I've used Cursor and Claude to build significant portions of all 6 of my SaaS products. AI coding tools are genuinely impressive in 2025. But there are specific places where they still break – and if you don't know where those are, you'll ship bugs you don't understand.
What AI Gets Right
UI scaffolding is where AI shines. Give it a component description and it produces clean, accessible JSX in seconds. CRUD operations, form validation, basic API routes all solid. I'd estimate AI handles 70% of the boilerplate in a typical MVP correctly on the first try.
Where It Still Breaks
Complex webhook logic. AI will write you a webhook handler that looks correct but misses idempotency, signature verification edge cases, or race conditions. Payment edge cases proration, failed payment retries, subscription pausing AI gets these wrong consistently. Custom business logic that requires understanding your specific domain.
The Architecture Problem
AI can't make architecture decisions for you. It doesn't know your scale requirements, your team's capabilities, or your business constraints. Ask it to design your database schema and it'll give you something technically correct but potentially wrong for your use case.
The Right Mental Model
Think of AI as a very fast typist who knows every syntax and pattern but has never shipped a product. It can write the code. You need to know if the code is right.
// AI will write this correctly:
const user = await supabase.from("users").select("*").eq("id", userId);
// AI will miss this:
// What if the webhook fires twice?
// What if the user was deleted between the webhook and this query?
// What if this runs during a database migration?I use AI to ship faster, not to replace judgment. Book a call if you want both.
