Essay

AI Slowed Experienced Developers by 19%. That's the Point.

2 min read

In 2025, researchers gave experienced developers the best AI coding tools available — Cursor, Claude, the works. Then measured how fast they completed real tasks.

  • The Expectation

    The developers predicted they'd be 24% faster.

  • The Reality

    They were actually 19% slower.

  • Cue the "AI is overrated" crowd. Wrong lesson.

    Personal Experience

    The 15-Minute "Quick Fix"

    I learned this the hard way. A bug fix. Twenty, maybe thirty lines of code. I know this codebase — I have the map in my head. By hand, it's a two-minute job. I hand it to AI. Fifteen minutes later, we're still going.

    Why? Because AI doesn't have the map. It has to discover the terrain, think about it, propose a fix — and still miss a spot. Now we're doing another round. What should have been a coffee-break fix became a small project.

    I've watched teammates hand AI a complex bug and get back something that technically works but is three times more complicated than it needs to be. AI considered forty different angles before settling on the most thorough — not the most appropriate — fix.

    Now there's a telephone game: developer explains to me, I explain back to the developer, developer explains to AI, we repeat.

    The bug had two lines. The fix has twenty. It is frustrating.

    So AI is bad? No. That's also the wrong lesson.

    Historical Context

    The Factory Motor Fallacy

    When factories switched from steam engines to electric motors in the early 1900s, most owners just replaced the one big steam engine in the center of the factory with one big electric motor.

    Same layout. Same process. Barely any improvement.

    It took almost thirty years before someone realized — electricity means you can put a small motor at every workstation. Completely reorganize the floor. Productivity went through the roof.

    The technology was right. The thinking was thirty years behind.

    We're doing the same thing with AI. Same old workflow, new tool dropped in the middle.

    The Solution

    Plan Before You Prompt

    Here's what actually changed things for me: I stopped using AI as a junior developer who just writes code. I started using it as a very well-read colleague who knows everything, reads fast, and understands even faster.

    1. Discuss

      Discuss the problem first.

    2. Align

      Close on the approach.

    3. Execute

      Then ask it to apply.

    That one habit change — plan before you prompt — unlocks what AI actually is. The tool didn't change. The conversation before the tool changed.

    The speedup is real. It just doesn't start when you open the chat window.

    Pankaj Sarda

    Engineering Leader

    Building systems in bits, atoms, and books. Bridging high-scale software infrastructure and real-world operations.

    Read full profile