There is a story about 2026 that almost everyone in my industry has already accepted, and I want to argue with it before it hardens into received wisdom. The story goes like this: artificial intelligence has finally become good enough to do the work of people, the layoffs prove it, and the sooner your organisation accepts that reality the smarter you look. It is a clean narrative. It fits on a slide. And I think it is mostly wrong, or at least wrong about the part that matters.
I run AI at a large Indian conglomerate, so I read these announcements the way a mechanic listens to an engine, not the way an investor reads a press release. When I look at the wave of cuts this year, roughly 120,000 tech roles gone across Amazon, Dell, Meta, Microsoft, Oracle and others, with AI named as the reason, I do not see a machine that has learned to do our jobs. I see a treasury decision in fancy dress. AI is real. That was never in doubt. What I want to know is what these companies are doing with their money, and why they need AI to explain it.
The story everyone has already bought
Start with the headline cases, because they set the tone for everyone below them. Oracle cut about 21,000 jobs, taking its workforce from roughly 162,000 down to 141,000, nearly 13 percent, and it put the AI rationale for that restructuring into its 10-K. Microsoft trimmed another 4,800 roles in the same season it is pouring capital into AI infrastructure. Each of these gets reported, reasonably enough, as evidence that the transformation has arrived and the humans are being released.
Here is where I part company with the consensus. When "AI" turns up as the stated reason for job cuts inside an SEC filing, I do not read that as an operational fact. I read it as a communications choice. A 10-K is not a confession, it is a positioning document, and executives know exactly how each sentence will land with analysts. "We are cutting staff because AI now does the work" is a sentence that makes a CFO sound visionary. "We over-hired during the boom and now we are correcting" is a sentence that makes the same CFO sound like he made a mistake. Both can describe the identical headcount number. Only one gets written down. So when I see the AI rationale formalised in a regulatory filing, my first instinct is not to believe it more. It is to ask what it is being used to avoid saying.
Follow the money
The detail that most of the layoff coverage glides past is the awkward one. A striking number of these firms were posting stable or even record revenue while they were cutting. Companies being hollowed out by obsolescence do not post record revenue. Companies reallocating capital hard do exactly that. Oracle spent about 1.8 billion dollars on restructuring, up from 374 million the year before, while committing to a heavy AI infrastructure buildout. Microsoft's cuts land in the same period as its own large spend on AI infrastructure. Read those two facts together and the real story is duller than the headline: they needed cash to buy compute, and payroll is the fastest line to cut.
Now hold that against the economics of intelligence, because this is the part that should embarrass the efficiency argument. The cost of AI output is falling off a cliff. OpenAI's newest release, GPT-5.6, is claimed to be about 54 percent more token efficient on coding tasks and roughly one-third cheaper than a leading rival. Take the efficiency story at face value for a second. If the marginal cost of a unit of AI work is collapsing this fast, the human wage bill was never the binding constraint in the first place. You do not need to fire 21,000 people to afford intelligence that is getting cheaper by the quarter. The savings are coming from the technology curve regardless. So the cuts are not paying for the AI at all. They are funding the buildout: the GPUs, the data centres, the multi-year capex commitments. AI is the alibi that makes the reallocation sound like foresight instead of a bet. I do not begrudge anyone the bet. I begrudge the labelling.
The real self-inflicted wound is the bottom rung
Where I get worried, and where I think the industry is sleepwalking into something costly, is a number nobody is cheering. Entry-level job postings have fallen by around 35 percent over eighteen months as AI adoption picked up. That is not a rounding error. That is the bottom of the pyramid being quietly kicked out. And it is happening at the exact moment when, on PwC's reading of over a billion job ads from six continents, the labour market is splitting into two tracks, with judgement and leadership commanding a rising wage premium.
The logic gap is almost comic. Everyone agrees that human judgement is the scarce, valuable thing AI cannot yet replace. Everyone also seems to be shutting down the mechanism by which judgement gets made. Twenty years of watching people grow has taught me one thing above the rest: nobody is born a senior. A fresher becomes a senior by doing the unglamorous work, getting it wrong, being corrected, sitting in the room while someone more experienced makes a hard call, and slowly accumulating the pattern recognition we later dignify with the word "judgement". That process has a name. It is apprenticeship. You cannot keep harvesting senior judgement forever if you stop planting juniors. Cutting the training ground to look efficient this quarter is not prudence. It is malpractice, and the invoice arrives on a delay long enough that the executive who ordered it will have moved on before it lands.
The counter-argument I take seriously
I do not want to strawman the other side, because there is a real argument there and I hold part of it myself. Agentic AI is not hype. It is a structural disruptor. Gartner reckons that up to 234 billion dollars of enterprise application software spend is exposed to agentic arbitrage between now and 2030, about 20 percent of enterprise application SaaS spend, as agents complete tasks across multiple systems and cut the need to click through screens. I believe that. And tools like ChatGPT's new work offering really are consolidating what used to be several analyst-level tasks into a single prompt. So tasks are being automated, and the per-seat economics that software companies built their models on are wobbling. I have sat through enough vendor decks to know the difference between that and a machine that replaces judgement.
But automating a task is not the same thing as eliminating the person who was going to grow into the role. This is the exact confusion driving the junior-hiring freeze, and it is a category error with enormous consequences. A first-year analyst is not valuable because of the specific reconciliation or the specific deck they produce this month. That output was always cheap, and now it is cheaper. The analyst is valuable because in four years they become the person who knows which reconciliation matters and why the deck is wrong. When you automate the task and then decide you therefore do not need the person, you have optimised away the seed because you did not like the price of the fruit. The task was never the point. The person was.
India's contrarian playbook
The clearest live experiment in the world on this question sits in my own backyard, which is one of the few advantages of writing from India right now. TCS just reported its AI business reaching a 2.6 billion dollar annualised revenue run rate, up 13.6 percent quarter on quarter, and rather than cut, it made a net addition of 9,279 employees while hiring and onboarding 14,000 campus graduates in the quarter, with its CEO stating plainly that he does not believe AI will reduce overall headcount. The Western coverage and the Indian coverage of the same technology barely sound like they are describing the same thing.
I do not think TCS is being sentimental, and I would be wary of anyone who frames it that way, as though hiring juniors were an act of charity rather than strategy. It is the sharper read of where AI value lands. If AI is expanding the total demand for AI-services work, which the revenue number strongly suggests, then the constraint on capturing that demand is delivery capacity, which means people. India's talent pyramid, the thing Western commentators have spent a decade describing as a low-margin liability, is precisely the asset you want when the pie is growing and you need people who can be made productive quickly. The game becomes using AI to compress a fresher's time-to-productivity, to get them doing useful, judgement-adjacent work in months instead of years, rather than using AI as the excuse never to hire them at all. That is a completely different use of the same tool.
My worry, and I will say it directly because I hear it in meetings, is that Indian firms will reflexively copy the Western script. It is a real temptation. The Western narrative is louder, it is written up admiringly in the outlets our boards read, and "we are cutting to fund AI" sounds sophisticated in a way that "we are hiring 14,000 freshers" does not, until you do the maths on 2029.
My take
So here is what I am telling my own leadership at Mahindra, and what I would tell any Indian enterprise leader who asks. Do not cut the bottom rung. Whatever else you do with AI this year, protect the graduate intake as if it were a strategic reserve, because it is one. Deploy the technology to shrink the time it takes a new hire to become useful, and treat that compression as the return. The efficiency dividend from AI is real, I am not romantic about it, and it will keep growing as the cost curves keep falling. The only open question is what you do with the dividend. The easy thing is to book it as a headcount saving, tell your investors a tidy story, and enjoy a clean cost line for a couple of years. I would rather spend it training the layer of people AI cannot replace, because I am going to need them and I cannot buy them fully formed on short notice.
The companies quietly starving their junior pipelines in 2026 are not going to feel it in 2026. That is what makes the mistake so seductive. They will feel it in 2029, when they go looking for the seasoned judgement they have been congratulating themselves on retaining, and discover they stopped growing it three years earlier. There will be no senior bench, because they fired the freshers who were supposed to become it, and you cannot capex your way out of a gap in human experience. Given a choice between the two experiments running in front of us, I know which one I would rather be running. I would take TCS's bet over Oracle's, and I do not think it is especially close.