The Smartest Thing I Did With AI Was Tell It to Get Out of the Way
Published: May 4, 2026 | Author: Todd Max Carey
Originally published on LinkedIn
TL;DR: When to NOT Use AI
After building a 60-engine AI automation system managing 33,000 contacts, I learned the hard way: AI is brilliant at diagnosis when you give it structure, but a liability for tasks that should just be Python scripts.
The Dream vs. Reality
I'm not a programmer. Two months ago I got embarrassingly excited about building my own AI system — autonomous scheduling, billing, outreach, research, content, monitoring. The dream.
I built the first part. It worked. I extended it. It broke. I fixed it. Extended it again. It broke differently. The first bit broke.
I had naively assumed the AI would just get what the system was supposed to do. It doesn't. It tries very hard and bills you enthusiastically for the attempt. The OpenRouter bills were a surprise!
The Breakthrough: Python Scripts > AI for Most Tasks
I moved to Claude Code to figure out what was going wrong. And slowly discovered something every real programmer already knows: the answer was to take AI out of most of it.
The parts that kept breaking were the parts using AI for things that didn't need AI. Tasks that could just be Python scripts on a schedule. No models. No tokens. Reliable, boring, fast.
Forcing Structure: The Seven Detectives Framework
The other breakthrough was forcing structure on the AI itself. I built what I call the Seven Detectives. Before suggesting any fix, Claude has to speak as all seven:
- Murphy maps every failure point
- Watson collects only facts
- Holmes finds inconsistencies
- The Architect checks downstream impact
- Socrates verifies against actual documentation
- Jobs asks if the fix is simpler or just messier
- Columbo asks the one question everyone forgot
Then Claude gives me a confidence percentage. Below 95%, we don't touch anything.
Results: 60+ Engines, 33,000 Contacts
Two months in, the system is finally real: 60+ engines, 33,000 contacts managed, cron jobs firing reliably, manageable bills.
The Lesson: Methodology > Model
AI is brilliant at diagnosis when you give it structure. It's a liability when you throw vague hope at it. The methodology matters more than the model.
Non-programmer building something with AI? Reach out.