I let AI write most of my code this year. Here's what it's actually good at, and what it's not.

I let AI write most of my code this year. Not some of it, most of it. Autocomplete finishing my sentences, whole functions generated from a comment, entire first drafts of features I would otherwise have spent an afternoon typing out by hand. If you had told me a few years ago that I'd be running a software company this way, I would have laughed and gone back to typing. Here's what an actual year of it taught me. No hype, no dunking on the tech, just what I watched happen at my own keyboard across a handful of different products.

What it's actually good at

Boilerplate is the easy answer, and it's true, but it undersells the real pattern. AI is genuinely great at anything where the shape of the answer is well known and the specific details are what take time. Scaffolding a new component. Writing the first pass of tests for logic I already understand. Converting a chunk of data from one format to another. Catching the dumb mistake before I do, the missing null check, the variable name that doesn't match the one three lines up. None of that is glamorous. All of it used to eat hours of a week.

It gets even more useful the bigger the gap is between what I already know and what I'm building. Touching an API I've never used, wiring up a config for a tool I only reach for twice a year, getting oriented in a library's conventions on day one instead of day three. That's the honest sweet spot: AI fills gaps in my knowledge faster than I can fill them myself.

What it's actually bad at

Judgment calls specific to the product and the people using it. Why this tradeoff over that one, given what our actual users actually do. AI doesn't know that, and it will confidently guess anyway if you let it. It's also weak on long changes that cross a lot of files in a codebase with years of history and its own particular quirks, the kind of thing where half the difficulty is remembering why a piece of code looks strange, not writing new code that looks normal.

The finding that surprised me least, once I'd lived through a version of it myself, came from METR, a nonprofit that studies AI capabilities. In a study published in July 2025, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," they ran a randomized controlled trial with 16 experienced developers working in codebases they knew well, 246 tasks in total. Going in, the developers expected AI tools to speed them up by about 24 percent. What actually happened was the opposite: they were about 19 percent slower with AI than without it. And here's the part that should make everyone running an engineering team sit up: after finishing the tasks, those same developers still believed AI had made them roughly 20 percent faster. The gap between what people feel is happening and what a stopwatch says is happening was the whole story.

The codebase you already know is the trap

I've felt my own version of that gap. On a system I've lived in for a long time, I already have the fix in my head before I finish reading the bug report. Writing a prompt, waiting for a response, reviewing it carefully enough to trust it, and redirecting it when it wanders off in a plausible but wrong direction all cost real minutes. When I already had the whole answer, those minutes are pure overhead. AI isn't competing with a blank page in that moment. It's competing with knowledge already sitting in my head, and knowledge sitting in your head is fast.

That's the piece most of the excitement skips over. The marketing pitch is that AI makes you faster, full stop. The truth, for anyone doing serious work, is that it depends heavily on whether you're filling a gap or competing with something you already know cold. Pretending otherwise is how teams end up slower and don't notice, exactly like the developers in that study.

The rule I actually use now

First draft of something new to me? AI writes it, and I review it harder than I'd review a junior engineer's pull request, because it will sound confident whether it's right or wrong. Change to a system I've worked in for years, where I can already see the fix? I write it myself, and let AI check my typing and catch edge cases afterward. That one distinction, gap versus no gap, has done more for how I actually use these tools than any strategy memo I could write.

None of this took a big AI rollout plan to figure out. It took months of paying attention to when the tools helped and when they quietly got in the way, and being honest with myself about the difference. If you're trying to work out where AI actually fits in how your team builds, before you commit real budget or headcount to it, the free AI readiness assessment takes about five minutes and gives you a straighter answer than most vendor pitches will.