3.4: AI Tool Selection

Choosing the Right AI Tool

AI Tool Categories

1. Code Assistants (IDE-integrated)

• GitHub Copilot, Tabnine, Amazon CodeWhisperer
• Best for: Real-time code completion while typing

2. Chat-Based AI (Conversational)

• ChatGPT, Claude, Gemini
• Best for: Requirements analysis, architecture design, code review

3. Specialized AI (Domain-specific)

• Sourcegraph Cody (codebase search), Codium AI (test generation)
• Best for: Specific tasks (tests, documentation, refactoring)

Tool Comparison for ASPICE

GitHub Copilot

Strengths:

• [PASS] Fast (inline suggestions as you type)
• [PASS] Context-aware (uses nearby code context, ~2-4KB window)
• [PASS] IDE-integrated (VS Code, IntelliJ, Vim)
• [PASS] Good at boilerplate, common patterns

Weaknesses:

• [FAIL] Limited context (2-4KB window)
• [FAIL] No conversational mode (can't ask "why?")
• [FAIL] Weak at safety-critical logic
• [FAIL] May suggest insecure code (CWE vulnerabilities)

Best Use Cases:

• CAN message parsing
• Doxygen header generation
• Test scaffolding
• Simple getters/setters

Pricing: €10/month (individual), €19/month (business) (Pricing as of December 2025; check vendor website for current rates)

ChatGPT-4

Strengths:

• [PASS] Conversational (iterative refinement)
• [PASS] Good at requirements extraction
• [PASS] Can explain code
• [PASS] Multi-language support

Weaknesses:

• [FAIL] No IDE integration (copy-paste workflow)
• [FAIL] Context limit (128K tokens, ~50K LOC)
• [FAIL] May hallucinate (invents nonexistent APIs)
• [FAIL] Data privacy concerns (cloud-based)

Best Use Cases:

• Extract requirements from Word documents
• Generate ADR templates
• Explain legacy code
• Generate test cases from requirements

Pricing: €20/month (Plus), API pricing available (pay-per-token) (Pricing as of December 2025; check vendor website for current rates)

Claude Sonnet

Strengths:

• [PASS] Large context (200K tokens, ~80K LOC)
• [PASS] Accurate (fewer hallucinations)
• [PASS] Good at code review (can review entire file)
• [PASS] Strong reasoning (architecture decisions)

Weaknesses:

• [FAIL] No IDE integration
• [FAIL] Slower than Copilot (batch workflow)
• [FAIL] More expensive (API pricing)

Best Use Cases:

• Code review (large files)
• Refactoring (understands full context)
• Architecture decision analysis
• Requirements traceability (cross-file)

Pricing: €20/month (Pro), API pricing available (pay-per-token) (Pricing as of December 2025; check vendor website for current rates)

Tabnine

Strengths:

• [PASS] Privacy-focused (on-premise option)
• [PASS] Trained on C/embedded code
• [PASS] IDE-integrated

Weaknesses:

• [FAIL] Less accurate than Copilot
• [FAIL] Smaller model (fewer capabilities)

Best Use Cases:

• Teams with strict data privacy (defense, medical)
• Embedded C projects
• Air-gapped environments

Pricing: Free (basic), €12–39/month (pro/enterprise) (Pricing as of December 2025; check vendor website for current rates)

Selection Decision Matrix

By ASPICE Process

By Task Type

Tool Combination Strategy

The Hybrid Approach

Use Multiple Tools for Different Tasks

Example Workflow: The following diagram shows how different AI tools are selected for different tasks in a typical development workflow, matching each tool's strengths to the task at hand.

Benefit: Use each tool's strengths (speed vs accuracy vs context)

Cost-Benefit Analysis

ROI Calculation

Scenario: Team of 10 engineers, 2-year project

Costs:

• GitHub Copilot: €10/month × 10 engineers × 24 months = €2,400
• ChatGPT-4 Plus: €20/month × 10 engineers × 24 months = €4,800
• Total: €7,200

Benefits (35% productivity gain, conservative estimate):

• Base calculation: 10 engineers × 1,600 hours/year × 2 years = 32,000 hours
• Time saved: 35% × 32,000 hours = 11,200 hours
• Hourly rate assumption: €100/hour (fully burdened rate including salary, benefits, and overhead)
• Cost savings: 11,200 hours × €100/hour = €1,120,000
• Net benefit: €1,120,000 − €7,200 = €1,112,800

ROI: 155x return on investment

Assumptions and Caveats:

• Productivity gain: 35% (conservative from 35–55% range based on GitHub Copilot study; actual results vary by team and complexity)
• Hourly rate: €100/hour fully burdened (salary + benefits + overhead)
• Measurement: Time savings measured via developer self-reporting and task completion metrics
• Does not include human review overhead (estimated 20% of saved time)
• Safety-critical code requires additional review (reduces net gain to 25–30%)
• Individual results vary based on code complexity and engineer experience

Data Privacy and Security

Considerations for Safety-Critical Projects

Cloud-Based Tools (GitHub Copilot, ChatGPT, Claude):

• [WARN] Code sent to cloud servers
• [WARN] Data privacy concerns (GDPR, customer contracts)
• [WARN] IP leakage risk (proprietary algorithms)

On-Premise Tools (Tabnine Enterprise):

• [PASS] Code stays on-premise
• [PASS] No cloud transmission
• [PASS] GDPR-compliant

Best Practices:

1. Review contracts: Check if customer allows cloud-based AI tools
2. Data sanitization: Remove customer-specific data before using AI
3. Air-gapped fallback: For defense/classified projects, use on-premise only
4. Non-disclosure: Ensure AI tool provider has NDA

Evaluation Process

How to Choose for Your Project

Step 1: Define Requirements

Requirements for AI Tool:

• IDE integration? (Yes/No)
• On-premise required? (Yes/No)
• Budget: €X/month per engineer
• Privacy: GDPR-compliant?
• Language: C, Python, etc.
• Context needed: Small (code completion) vs Large (code review)

Step 2: Trial Period

• Test 2-3 tools for 1 month
• Measure: Time saved, defects caught, engineer satisfaction
• Compare cost vs benefit

Step 3: Team Feedback

• Survey engineers: "Which tool helps most?"
• Collect data: Code review time, defect density, test coverage

Step 4: Decision

• Choose tool(s) based on ROI, engineer preference, compliance
• Document decision (ADR if significant cost)

Summary

Tool Categories: Code assistants (Copilot), chat-based (ChatGPT, Claude), specialized (Codium, Sourcegraph)

Top Recommendations:

• Code completion: GitHub Copilot (fast, IDE-integrated)
• Requirements/Review: ChatGPT-4 or Claude Sonnet (large context)
• Privacy-sensitive: Tabnine (on-premise option)

Hybrid Strategy: Use multiple tools for different tasks (code, requirements, review)

ROI: 155x return (€7,200 cost → €1,112,800 benefit for 10-engineer team, based on conservative 35% productivity gain)

Privacy: Check customer contracts, sanitize data, consider on-premise for classified projects

Evaluation: Define requirements, trial period, team feedback, measure ROI

Chapter 35 Complete: Working with AI Assistants covers effective prompting, reviewing AI output, HITL decision-making, and tool selection.

Next: Chapter 36 — Putting It All Together (end-to-end example, common pitfalls, continuous learning)