Let me tell you about the four most dangerous words in modern software development:
"Just use AI."
I've heard this phrase twice in the last quarter. Both times, it nearly broke me.
Story #1: The Authentication Migration
Our client in medianeth, approached us with what they described as a "simple task."
"We're using Clerk for auth and we are having an issue with it. We want to switch to BetterAuth. Just use AI to do it should take like an hour, right?"
I stared at my screen. Then I laughed. Then I cried a little.
Let me explain what "switching authentication providers" actually means:
What the client imagined:
- Tell AI to do it
- Copy paste
- Deploy
- Go to lunch
What it actually involves:
- Audit existing auth flow — Every protected route, middleware check, session handler
- Map Clerk's API to BetterAuth's API — Different token formats, different session management, different OAuth callback patterns
- Migrate user data — Password hashes aren't transferable between providers. You need a migration strategy.
- Update every component —
<SignIn />,<UserButton />,useUser(),useAuth(),currentUser()— all different in BetterAuth - Test every edge case — Password reset, email verification, OAuth reconnection, session expiry, multi-device logout
- Handle existing sessions — Do you force-logout everyone? Graceful migration?
- Update environment variables — Different keys, different webhook formats
- Fix the 47 things that break — Because something always breaks
The AI can help with some of this. It can translate syntax. It can suggest patterns.
But it cannot:
- Understand your business logic
- Test in your staging environment
- Migrate your production users without data loss
- Debug the OAuth redirect that fails only in Safari on iOS 16.3
Did You Know?
Trivia: The average authentication migration takes 2-4 weeks for a production application. Not because the code is hard, but because auth touches everything. One missed middleware? Your entire API is exposed. One wrong redirect? Users can't log in.
Story #2: The Impossible Timeline
A different client, different project. They wanted:
- A full web application
- A React Native mobile app (iOS + Android)
- Admin dashboard
- Payment integration
- Real-time notifications
Their expected timeline? 30 days.
When I asked how they arrived at that number, they said:
"Well, with AI now, development is like 10x faster, right? So a 10-month project should take a month."
I wish I was making this up. I wish this was a comedy skit.
(Narrator: It was not a skit. The client was dead serious.)
The Math They Did:
Traditional timeline: 10 months
AI multiplier: 10x faster (apparently)
New timeline: 10 / 10 = 1 month
The Math That Actually Applies:
Traditional timeline: 10 months
AI benefit: Faster boilerplate, better autocomplete
Actual time saved: Maybe 20-30%
New timeline: 7-8 months
AI doesn't eliminate:
- Requirements gathering — Still need to understand what to build
- Architecture decisions — AI can't decide if you need WebSockets or polling
- Design iteration — Clients change their minds (often)
- Testing — AI-generated code has bugs too
- Deployment — Still need to configure servers, CI/CD, DNS
- Client feedback loops — "Can you move the button 2px to the left?"
The "One-Shot" Fallacy
Here's what clients don't understand about AI-assisted development:
AI is not one-shot. It's iterative.
When I use Claude or Cursor to write code, I'm not typing a prompt and deploying the result. I'm:
- Prompting for initial code
- Reading and understanding it
- Identifying issues
- Prompting for fixes
- Manually adjusting edge cases
- Testing
- Finding bugs
- Prompting for more fixes
- Eventually rewriting half of it anyway
The AI is a collaborator, not a replacement. It's like having a very fast, very knowledgeable junior developer who occasionally hallucinates entire functions that don't exist.
The Dunning-Kruger of AI
Non-technical clients see AI demos and think:
"It wrote a website in 30 seconds! Software must be easy now!"
What they don't see:
- The demo was a simple landing page, not a production app
- The developer prompted it 15 times before recording
- It still has CORS errors and broken mobile responsiveness
- Nobody tested it with real users
How I Handle "Just Use AI" Now
When a client says those words, I do three things:
1. Educate, Don't Argue
"AI helps me work faster, but it doesn't eliminate complexity. Here's what's actually involved..."
2. Itemize the Work
I break down every task in painful detail. When they see 47 line items, they understand.
3. Charge Accordingly
I've started adding a "complexity surcharge" when clients underestimate work based on AI assumptions. It's a polite way of saying: "Your lack of understanding will cost us both time."
The Silver Lining
Look, AI has made me faster. Genuinely.
- Boilerplate? Gone in seconds.
- Documentation? AI reads it for me.
- Regex? I never learned it and never will.
- Debugging? "Why doesn't this work" + code = instant hypotheses.
But faster doesn't mean instant. 10x autocomplete speed doesn't mean 10x project completion speed.
The bottleneck was never typing. It was thinking.
And AI can't think for you.
To My Fellow Developers
If a client tells you to "just use AI," send them this article.
Then send them an invoice. A big one.
They'll learn eventually. Or they won't. But at least you'll be paid for the chaos. 💰
