Stuck at 80: the AI-built site problem
The story arrives at our door in almost the same words every time. The demo took a weekend. It looked real, real enough to show people. Then came login, or Stripe, or the first deploy, and now every fix breaks something else, the credits are gone, and the site that was almost done in March is still almost done in October.
If that's you, the first thing worth saying is that you didn't do anything wrong. You picked Lovable or Bolt or v0 because they work, and the weekend demo was not a trick. These tools are genuinely good at the first 80 percent of a website. The trouble is that the last 20 percent isn't the same job done slower. It's a different job.
What the first 80 percent is made of
Layouts, pages, copy, buttons, a pricing table, a contact form that looks like a contact form. This is pattern work. The tools have seen millions of websites and yours, at this stage, can safely resemble them. When something looks wrong, you can see it's wrong, say so in plain English, and watch it get fixed. The feedback loop is instant and visual, and the tools are built for exactly that loop. That's why the demo weekend feels like magic. It basically is the tool's home turf.
What the last 20 percent is made of
Then you ask for accounts. Or payments. Or you push to production and the thing that worked in preview doesn't work at the real address. Auth, Stripe, database permissions, webhooks, deploys, error handling: the boring reliability work. This layer has two properties the first layer doesn't.
First, you can't see it. A misaligned header announces itself. A broken session token doesn't; it just logs someone out at checkout, sometimes, on some phones. You can't point at the problem in plain English because nobody can see where it lives.
Second, it's connected. The visual layer is mostly independent pieces, so regenerating one rarely hurts another. The reliability layer is one tangle of state, where the login touches the database, which touches the permissions, which touch the payment flow. A tool that fixes by regenerating, rather than by reasoning about your one specific tangle, will sometimes regenerate the knot tighter. From the outside, that feels like "fix one thing and it breaks ten others," which is exactly how people describe it.
You can read this pattern in public. On Vercel's own community forum, one v0 user describes spending over $300 trying to fix a single parser bug in an app they'd already put $2,000 into: "it's stuck in this infinite loop," "I'm seriously about to give up on this." A whole cottage industry now exists around finishing Lovable projects, which tells you how common the stall is. And the supply side keeps growing: Lovable reportedly raised $400 million in August at a valuation north of $13 billion, which means a lot more weekend demos are on their way to their own month ten.
None of this is a scandal. The tools sell the first 80 percent and deliver it. The stall happens because the remaining work rewards a different skill: holding the whole tangle in your head, knowing which failures only show up in production, and being patient about plumbing nobody will ever compliment.
What to actually do
You have three honest options.
Push through. Sometimes the right call, if the remaining problems are small and you can budget hours and credits without resentment. Set a limit first: a number of dollars or evenings after which you stop and reassess, instead of discovering the limit the way the forum posters did.
Start over conventionally. Painful, and occasionally correct, especially if the foundation (the database schema, the auth setup) is itself the tangle. The demo still earned its keep: you now know exactly what you want, which is worth more than most people admit.
Hire the last 20 percent out. Keep what works, hand the tangle to someone who untangles auth and payments and deploys for a living, and get your evenings back. This is usually the cheapest option per month of your life, which is the unit that actually matters by month ten.
Before you pick, it helps to know what you're actually holding. SalesLeaks scores an AI-built site the same way it scores any site: twelve plain questions about whether it can win a customer, not a review of your code. Three minutes, and the score shows before we ask you for anything.