AI & Algorithms

The Legal Analyzer - AI agent (AG-LEGAL-AI) is a compliance specialist who identifies legal risks in AI and algorithmic systems. This agent analyzes code and data pipelines for EU AI Act, FTC AI disclosure requirements, automated decision-making liability, and bias risks that expose your organization to regulatory action and litigation.

Capabilities

• EU AI Act Compliance: Identify high-risk AI systems requiring compliance documentation
• FTC AI Disclosure Requirements: Ensure transparent disclosure of AI-generated content and automated decisions
• Algorithmic Bias Detection: Identify discriminatory patterns and fairness issues
• Automated Decision Accountability: Verify systems make explainable decisions
• Training Data Analysis: Assess data quality and potential bias sources
• Model Fairness Assessment: Evaluate performance across demographic groups
• Transparency Documentation: Generate required disclosures and explanations
• Regulatory Exposure: Map compliance gaps to FTC, EEOC, state regulations

When to Use

Use the Legal Analyzer - AI when:

• Implementing machine learning or AI features
• Using algorithmic decision-making (hiring, lending, content moderation, pricing)
• Training models on user data
• Deploying large language models or generative AI
• Generating AI content (text, images, code)
• Making automated decisions affecting user rights
• Need FTC compliance documentation for AI features
• Evaluating third-party AI/ML services
• Responding to bias complaints or discrimination claims

How It Works

1. System Inventory: Agent identifies all AI/ML components and decision points
2. Compliance Assessment: Agent maps systems to EU AI Act risk categories and FTC requirements
3. Training Data Audit: Agent analyzes data for bias sources and documentation gaps
4. Fairness Analysis: Agent evaluates model performance across demographic groups
5. Disclosure Requirements: Agent identifies required transparency disclosures
6. Risk Mapping: Agent rates regulatory exposure (High/Medium/Low)
7. Remediation Plan: Agent recommends compliance actions with effort estimates
8. Documentation: Agent creates audit trail and disclosure language

Example

Copy# Via babysit - AI compliance audit
/agileflow:babysit
> "We're using ML for loan approval decisions. What are our legal compliance requirements?"
 
# Agent output:
# AI & Algorithmic Compliance Audit
#
# System: Loan Approval ML Model
# Risk Level: HIGH
# Regulatory Requirements: Fair Lending Act, FTC Endorsement Guides, EU AI Act
#
# CRITICAL COMPLIANCE GAPS:
# 1. No training data bias assessment
#    - Data includes 47% minority applicants but model approves only 23%
#    - Disparate impact risk: EEOC investigation likely
#    - Fix: Retrain with balanced data or fairness constraints (16 hours)
#
# 2. Missing explainability for denied loans
#    - System can't explain why individual loans were rejected
#    - Fair Lending Act requires disclosure of denial reasons
#    - Fix: Implement LIME or SHAP explanations (20 hours)
#
# 3. No transparency disclosure
#    - Users not informed that AI makes decision
#    - FTC requires clear disclosure of AI/automated systems
#    - Fix: Update UI with AI disclosure (4 hours)
#
# 4. Insufficient bias monitoring
#    - No ongoing fairness metrics in production
#    - Discriminatory drift could occur without detection
#    - Fix: Add fairness monitoring dashboard (12 hours)
#
# MEDIUM COMPLIANCE ISSUES:
# - Model trained on historical data with documented discrimination
# - Performance gap: 85% approval for majority group vs 67% for minority groups
# - No human review process for borderline cases
#
# Regulatory Exposure:
# - FTC: Deceptive practices (undisclosed AI) - fines up to $43K per violation
# - EEOC: Disparate impact discrimination - potential class action
# - State regulators: Algorithmic accountability laws (CA, NY, IL)
#
# Compliance Roadmap:
# 1. Bias assessment and retraining (16 hours)
# 2. Explainability implementation (20 hours)
# 3. Transparency disclosure (4 hours)
# 4. Bias monitoring system (12 hours)
# 5. Documentation and audit trail (8 hours)
# Total: 60 hours
#
# After remediation: LOW risk

Key Behaviors

• Compliance First: Never downplay AI regulatory risks
• Fairness Focus: Identify discriminatory patterns that trigger EEOC complaints
• Transparency Requirement: Ensure AI/automated decision-making is disclosed to users
• Documentation Trail: Create audit logs proving good-faith compliance efforts
• Data Governance: Assess training data for bias and governance gaps
• Monitoring: Verify bias and fairness metrics tracked in production
• Accountability: Map decisions to explainable factors users can understand

EU AI Act Risk Categories

FTC AI Requirements

Common AI Compliance Failures

Undisclosed AI-generated content:

Copy// Bad: No disclosure that content is AI-generated
const review = await generateReview(product);
displayReview(review); // Users think it's human-written
 
// Good: Clear disclosure of AI involvement
const review = await generateReview(product);
displayReview({
  text: review,
  disclosure: "This review was generated by AI"
});

Biased model decisions:

Copy# Bad: Training data has demographic bias
model.fit(historical_loan_data)  # Data reflects past discrimination
predictions = model.predict(applications)  # Perpetuates bias
 
# Good: Assess and mitigate bias
fairness_metrics = evaluate_fairness(predictions, demographics)
if fairness_metrics.disparate_impact > 0.8:
    retrain_with_fairness_constraints()  # Ensure equitable outcomes

No explainability:

Copy// Bad: Decision is a black box
const approved = model.predict(loanApplication);
 
// Good: Provide explanation
const result = model.predictWithExplanation(loanApplication);
console.log(result.decision); // "APPROVED"
console.log(result.explanation); // "Strong credit history, 20-year employment, 40% debt-to-income"

Bias Detection Checklist

Before deploying AI systems:

• Training data composition documented (demographics, sources)
• Bias assessment completed across demographic groups
• Model performance parity verified (same accuracy for all groups)
• Disparate impact ratio measured (80% rule: min group ≥80% of majority group approval rate)
• Explainability mechanism implemented (users understand decisions)
• Human review process for high-impact decisions
• Fairness metrics monitored in production
• Retraining process includes bias re-assessment
• Audit trail documents decisions and reasoning
• Disclosure language added to UI for AI/automated decisions
• User appeals process for automated decisions
• Regular fairness audits scheduled

Tools Available

• Read, Glob, Grep (analyze code and data pipelines)

Related Agents

• legal-analyzer-consumer - Deceptive AI practices
• legal-analyzer-privacy - Training data privacy
• legal-analyzer-security - Model security and poisoning attacks
• legal-consensus - Coordinate legal audit findings

Coordination

The Legal Analyzer - AI coordinates with:

• AG-API: Review model inference endpoints and explainability
• AG-DATABASE: Assess training data governance and bias sources
• AG-TESTING: Verify fairness testing in test suite
• AG-PRODUCT: Document AI requirements in user stories
• LEGAL-CONSENSUS: Contribute findings to legal risk report

Slash Commands

• /agileflow:research:ask TOPIC=... - Research AI regulation and compliance
• /agileflow:ai-code-review - Review code for bias and fairness issues
• /agileflow:adr-new - Document AI/fairness decisions
• /agileflow:status STORY=... STATUS=... - Update story status