The Judgement Deficit

India's biggest organisational invention was never a product. It was a shape.

The offshore services pyramid, wide at the base with freshers, thick with team leads, narrow at the top, lets a handful of firms hire graduates by the hundred thousand and still promise a bank in New Jersey that delivery would not slip. It ran on engineered replaceability. Knowledge was pushed out of people into process documents and delivery frameworks, so that double-digit annual attrition was a nuisance, not a disaster.

That structure is under real pressure now, and we are watching the wrong part strain.

Look at FY26. TCS closed the year with constant currency revenue down 2.4 per cent and an operating margin at a four-year high of 25 per cent, while cutting headcount by 23,460 to 584,519. Infosys grew 3.1 per cent but shed 8,440 people in the March quarter, though full-year headcount was still up about 5,000; its CFO called that seasonality. HCLTech's C Vijayakumar gave the phenomenon a name on his April earnings call: "AI deflation," worth 3 to 5 per cent off next year's revenue. His illustration was blunt. A deal that once fetched a hundred million dollars now fetches perhaps eighty, and takes 25 to 30 per cent more effort to win. TCS's K Krithivasan used a harsher word: degrowth.

The leaders are defending profit while pricing falls. At TCS, the defence ran straight through the workforce.

The usual conclusion: AI is eating entry-level jobs. True, and not the interesting part.

The interesting part is what those jobs were quietly doing. The pyramid was never only a cost structure. It was an apprenticeship system, likely the largest India has ever run. Ask anyone who spent their first three years inside one of these firms. Judgement was not taught; it accumulated, through work nobody writes case studies about: fixing a broken data feed at eleven at night, sitting silently through a client escalation, learning which deadlines are real and which are theatre, being wrong and surviving it. The base of the pyramid produced capable forty-year-olds as a side effect of billing hours. It was never designed to do that, which is exactly why nobody is defending it now.

That side effect is what disappears. Automation is arriving exactly where the learning happened: high-volume, low-stakes work where a twenty-three-year-old could afford to be wrong. We are removing the cheap-mistake layer while still expecting organisations to produce people who are reliably right when stakes are high. Nobody has a plan for this, because it never had a budget line.

The firms see part of the gap. TCS logged 69 million learning hours in FY26, up 23 per cent, with over 270,000 employees raising their AI proficiency. That is real money. But training is not consequence, and judgement is only ever made out of consequence.

The GCC boom hides the problem rather than solving it. Zinnov and NASSCOM count 2,117 global capability centres in India in FY2026, roughly a third more than in FY2021. Those centres are not buying capacity. They are buying judgement: people who can hold a global mandate, absorb ambiguity, and be trusted without supervision. That judgement was manufactured by the services pyramid over two decades. It is a stock being drawn down, not a flow being replaced. If the base stops seasoning people at scale, where does the GCC leadership of 2040 come from?

If that is right, three design choices follow.

First, treat judgement as something you manufacture, at a cost. If the model writes the first draft, juniors must review it, argue with it, and sign their name under it. Consequence is the input, not tasks.

Second, expect hierarchy to survive, but with a different job. Its supervisory purpose is being automated. Its teaching purpose is not. The question defining a manager stops being how many people report to you, and becomes how many got measurably better this year.

Third, separate the process from memory. Firms running attrition of 13.7 per cent (TCS) or 12.6 per cent (Infosys) never had institutional memory; they had documentation, which anyone with a subscription can now copy in an afternoon. Memory cannot be copied: why we walked away from that client, which shortcut nearly cost a licence. Trust lives in that residue. So does endurance.

Twenty years out, the institutions still standing will not be the flattest or most automated. They will be the ones who learned to produce judgment after the cheap work it came from went away. India built the biggest such machine in the world by accident. We should decide what replaces it on purpose.