The Model
- Framework
The Three-Layer Model
A decision-systems framework for AI adoption: automate execution, protect judgment, embed compliance as the architectural foundation. This page explains each layer in depth, with implementation guidance and failure patterns to watch.
Why This Model Exists
- Layer 1: Execution
- Layer 2: Judgment
- Layer 3: Compliance Foundation
- How the Layers Interact
- Implementation Principles
Why This Model Exists
Most AI adoption conversations collapse three distinct concerns into one bucket: doing work, making decisions, and being accountable. Vendors talk about “AI-powered workflows” as if the technology handles all three seamlessly.
The reality is messier. AI is genuinely excellent at some things and genuinely dangerous at others. The things it’s excellent at (execution) and the things it’s dangerous at (judgment) are easy to confuse because they often look similar from the outside.
The Three-Layer Model exists to make these distinctions explicit. It gives organizations a vocabulary for talking about what AI should and shouldn’t do.
- Layer 1
Execution
The execution layer is where AI genuinely shines. These are tasks that are repeatable, definable, and verifiable—where the goal is throughput and consistency rather than judgment and creativity.
Characteristics
Definable: You can specify what “correct” looks like before the task runs.
Repeatable: The task is performed frequently, with similar inputs producing similar outputs.
Verifiable: You can check whether the output is correct without exercising professional judgment.
Examples Across Industries
Immigration Consulting
Extracting data from intake forms, validating document formatting, routing cases by visa category, generating checklists, scheduling appointments.
Legal Services
Extracting key dates and parties from contracts, organizing discovery documents, generating first drafts of standard agreements.
Healthcare Administration
Processing insurance claims, extracting patient data from forms, scheduling based on availability, routing referrals.
Financial Services
Processing transaction data, generating account statements, routing customer inquiries, initial KYC data collection.
Implementation Guidelines
Define success criteria explicitly.
Before automating any task, write down what "correct" looks like.
Build validation into the workflow.
AI outputs should be validated automatically wherever possible.
Monitor for drift.
AI performance can degrade over time as inputs change.
Preserve human override.
Even in pure execution tasks, humans should be able to override AI outputs.
- Layer 2
Judgment
The judgment layer is where human authority must be preserved. These are tasks that require context, discretion, and accountability.
Characteristics of Judgment Tasks
Contextual: The right answer depends on circumstances that vary case by case.
Discretionary: Reasonable professionals might reach different conclusions.
Accountable: Someone must stand behind the decision.
How AI Should Support (Not Replace) Judgment
AI has a legitimate role in the judgment layer—but that role is support, not substitution.
Information Surfacing
AI can pull relevant information from large datasets. The human still evaluates.
Option Generation
AI can suggest possible approaches or flag considerations that might be overlooked.
Risk Flagging
AI can identify patterns that suggest elevated risk.
Consistency Checking
AI can compare current decisions against past decisions for similar cases.
Implementation Guidelines
Make AI support explicit.
Label AI-generated information clearly.
Require active decision.
Design workflows that require explicit human judgment.
Log what humans actually do.
Track whether humans modified AI suggestions.
Set minimum engagement thresholds.
For high-stakes decisions, require minimum review times.
- Layer 3
Compliance Foundation
The compliance layer is fundamentally different. It's not a set of tasks—it's a set of properties that the system must have.
Core Properties
Traceability: Every significant action can be traced to its source.
Auditability: The trace record is structured for review. Logs are immutable.
Decision Rights: The system knows who is authorized to do what.
Escalation Paths: When the system encounters uncertainty, it routes to human judgment.
Minimum Viable Compliance Controls
Decision Logging
Every decision logged with timestamp, actor, inputs, and outcome
AI Output Tracking
Every AI-generated output tagged with model/version and confidence score
Human Override Recording
When humans modify AI outputs, changes logged with reasons
Approval Gates
Judgment-layer decisions pass through explicit approval with verified identity
Escalation Triggers
System automatically escalates when defined conditions are met
Reconstruction Capability
Any significant decision can be fully reconstructed within 24 hours
- Immutable Logs: Audit logs are tamper-evident—cannot be modified after the fact
How the Layers Interact
A typical workflow touches all three layers, and the interactions between layers are where many problems emerge.
Execution:
Client submits intake form. AI extracts data, validates formatting, routes to appropriate queue.
Judgment:
Professional reviews the case for complexity indicators. Decides whether to accept engagement.
Compliance (throughout):
System logs intake receipt, routing decision, professional review with timestamp and identity, accept/reject decision with reasoning.
Layer Bleed: The Primary Risk
The biggest risk is “layer bleed”—when execution activities silently expand into judgment territory without anyone adjusting the compliance constraints. For example: An AI that “extracts data from forms” starts to “assess case complexity” because complexity indicators are extractable. A judgment function has been embedded in an execution wrapper.
For example: An AI that “extracts data from forms” (execution) starts to “assess case complexity” (judgment) because complexity indicators are extractable. The complexity assessment is still treated like data extraction—no human review required, minimal logging—but it now affects how cases are routed. A judgment function has been embedded in an execution wrapper.
Layer Bleed
Boundaries blur silently
Implementation Principles
Map Before You Automate
Before introducing AI to any workflow, map it explicitly. Identify which tasks are execution and which are judgment.
Principle 1
Automate Execution First
Start with pure execution tasks. Get the automation working well. Build confidence.
Principle 2
Protect Judgment Explicitly
When AI enters judgment-adjacent territory, add explicit protections: clear labeling, required human decisions, logging.
Principle 3
Embed Compliance from Day One
Build logging, audit trails, and escalation triggers from the start. Retrofitting compliance is expensive and incomplete.
Principle 4
Monitor and Audit Continuously
Track human override rates. Monitor review times. Test reconstruction capability. The system will drift if you don't watch it.
Principle 5
- Your Next Step
Let's Build Your Advantage
If you are ready to move beyond discussion and start implementing intelligent solutions that deliver a measurable impact, let's talk. I am selective about the projects I take on, focusing on partnerships where I can create significant, lasting value.