Everyone wants to know which AI to use. It’s the wrong question, and it’s quietly become one of the most expensive mistakes I see.
The pattern is always the same. A team picks a model—usually whatever’s newest or whatever IT approved—then throws entire workflows at it. Summarize these interviews. Write this report. Analyze this data. When outputs disappoint, they blame the model, upgrade to something more expensive, and repeat. I’ve seen companies burn through three enterprise contracts in a year this way.
Almost nobody tells you this explicitly: AI doesn’t fail at the workflow level. It fails at the task level. And most workflows contain five or six tasks pretending to be one.
“Generate a PRD” sounds like one thing. It’s actually customer synthesis, UI analysis, feature design, roadmap alignment, and document construction. Each requires different capabilities. When you throw them all at ChatGPT, you get a document that looks professional and falls apart under scrutiny. Not because the model is bad—because you asked a specialist to be a generalist.
McKinsey says 80% of organizations use AI somewhere. BCG says 74% haven’t seen tangible value. It’s not just skills or data quality. In the implementations I see, the deeper gap is that we’re planning at one level of abstraction and executing at another.
Here’s what I’ll cover:
The task decomposition framework I use to break workflows into AI-ready pieces—with real examples from regulatory reporting, customer success, and product development
Which models I’ve found work best for which cognitive tasks—not benchmarks, but patterns from dozens of implementations
What multi-model setups actually cost in practice, and why they quietly beat “one model for everything” once you do the math
How to build model intuition through deliberate practice—the skill that separates teams getting value from teams getting frustrated
No theory. No benchmarks. Just what I’ve seen work.
Listen to this episode with a 7-day free trial
Subscribe to Nate’s Substack to listen to this post and get 7 days of free access to the full post archives.
