I've been watching the "AI builds your SaaS in a weekend" claims for a couple years now, and I used to roll my eyes at most of them. Too much of it read like marketing copy for tools trying to sell you a subscription. But somewhere in the last year, the story quietly stopped being hype and started being... true, at least for a specific kind of developer doing a specific kind of work. So this week I want to talk about what's actually changed, and what I think people are getting wrong about it.
THE REAL SHIFT ISN'T THE AI, IT'S THE WORKFLOW
Everyone focuses on the model. Cursor vs. Copilot vs. Claude Code vs. whatever agent launched this month. That's the boring part of the story, honestly. The interesting part is that the unit of work changed. Two years ago, AI helped you write a function. Now it plans a feature, touches a dozen files, runs the tests, and shows you a diff before anything ships. That's a different job. You're not typing code anymore, you're reviewing decisions.
That shift is why the productivity numbers are all over the place depending on who you ask. A McKinsey survey of thousands of developers earlier this year found AI tools cutting time on routine coding work by close to half, with meaningful drops in time-to-production too — but the same research flagged that review time and bug counts both crept up whenever developers got sloppy about actually checking the AI's work. Meanwhile, a well-known randomized trial found close to the opposite on mature, unfamiliar codebases: experienced open-source developers ended up slower with AI assistance than without it, even though they'd predicted the opposite going in.
My honest take: both are correct, for different situations. Greenfield SaaS product, familiar stack, you know roughly what you're building — AI is a genuine multiplier. A gnarly five-year-old repo with tribal knowledge baked into it — AI can actively slow you down while you fight it. The "10x" headline is real, but it's a 10x on the first mile, not necessarily the whole marathon.
WHAT THE ACTUAL 10X LOOKS LIKE DAY TO DAY
The founders and indie developers I'd trust on this aren't using AI to write clever algorithms. They're using it to kill the parts of building a SaaS product that were never really about engineering skill in the first place: auth boilerplate, CRUD scaffolding, Stripe integration, admin dashboards, test scaffolding, the fifteenth "add a filter to this table" ticket. That stuff used to eat weeks. Now it's a prompt and a review pass.
That's also why the stack people reach for has narrowed. A lot of solo builders are converging on the same handful of pieces — something like Supabase, Next.js, and an AI API bolted on — not because it's technically superior to everything else, but because it's the path of least resistance for an AI agent that's read a thousand tutorials on exactly that combination. There's a real lesson in there: part of "building faster with AI" is choosing boring, well-documented tools the model already knows cold, not the interesting new framework you were excited about.
THE PART NOBODY WANTS TO SAY OUT LOUD
Here's my actual opinion, and it might be unpopular: the biggest bottleneck left isn't the AI's coding ability anymore. It's your own judgment about what to build. AI collapsed the cost of building, which means the cost of building the wrong thing also collapsed — you just get there faster now. I've seen people ship an AI-generated MVP in a weekend, and then spend three months confused about why nobody's using it, because they skipped the part where you talk to five potential customers before writing a line of code. AI didn't remove that step. It just made it easier to skip.
The other quiet cost is trust debt. If you never really read the diffs, you don't actually know your own codebase anymore. That's fine until the day something breaks in production at 2am and you're staring at a function you technically own but have never seen before. I don't think that's a reason to slow down. I think it's a reason to treat AI-generated code the way a senior engineer treats a junior's pull request — fast, but never a rubber stamp.
IF YOU'RE STARTING THIS WEEK
Pick one tool, not five. Use it on something with a tight, well-understood scope first — an internal tool, a small feature, not your entire product. Read every diff for the first month, even the boring ones, until you've built a feel for where it's reliable and where it quietly makes things up. And keep talking to users at the same pace you always should have. The tools got faster. Good judgment didn't get any less necessary — if anything, it got more valuable, because it's now the only part of the process AI can't do for you.
That's the real 10x: not that the AI writes the code, but that it buys you back the time to actually think about whether you're building the right thing.