Every software engineer I know has an LLM open all day now.
Copilot. ChatGPT. Cursor. Pick your weapon. The tab is always there, right next to the terminal.
On paper, this should feel great. Faster coding. Less grunt work. More output per hour. The pitch writes itself.
But something feels off — especially in Indian IT services.
When everyone moves faster, the bar quietly moves up. Speed stops being a win and starts becoming the baseline. You're no longer impressive for shipping quickly. You're just keeping up.
And the job itself has changed shape.
You're not just writing code anymore. You're reviewing AI output. Double-checking logic that looks right but smells wrong. Owning mistakes you didn't fully write. The grunt work didn't disappear. It just put on a different outfit.
So I started digging into whether the speed gains are even real.
The numbers don't say what you'd expect
GitHub's own research claims Copilot makes developers 55% faster. That's the headline number everyone quotes. But that was measured on isolated, controlled tasks. Small stuff. Boilerplate. Greenfield.
When METR ran a study in early 2025 with experienced open-source developers — on real codebases, real complexity — the result was different. Developers using AI were 19% slower. Not faster. Slower.
The kicker? Those same developers believed they were 24% faster. The gap between perception and reality is wild.
And when teams tested this on actual projects, the gains shrank to 10-15%. Frontend monoliths saw maybe 20-25% improvement. Backend microservices? 5-7%. Barely worth the context-switching cost.
AI is fast when the task is small, isolated, and repetitive.
AI struggles when the codebase is large, context-heavy, and messy.
Guess which one describes most Indian IT services work.
The real problem isn't the tool
Legacy systems. Complex integrations. Clients who change specs mid-sprint. That's the daily reality for most engineers in Indian IT services. That's not where AI shines.
But the companies saw the pitch. "AI makes developers 55% faster." And they did what Indian IT services companies always do — they optimized for margin.
Fewer people. Same deadlines. More expected output. AI became the justification, not the solution.
NASSCOM is already warning that hundreds of thousands of mid-level jobs could vanish without reskilling. Not because AI replaced those people. Because management decided AI replaced those people.
Copilot didn't decide you should own mistakes you didn't write. Someone with a Jira board did.
So what actually shifted?
I don't think AI shifted productivity. I think it shifted pressure.
The engineer's job got harder to define. You're part writer, part reviewer, part debugger of someone else's confident-sounding nonsense. The output looks clean. The responsibility is murky.
And the mental load didn't go down. If anything, there's a new kind of fatigue — prompt fatigue, verification fatigue, the constant question of "is this actually correct or does it just look correct?"
The tools are genuinely useful. I'm not anti-AI. I use them every day.
But useful and liberating are not the same thing.
AI made certain tasks faster. It did not make the job easier. And in an industry that was already optimized for squeezing humans, handing management a new justification to squeeze harder was never going to end well.
The pressure didn't go away.
It just changed its shape.