Architecture Comparator

Compare the risk profiles of different delegation architectures to make informed design decisions.

How It Works

This tool provides side-by-side comparison of delegation architectures:

1. Component Breakdown: See which AI/human components each architecture uses
2. Risk Visualization: Compare expected vs. worst-case risk
3. Mitigation Impact: See how different safety measures affect overall risk
4. Trade-off Analysis: Understand the cost-benefit of each approach

---

Architecture Comparison

Compare risk profiles of different delegation architectures. Darker bars show expected monthly risk; lighter bars show worst-case potential.

Baseline (Human Only)

Traditional human-driven process with no AI delegation

$72/mo

expected

Simple AI Assist

AI provides suggestions, human makes all decisions

$79/mo

expected

Autonomous AI

AI handles routine tasks, human handles exceptions

$151/mo

expected

Full Automation

End-to-end AI with minimal human intervention

$602/mo

expected

Trade-off Analysis

Architecture	Expected Risk	Worst Case	Components	Mitigations
Baseline (Human Only)	$72/mo*	$2,000	1	2
Simple AI Assist	$79/mo	$2,500	2	2
Autonomous AI	$151/mo	$4,800	3	3
Full Automation	$602/mo	$8,000	3	3

Component Type Legend

Deterministic(~0.1% base)

Narrow ML(~3.0% base)

General LLM(~10.0% base)

RL Agent(~20.0% base)

Human(~5.0% base)

---

Default Architectures Explained

Baseline (Human Only)

Traditional human-driven process with no AI delegation. Establishes the risk level you’re comparing against.

Characteristics:

• Single point of failure (human error)
• Predictable but expensive
• Limited scalability
• Well-understood failure modes

Simple AI Assist

AI provides suggestions and recommendations, but humans make all decisions. Common for high-stakes domains.

Characteristics:

• AI errors caught by human review
• Slower than autonomous but safer
• Good for building trust in AI systems
• Human bottleneck remains

Autonomous AI

AI handles routine tasks independently; humans handle exceptions. Balances efficiency with safety.

Characteristics:

• Higher throughput for routine work
• Complex failure modes (routing errors)
• Requires robust exception handling
• Multiple points of mitigation

Full Automation

End-to-end AI with minimal human intervention. Maximum efficiency, maximum risk.

Characteristics:

• Highest potential damage
• Requires extensive mitigation
• Suitable only for well-understood domains
• Fastest degradation if poorly designed

---

Interpreting the Comparison

Expected vs. Worst Case

• Expected Risk (dark bar): Average monthly risk given probability distributions
• Worst Case (light bar): Maximum possible damage if everything fails

A wide gap indicates high tail risk.

Component Types

Risk varies significantly by component type:

Type	Base Failure Rate	Typical Use
Deterministic	~0.1%	Rule-based routing, validation
Narrow ML	~3%	Classification, detection
General LLM	~10%	Generation, reasoning
RL Agent	~20%	Autonomous decision-making
Human	~5%	Review, exception handling

Mitigation Stacking

Each mitigation reduces risk multiplicatively. With 3 mitigations at 85% effectiveness each:

final_risk = base_risk × 0.85³ = base_risk × 0.61

More mitigations help, but with diminishing returns.

---

When to Choose Each Architecture

Choose Human Only When:

• Stakes are extremely high
• Decisions require nuanced judgment
• AI systems aren’t well-calibrated for your domain
• Regulatory requirements demand human oversight

Choose AI Assist When:

• AI can improve human decision quality
• Speed is important but not critical
• Building organizational trust in AI
• Failure costs are moderate

Choose Autonomous AI When:

• High volume of routine decisions
• Clear criteria for “routine” vs “exception”
• Good monitoring and fallback in place
• Moderate-to-high risk tolerance

Choose Full Automation When:

• Domain is well-understood with clear boundaries
• Extensive testing and validation completed
• Strong mitigation stack in place
• Benefits significantly outweigh risks

---

Migration Paths

Human → AI Assist

1. Start with AI suggestions for low-stakes decisions
2. Track AI accuracy vs. human decisions
3. Gradually expand scope based on performance
4. Maintain human review throughout

AI Assist → Autonomous

1. Identify truly routine tasks (over 95% predictable)
2. Implement robust exception detection
3. Add monitoring and alerting
4. Pilot with limited scope, then expand

Autonomous → Full Automation

1. Achieve consistent performance metrics
2. Reduce human exception handling rate to less than 5%
3. Implement comprehensive mitigation stack
4. Establish clear rollback procedures

---

Customization

The default architectures serve as templates. To analyze your specific situation:

1. Use the Risk Calculator to model each architecture
2. Use the Sensitivity Dashboard to identify key parameters
3. Use the Trust Updater to calibrate component reliability
4. Document your analysis in the a decomposition worksheet