On April 23, Infosys's CEO disclosed that AI had cut the cost of transformation programs by 60%. In the same quarter, headcount fell by 8,440. The company called it a demand-supply equation. The arithmetic suggests something more structural.
On April 23, a journalist on Infosys's Q4 earnings call asked CEO Salil Parekh a direct question. Was AI cannibalizing Infosys's own services revenue? Were clients paying less because AI was doing more of the work?
Parekh answered honestly. He said yes, there was compression in tech services and BPM work. And then he said the specific thing that has been sitting with me since:
The estimates of the cost has come 60% lower than what they would have done before the AI approach.
Sixty percent. Not a pilot result. Not a projected figure. A disclosed outcome, on the record, on a public earnings call.
The stock moved. The transcript went live. And then the industry largely moved on, the way industries do when the data confirms something everyone suspected but nobody wanted to name first.
I don't think we should move on.
The Word "Compression" Is Doing a Lot of Work
In the language of earnings calls, "compression" is a financially precise term. It means the same output achieved with less input. It is accurate, neutral, and designed to travel well.
What it describes, in this case, is something more specific: a transformation program that a client used to pay $10 million for can now be delivered for $4 million when AI handles the analytical scaffolding, the documentation, the structured coding, and the BPM workflow design that used to require large teams working through structured phases over twelve to eighteen months.
The $6 million doesn't disappear. Some of it becomes margin. Some of it becomes pricing advantage in competitive bids. And some of it used to be headcount.
In Q4 FY26 - the three months from December 2025 to March 2026 - Infosys's employee count fell from 337,034 to 328,594. That's 8,440 people in a single quarter, according to the company's own fact sheet. When the CFO was asked about it directly, he called it a "demand-supply equation." He's right. What changed is which variables are in that equation now.
The Pyramid That Built an Industry
India's IT export sector - the $250B+ industry that made software engineering a generational career anchor for millions of families - was built on a specific workforce pyramid. Large base of junior engineers executing structured, repeatable technical work. Smaller layer of mid-level architects, project leads, and domain specialists. Thin top layer of client relationship owners and senior delivery leaders. The pyramid worked because the base was large, trainable, and economically viable in the global delivery model that companies like Infosys mastered.
The work at the base - the first drafts of code, the documentation passes, the BPM workflow design, the structured analysis that becomes a deliverable - is precisely the work Parekh was describing when he named where AI is creating compression. Not the top of the pyramid. Not the relationship layer. The base. The part that produced the talent that eventually filled the middle.
That's the part nobody wants to say out loud.
Infosys simultaneously announced a partnership with Cursor to equip over 100,000 of its software engineers with agentic coding tools. That number is not a pilot - it's the operating model. The hypothesis embedded in that decision is that engineers augmented by AI can cover more ground than engineers working without it, which means you need fewer engineers per unit of delivery. This is the same logic that created the pyramid in the first place. It's now being applied to compress it.
What the Transcript Doesn't Say
Infosys also announced plans to hire more than 20,000 freshers in FY27. This sounds like continuity - the pipeline stays open, the entry point holds. But hold that number against the context and something shifts.
The 8,440 who left in Q4 weren't all juniors. And the 20,000 arriving into a workforce where 100,000 of their senior colleagues are being equipped with AI tools that do what junior engineers used to do will not experience the same ramp that built the mid-level layer in the previous decade. The structured repetition - two years of foundational testing, documentation, and incremental coding that trained pattern recognition before anyone was trusted to lead a stream - is compressing. Not disappearing. Compressing. And the compression is not neutral.
I spent time at Zendesk watching large IT services firms deploy enterprise software for clients at scale. The delivery teams that executed best weren't the ones with the best tooling or the most certified architects. They were the ones where the mid-level had accumulated enough repetitions to anticipate the failure mode before the client named it. That institutional radar didn't come from training. It came from being the junior on enough projects where things went sideways, understanding why, and carrying that read into the next engagement. The pyramid produced those people. What produces them next is a question I haven't heard anyone at the senior level answer concretely.
What Attrition Is Actually Telling You
One number in the Infosys fact sheet gets less attention than it should. Attrition fell - from 14.1% a year ago to 12.6% at the end of March 2026.
In a demand environment where talent is being pulled toward better opportunities, falling attrition signals satisfaction. Retention. Organizational health. In an environment where external opportunities are contracting - where the same AI compression Infosys is running internally is running across competitors, mid-tier firms, and domestic startups - falling attrition signals something different. People are staying because the calculus outside has changed. That's not a bad thing for the quarter. It is a different thing. And it's worth reading accurately.
I keep revising how I think about this. Some months I'm persuaded the industry adapts faster than I'm crediting - that the new pyramid just looks different, with AI-augmented engineers at the base doing work the previous base couldn't have. Other months I do the arithmetic again and it doesn't close the same way.
The Question That Isn't Being Asked
The earnings call conversation was appropriate for an earnings call: margins, guidance, AI revenue trajectory, client demand signals. These are the right questions in that context.
What the format doesn't create space for is the structural question underneath: what does a 60% cost compression in core services mean for the career contract that India's IT sector represents to the people inside it? Not the company. The people. The 328,000 currently employed. The 20,000 arriving next year. The several million who built careers on a model that assumed a base layer of foundational technical work would always exist and always grow.
That question doesn't fit in a guidance range.
I'm not saying Infosys is in crisis or making wrong decisions. They are doing what a well-run company under competitive pressure does: deploy the tools that change the cost structure, retain the margin, and adjust the headcount to match the new equation. That's the job. The harder question belongs to the industry, to policymakers, and to everyone who has a stake in what happens to the people the pyramid was designed to develop.
What Comes After Compression
Salil Parekh used "compression" because it is the accurate word. It is. Compression means the same or better output with less input. In manufacturing, that's called efficiency. In knowledge work, the input being compressed includes the hours that teach people how to think in a domain before they're trusted to lead in it.
Infosys's 100,000-engineer agentic AI deployment is the largest real-world experiment in the industry on this question. What it produces - a generation of engineers who are faster and shallower, or a generation that is faster and still substantive - will tell us a lot about whether the pyramid adapts or hollows out. We won't know for two or three years.
In the meantime, the word "compression" is on the transcript. The arithmetic is available to anyone who wants to do it. The industry moved on. I think it's worth doing the math.
