Chapter 1: Blockers

The Reliability Crisis in Agentic AI

Chapter 1: The Deterministic Gap

Why Probabilistic Intelligence Breaks Enterprise Systems

🎯 Difficulty Level: Easy
⏱️ Reading Time: 15 minutes
👤 Author: Rob Vettor
📅 Last updated on: March 8, 2026

The Problem

blockers blocker-mitigation

Your slide is very close conceptually, but the wording can be tightened so the logic is crystal clear and defensible to architects. Right now a few phrases blur the boundary between what can truly be deterministic vs. what can only be constrained.

The key distinction you want is this:

The model remains probabilistic. The system around it can be engineered deterministically.

That idea is extremely strong, but the slide should express it more precisely.

The Core Concept (What the Slide Should Say)

There are two layers in an AI system:

1️⃣ Deterministic System Controls

These can be engineered deterministically.

Security controls

Compliance enforcement

Data quality validation

Governance workflows

Policy enforcement

These are software engineering problems, not model problems.

2️⃣ Model Behavior

These cannot be made deterministic, but they can be constrained and governed.

Predictable AI behavior

Content safety enforcement

These depend on probabilistic reasoning, so the system must wrap them in deterministic controls.

The Most Important Phrase on the Slide

Your bottom-right label:

CONSTRAIN — NOT ELIMINATE

That is exactly correct.

The model cannot become deterministic.

But the system can constrain the risk envelope.

Where the Slide Can Be Improved

The phrase:

"Can be engineered deterministically"

is correct but could be clearer.

What you really mean is:

Can be enforced deterministically by system controls

Because security, compliance, etc. are enforcement mechanisms, not properties of the model.

A Cleaner Version of the Slide Logic Title

What Can Be Engineered Deterministically?

Deterministic System Controls

(Fully enforceable through software and governance)

Security Controls

Compliance Enforcement

Data Quality Validation

AI/Data Governance

Policy Enforcement

Model Behavior

(Probabilistic — must be constrained)

Predictable AI Behavior

Content Safety

Approach

Guardrails

Structured outputs

Validation layers

Retry budgets

Human approval gates

One Extremely Powerful Sentence for This Slide

This might actually be the core thesis of your book:

Reliable AI systems emerge when deterministic controls surround probabilistic models.

Why This Slide Is Strong

It makes a very important architectural claim:

AI reliability does not come from the model.

It comes from system design.

That’s the exact mental shift most enterprises haven’t made yet.