burnout rates than non-users. Five years of workforce predictions said virtually nothing about what happens inside the human being going through this transition. The skills and structures problem was modelled extensively. The identity and wellbeing problem was ignored.

The entry-level pipeline collapse was not modelled. Every prediction framework assumed the career ladder stayed intact - that AI would augment experienced workers while new workers learned through entry-level roles and apprenticeship. Entry-level job postings in knowledge work have fallen sharply since 2022. One in three companies expects to eliminate entry-level roles entirely by end of 2026. The rungs are gone. No one predicted this.

The speed was wrong by an order of magnitude. Every framework assumed transition windows of three to five years. This month, Oracle moved in one morning.

The numbers that tell the story

Four numbers from Issue 003 of reAImagine.work:

Zero of four major prediction leaders predicted the entry-level pipeline collapse.

Zero mentioned India, GCC, or Africa substantively.

45% higher burnout among the very workers organisations most need to retain - the frequent AI users.

6am - the hour at which Oracle demonstrated what AI-accelerated capital reallocation actually looks like in practice.

They predicted years. It happened in hours.

The most damaging gap between prediction and reality is not the mechanism or the geography. It is the speed.

When I was selling Bersin by Deloitte to CHROs across South Asia, the implicit message was: you have time to prepare. The frameworks would unfold over years. Reskilling programmes would catch up. Organisations would adapt through considered transformation programmes.

Oracle’s 6am email invalidates that assumption permanently. When a company can eliminate 30,000 roles - while simultaneously reporting near-doubling profits - before the working day begins, the three-to-five-year preparation window is a fiction.

The CHROs who planned reskilling programmes for 2027 are already late.

What they said versus what happened

I want to be precise about what I am not saying. The WEF survey covers hundreds of companies. Bersin has been tracking HR technology for decades. Deloitte's Human Capital Trends is the largest longitudinal study of its kind. Ulrich's organisational capability frameworks remain structurally sound. These are serious people doing serious work — and that is exactly why the blind spots matter. The predictions were built on the best available data from the best-resourced institutions in the field. That data was Western, optimistic, and structurally blind to the majority of the global workforce. The problem is not carelessness. The problem is the view from Davos.

Why I published this in reAImagine.work

I started reAImagine.work because I could not find the publication I needed as a workforce strategy consultant operating across India, the GCC, and Africa. The publications I had helped build - or sell - were not covering the signals that mattered for my clients. They were not asking the questions that the CHROs I was coaching were actually asking.

reAImagine.work publishes every Friday. It covers AI and the world of work through the specific lens of India, the GCC, and Africa - the regions that the major prediction frameworks consistently ignore. It is free. No paywall. No corporate sponsor shaping the editorial line.

Issue 003 - published today - is our most ambitious issue yet. Beyond the prediction report card, it covers:

The physical AI inflection point - why Jensen Huang’s “ChatGPT moment for physical AI” declaration is now supported by production-scale deployment data across manufacturing, logistics, and energy.

Gemma 4 - Google’s release of a frontier AI model that runs entirely offline on a smartphone in 140 languages for free, and what it means for the AI access gap in emerging markets.

Deepfake hiring fraud - why your video interview process is now an attack surface, and what Cloudflare’s 2026 Threat Report says about organised identity fraud in remote hiring.

AI Replacement Dysfunction - the clinical literature on what AI transition anxiety is doing to the workforce your organisation most needs to retain.

The full issue also includes a nine-sector breakdown of how this week’s signals land differently in financial services, healthcare, manufacturing, energy, and five other sectors - and a regional read across India, the GCC, and Africa that you will not find anywhere else.

Read Issue 003

The full report card - with the detailed evidence behind each grade, the four blind spots analysis, and the reAImagine.work prediction check against our own Issue 001 forecasts - is live now.

reaimagine.work/articles/003-conversation

If you are a workforce leader, CHRO, or consultant operating in India, the GCC, or Africa - or if you advise organisations that do - reAImagine.work is the weekly read that covers the signals your current publications are missing.

Subscribe free at reaimagine.work

New issue every Friday. No paywall. No fluff.

Debu Mishra is a Dubai-based workforce strategy consultant with 30 years of experience across India, MENA, and Africa. He is the founder of reAImagine.work and Shim Partners Management Consultancy FZCO.