Every step in any process requires one of three types of intelligence.
1. Rules (Deterministic)
- What it is: Logic.
IF/THEN. Math. Same input, same output, every time. - The Trap: Using a language model to check if
2+2=4. It doesn’t make your system “smarter.” It makes it unpredictable, slow, and expensive. - The Rule: If a simple script can do it perfectly, do not use AI.
2. Models (Probabilistic)
- What it is: Pattern recognition. Summarization, classification, prediction.
- The Trap: Treating a guess as a fact. A model’s output is a high-probability suggestion, not ground truth.
- The Rule: Validate everything. Check the confidence. Structure the output. Plan for when the model is wrong.
3. Minds (Human)
- What it is: Accountability. Judgment, empathy, ethics, and owning the consequences.
- The Trap: Automating a decision you can’t explain to your boss, your customer, or a regulator.
- The Rule: A model can support a high-stakes decision. It cannot own it. Someone has to be responsible.
Your Sanity Check
Next time you’re in a meeting about an AI agent, ask these questions. If the team doesn’t have good answers, you have a problem.
“Which steps here are just Rules?”
Force the team to find the simple logic first. This will slash your budget and increase stability.“What happens when the Model is only 40% confident?”
If there’s no fallback path to a Rule or an escalation to a Mind, the design is too fragile.“How do we validate the Model’s output before the next step runs?”
If the answer is “we trust it,” they are building on sand. You need a validation layer—another Rule—to protect the rest of the system.“Who owns the outcome when this makes a mistake?”
This question uncovers which steps truly belong in the “Minds” zone. If no one can answer, you are automating accountability away. That never ends well.
The Goal Isn’t Full Automation
The boundaries between these zones will shift. As models improve, some “Mind” tasks will become “Model” tasks. That’s fine.
This isn’t a static map. It’s a diagnostic tool.
The goal isn’t to remove humans from the loop. It’s to use them for the one thing AI can’t do: take responsibility.
Stop building agents.
Start building systems.