KJ Huang

Building an app from scratch with AI is feasible, but converting an existing product still requires at least 50% human involvement. That’s my conclusion after using AI to build a native mobile app: AI is remarkably effective going from 0 to 1, and remarkably fragile when legacy code enters the picture. Here’s what I learned, and how I’d describe the current state of AI-assisted development.

Where Does AI Excel? Going from 0 to 1

AI has proven to be more helpful than I expected when it comes to going from 0 to 1. Especially with SwiftUI, AI can be incredibly efficient at building UI, setting up the basic architecture, and even connecting APIs.

But when I tried to convert an existing product into a native app, that’s where the real challenges started to emerge.

Why Does Converting an Existing Product Still Need Humans?

1. AI struggles to understand the true logic behind legacy code

Even if I carefully explained the logic behind my old website, AI often generated code that looked plausible but didn’t actually work. Without proper context, AI can’t truly grasp what the legacy system was doing.

2. AI’s output still needs a real engineer to review and refine

Whether it’s data synchronization, permission management, or app lifecycle handling, AI often produces solutions that appear correct but are fundamentally unstable. I had to debug and rewrite large parts manually, and in those moments only a human developer can truly understand what needs to be done.

You could argue that this is because the original code wasn’t written with AI in mind. But this highlights a major commercial challenge: how do we integrate AI into existing codebases effectively?

3. From scratch, AI is an amazing collaborator

With no technical debt, no legacy systems, and no confusing data schemas, AI can deliver fast and solid results, as long as you clearly articulate what you need. In these scenarios, the initial output is often good enough to test directly.

My Assessment

I would describe the current state of AI-assisted development like this: “AI is a brilliant intern who can write code, but you need to be the product owner who knows the logic.” It doesn’t know why you designed a data structure a certain way or why a specific condition exists. But if you give it clear direction, it can implement it quickly.

I still firmly believe: AI is an amplifier for engineers, not a replacement.

Update (2026): a year later, the picture changed dramatically. Read the follow-up: AI 今年明顯變聰明非常多 — App 開發實戰分享.

Frequently Asked Questions

Can AI build an app from scratch?

Yes. For greenfield projects with no technical debt or legacy systems, AI can build UI, set up architecture, and connect APIs efficiently, and the initial output is often good enough to test directly, as long as you articulate clearly what you need.

Why does AI struggle with legacy code?

AI lacks the context behind legacy systems: why a data structure was designed a certain way, or why a specific condition exists. It generates code that looks plausible but often doesn’t work, especially around data synchronization, permissions, and app lifecycle handling.

Will AI replace software engineers?

Based on this experience, no. AI is an amplifier for engineers, not a replacement. It works like a brilliant intern who can write code fast, but it needs a human who owns the product logic to direct, review, and refine its output.

If you’re also exploring how to use AI to build apps, or looking to bring an existing product into the native mobile world, I’d love to connect and exchange ideas.