AI App Builders

AI Can Write the Code. Why Is Shipping Still So Slow?

Let’s start with a scene most founders know by heart. You’ve got an idea, a free afternoon, and zero patience. You type a few sentences into an AI app builder, grab a coffee, and come back to a clickable prototype with a login page, a few forms, and a database humming underneath.

Feels like magic, right?

Then reality shows up. Somebody has to review that code. It still needs to be tested, and a human has to sign off before it goes live. That’s where the magic starts to wobble.

The hype has numbers behind it

“Vibe coding” started as a throwaway phrase from AI researcher Andrej Karpathy. By late 2025, Collins Dictionary had made it Word of the Year. When the dictionary people notice your slang, it’s no longer slang.

Developers jumped in with both feet. Stack Overflow’s 2025 survey drew more than 49,000 responses, and 84% of developers said they use AI tools or plan to, up from 76% the year before. Among professional developers, 51% use them every single day.

The speed isn’t imaginary either. In a GitHub experiment, developers built an HTTP server in JavaScript, and the group with Copilot finished in about 1 hour 11 minutes versus 2 hours 41 minutes for everyone else. That works out to 55.8% faster. More of them crossed the finish line too: 78% completed the task, against 70% without the tool.

Mind the fine print, though. That was one task on one afternoon. A drop in the bucket compared with a real product.

What these builders actually do for you

Think of the chores every developer has done a hundred times: login systems, forms, database calls, API connections, page layouts. Nobody brags about them, but they swallow hours. An AI builder churns out that groundwork, and you steer it in plain English. “Make the button larger.” “Add a login page.” No hand-editing required.

That opens the door for people who’ve never written a line of code, mostly to build internal tools, dashboards, and MVPs. Demos get better too. Stakeholders click through something real instead of nodding politely at slides.

Startups noticed early. Y Combinator’s Jared Friedman said about a quarter of the Winter 2025 batch had codebases that were 95% AI-generated, excluding imported libraries. He was quick to add that these weren’t greenhorns. They were technical founders who could have built the whole thing themselves.

Now here’s the catch. Everything above happens at the front of the line: prompt, generate, preview, tweak. The back of the line is a different animal.

Where the speed goes to die

MIT Sloan’s Mert Demirer, Penn’s Leon Musolff, and Liyuan Yang tracked more than 100,000 GitHub developers through six stages of work, from typing the first line of code to releasing the finished product. Demirer says developers tell him they now do in minutes what used to take a full day. The study backs the first half of that. Developers using AI agents logged up to 180% more coding activity.

But look at what came out the other end:

MeasureChange with AI tools
Coding activity (autocomplete only)+40%
Coding activity (plus sync agents)+140%
Coding activity (plus async agents)up to +180%
Projects+50%
Releases+30%

(No release figure for async agents, because they can’t release software on their own.)

Why the gap? Plain old human bottlenecks. Code review, merging, and release still run on the same slow, careful processes they ran on before AI showed up. Pour more water into a funnel, and it only comes out the bottom at one speed.

The app stores add an odd twist. New apps have poured in since early 2025, yet downloads and reviews barely budged. Translation: a lot of apps nobody uses. Building got cheap. Earning attention didn’t.

Developers say “yes, but” to AI

Honestly, the survey data reads like a love-hate letter. In that same Stack Overflow study, 46% of developers said they don’t trust AI output to be accurate, up from 31% the year before. Only about 3% say they “highly trust” it. And 61.3% want to fully understand the code they ship.

Look at where they draw the line, too. Roughly 76% don’t plan to use AI for deployment and monitoring, and 69% won’t touch it for project planning. Developers will let AI draft the code. They won’t let it hold the keys to production.

That wariness turns every generated change into a small detective job, and the security and technical-debt risks of vibe coding explain why reviewers dig so hard. What did the tool touch? Why? Most builders tuck all of that behind a chat box, which doesn’t help. A few lay it out in the open. ProjectAAL shows the underlying operations on every file, along with token costs, execution details, and a re-roll option, so a reviewer has something solid to check and the developer keeps a firm grip on the codebase.

So what do you do with all this?

Money first. BLS puts the median annual wage for software developers, QA analysts, and testers at $134,040 as of May 2025, more than two and a half times the $50,980 median across all U.S. occupations. Every hour a reviewer burns on murky generated code costs real money.

And the jobs aren’t going anywhere. BLS counted about 1.9 million of these roles in 2025 and projects 10% growth through 2035, with around 106,100 openings a year. Somebody has to be at the wheel, and judging whether code is safe to ship now ranks among the skills a computer science degree builds toward.

Skip that judgment and the bill arrives later. Freelance marketplaces already report a surge in requests to fix AI-generated code that shipped without anyone fully understanding it.

A few rules of thumb. Test everything the AI hands you, because faster doesn’t mean right. Keep a human responsible for security, scalability, architecture, data protection, and maintenance. No builder owns those for you, and the devil is in those details.

Confirm code export before your first prompt, too. Some platforms give you clean, standard code you can take anywhere. Others lock your app in their own hosting. Don’t put all your eggs in one basket and find out later you can’t carry the basket out the door.

One last tip from Demirer: shrink the team and aim the spare hours at the tail end of the pipeline, meaning merging, releasing, and upkeep.

Writing code was never the whole job. AI just made that impossible to ignore.

Related: AI in Construction Specs: Why Better Tools Won’t Fix Bad Document Control

Tags: