Cost Effectiveness Framework
Reading time: Approximately 18 minutes
AIDLC adoption is not a technical transition but cost structure redesign. However, lack of actual data creates difficulties in RFP estimation, ROI justification, and budget securing. This document provides a practical framework for quantifying AIDLC cost effectiveness and reflecting it in project proposals.
1. RFP Cost Estimation Dilemma
1.1 Fixed-Price Bidding Market Reality
The Korean SI market is dominated by fixed-price bidding. Issuers present detailed RFPs, and bidders submit fixed amounts. Post-contract cost overrun risk is borne entirely by bidders.
Traditional Estimation Formula:
Total Cost = Σ(MM by Role × Monthly Rate × Duration)
+ Infrastructure Cost
+ Risk Buffer (10-20%)
Problems arising with AIDLC adoption:
-
How to quantify AI productivity?
- The fact that "AI generates code" alone cannot justify MM reduction
- Issuers expect "AI = automatic cost reduction", but bidders estimate conservatively due to lack of actual data
-
Additional cost items emerge:
- Ontology design: Initial 2-4 weeks
- Harness engineering: Continuous effort
- AI tool licensing: Claude Team ($30/user/month), LiteLLM Pro, etc.
- Training: 1-2 weeks for developer AIDLC transition training
-
Increased risk:
- AI output quality variability
- Harness integration complexity in legacy environments
- Organizational transition resistance
As a result, many bidders cannot reflect AIDLC cost savings in proposals and revert to traditional estimation methods.
1.2 Cost Reduction vs Quality Improvement Trade-off
AIDLC provides two types of value:
- Cost Reduction: Complete same scope with less effort
- Quality Improvement: Achieve higher quality/scope with same effort
In fixed-price bidding, cost reduction leads to bidder margin increase, but issuers prefer quality improvement. Proposals must clearly present this balance.
Example: ₩5B Project
| Scenario | Approach | Proposal | Actual Effort | Bidder Margin | Issuer Value |
|---|---|---|---|---|---|
| Traditional Method | No AIDLC | ₩5B | ₩5B | ₩1B (20%) | Meet basic requirements |
| Emphasize Cost Reduction | Apply AIDLC | ₩4B (-20%) | ₩3.2B | ₩0.8B (20%) | Meet basic requirements |
| Emphasize Quality Improvement | Apply AIDLC | ₩5B | ₩3.5B | ₩1.5B (30%) | High quality + additional features |
| Balance | Apply AIDLC | ₩4.5B (-10%) | ₩3.3B | ₩1.2B (27%) | Basic + partial high quality |
In most cases, the balance scenario is most realistic. Issuers realize cost reduction while receiving quality improvement guarantees.
2. AIDLC Cost Model Framework
2.1 Phase-by-Phase Cost Reduction Structure
AIDLC produces differentiated effects across RUP-based 4 phases:
Inception Phase
Traditional Method:
- Requirements analysis: 4 weeks
- Domain modeling: 2 weeks
- Architecture design: 3 weeks
- Total 9 weeks
AIDLC Method:
- AI requirements analysis: 1 week (AI auto-decomposes Intent → Unit)
- Ontology engineering: 2 weeks (Formalize domain model as ontology)
- Architecture design: 2 weeks (AI proposes reference architecture)
- Total 5 weeks (-44%)
Cost Reduction Mechanism:
- AI auto-detects requirements ambiguity and generates clarification questions
- Ontology structures domain knowledge for repeated reuse
- AI immediately proposes reference architecture patterns
Elaboration Phase
Traditional methods consume much time on prototype development and architecture verification. AIDLC has AI rapidly generate prototypes and automatically verify domain accuracy based on ontology.
Cost Reduction: Approximately 30-40%
Construction Phase
This phase concentrates 60-70% of project costs. It's AIDLC's core value delivery point.
Traditional Method:
- Developers write specification → code → unit test → review → fix
- Feedback loop: Daily basis
- Code generation speed: 100 LOC/day
AIDLC Method:
- AI auto-generates code based on ontology
- Harness provides runtime verification and immediate feedback
- Feedback loop: Minute basis
- Code generation speed: 500 LOC/day (+400%)
Cost Reduction: Approximately 40-60%
Cautions:
- Harness overhead increases in legacy environment integration
- AI-generated code quality varies for complex business logic
- Review effort decreases, but ontology accuracy verification effort added
Transition (Operations) Phase
Traditional Method:
- Manual deployment checklist
- Manual alert monitoring
- Manual diagnosis and recovery on failures
AIDLC Method:
- GitOps automatic deployment (Argo CD)
- AI Agent autonomous incident response
- 73% MTTR reduction (45min → 12min)
Cost Reduction: Approximately 50-70% (operations personnel optimization)
2.2 Cost Increase Factors
AIDLC adoption doesn't always bring cost reduction only. The following factors increase costs:
Initial Investment
| Item | Scale | Cost |
|---|---|---|
| Ontology design | 2-4 weeks (1 architect + 0.5 domain expert) | ₩50-100M |
| Initial harness engineering setup | 1-2 weeks (2 DevOps) | ₩20-40M |
| AIDLC training | 10 developers × 1 week | ₩30M |
| AI tool licensing (first 3 months) | Claude Team 10 people × $30 × 3 months | ₩10M |
| Total | ₩110-180M |
For medium projects (₩2B), initial investment ratio is 5.5-9%. This cost is concentrated during the first Bolt cycle (2-4 weeks), after which savings accumulate.
Ongoing Costs
- AI tool licensing: ₩3M/month (10 developers basis)
- Ontology maintenance: 4 hours/week (0.1 MM architect)
- Harness operations: 8 hours/week (0.2 MM DevOps)
For 6-month projects, ongoing costs add approximately ₩70-100M.
Break-Even Point
Point where initial investment + ongoing costs are offset by savings:
Small Projects (<₩1B): 2-3 months later
Medium Projects (₩1-5B): 1-2 months later
Large Projects (₩5B+): 1 month later
Break-even point comes faster with larger projects. This is because AIDLC produces greater effects on large-scale repetitive work.
3. Expected Benefits by Project Scale
3.1 Small Project (Under ₩1B)
| Item | Traditional Cost | AIDLC Cost | Savings Rate |
|---|---|---|---|
| Total MM | 50 MM | 40 MM | 20% |
| Total Cost | ₩800M | ₩650M | 18.75% |
| Initial Investment | 0 | ₩120M | +₩120M |
| Net Savings | - | ₩30M | 3.75% |
Characteristics:
- Initial investment weight is high, limiting savings effect
- Inception phase compression effect largest (requirements analysis time reduction)
- Harness setup overhead relatively high
Recommendations:
- Design ontology lightweight (focus on core entities)
- Implement harness essential validation only
- Reuse existing reference architecture
3.2 Medium Project (₩1-5B)
| Item | Traditional Cost | AIDLC Cost | Savings Rate |
|---|---|---|---|
| Total MM | 250 MM | 175 MM | 30% |
| Total Cost | ₩4B | ₩2.8B | 30% |
| Initial Investment | 0 | ₩150M | +₩150M |
| Net Savings | - | ₩1.05B | 26.25% |
Characteristics:
- Construction phase acceleration effect materializes
- Ontology reuse effect accumulates
- Harness investment vs ROI clear
Recommendations:
- Design ontology systematically (cover entire domain)
- Expand harness progressively (core → entire)
- Actively adopt AI review automation
3.3 Large Project (₩5B+)
| Item | Traditional Cost | AIDLC Cost | Savings Rate |
|---|---|---|---|
| Total MM | 600 MM | 400 MM | 33.3% |
| Total Cost | ₩10B | ₩6.5B | 35% |
| Initial Investment | 0 | ₩200M | +₩200M |
| Ongoing Cost | 0 | ₩150M | +₩150M |
| Net Savings | - | ₩3.15B | 31.5% |
Characteristics:
- Operations automation effect accumulates (long-term operations cost reduction)
- Ontology consistency effect maximized in multi-team parallel development
- Service-to-service harness integration effect in MSA environment
Recommendations:
- Expand ontology to enterprise level
- Integrate harness with service mesh
- Actively adopt AI Agent autonomous operations
3.4 Complex Program (₩10B+)
Complex programs with multiple bundled projects achieve additional effects through ontology reuse and harness platformization.
Example: Bank Next-Gen System (₩30B, 3 years)
| Phase | Traditional Cost | AIDLC Cost | Savings Rate |
|---|---|---|---|
| Phase 1 (Accounting) | ₩10B | ₩6.8B (-32%) | 32% |
| Phase 2 (Loans) | ₩10B | ₩6.0B (-40%) | 40% |
| Phase 3 (Deposits) | ₩10B | ₩5.5B (-45%) | 45% |
| Total | ₩30B | ₩18.3B | 39% |
Savings rate increases as phases progress due to:
- Ontology accumulates, increasing reuse rate
- Harness becomes platformized, reducing setup time
- Team AIDLC proficiency improves
4. Ontology ROI
Ontology is AIDLC's core investment item. Quantifying initial investment vs long-term effects.
4.1 Initial Investment
| Activity | Effort | Cost (₩10M/month basis) |
|---|---|---|
| Domain analysis | 1 week (1 architect) | ₩2.5M |
| Entity/relationship modeling | 1 week (1 architect) | ₩2.5M |
| Define validation rules | 1 week (0.5 architect + 0.5 domain expert) | ₩2.5M |
| Ontology documentation | 1 week (1 technical writer) | ₩2M |
| Total | 4 weeks | ₩9.5M |
For medium projects (₩2B), this is 4.75% initial investment.
4.2 Long-Term Effects
According to McKinsey research "The economic potential of generative AI" (2023), domain-specific AI achieved 87% accuracy vs 62% for general AI. Ontology is the mechanism that transforms AI to domain-specific.
Error Rate Reduction
Hamel & Patil (2024) "The Strawberry Manifesto" experimental data:
| Condition | Error Rate (Day 31) | Total Cost | Improvement |
|---|---|---|---|
| No Feedback Loop (Baseline) | 8.3% | $25K | - |
| Feedback Loop (AI only) | 7.9% | $28K | 5% improvement |
| Feedback Loop + Ontology | 1.2% | $30K | 85.5% improvement |
With ontology, AI self-corrects in domain context, reducing error rate 7x.
Project Conversion:
- Medium project defect fix cost: Average ₩300M (7.5% of total cost)
- 85% error reduction with ontology: ₩255M savings
- ROI vs ₩9.5M ontology initial investment: 2,584%
Review Effort Savings
When AI generates domain-accurate code based on ontology, reviewers focus on business logic verification.
| Review Item | Without Ontology | With Ontology | Savings Rate |
|---|---|---|---|
| Domain consistency verification | 30 min | 5 min | 83% |
| Coding convention verification | 15 min | 3 min | 80% |
| Security/performance verification | 20 min | 15 min | 25% |
| Business logic verification | 30 min | 30 min | 0% |
| Total | 95 min | 53 min | 44% |
For medium projects with 500 PRs:
- Traditional review time: 500 × 95min = 791 hours = 4.4 MM
- AIDLC review time: 500 × 53min = 442 hours = 2.5 MM
- Savings: 1.9 MM (approximately ₩30M)
4.3 Ontology Maintenance Costs
Ontology is not static. Continuous updates needed as requirements change.
| Activity | Frequency | Effort/time | Monthly Effort |
|---|---|---|---|
| Add new entities | Weekly | 2 hours | 8 hours |
| Modify relationships | Bi-weekly | 1 hour | 2 hours |
| Add validation rules | Bi-weekly | 1 hour | 2 hours |
| Total | 12 hours/month (0.3 MM) |
6-month project maintenance cost: 1.8 MM (approximately ₩30M)
Net ROI Calculation:
- Initial investment: ₩9.5M
- Maintenance (6 months): ₩30M
- Total investment: ₩39.5M
- Effect (error savings + review savings): ₩255M + ₩30M = ₩285M
- Net ROI: 621%
5. Harness ROI
Harness automates feedback loop between AI and operating environment. Without harness, runtime errors accumulate and can derail projects.
5.1 Cost of Harness Absence: Fintech Runaway Case
Actual case introduced in Hamel & Patil (2024) "The Strawberry Manifesto":
Scenario: AI Agent develops email reminder feature for fintech app
Developing without harness:
- AI generates code → Manual test → Deploy
- Runtime error occurs → AI checks error log → Fix → Redeploy
- Slow feedback loop causes AI to repeat same mistakes
Result:
- API calls: 847 times (7x normal)
- LLM cost: $2,200 (normal $300)
- Generated emails: 14 (all incomplete)
- Project failure
After harness introduction:
- Harness provides immediate feedback of runtime errors to AI
- API calls: 123 times
- LLM cost: $320
- Generated emails: 50 (all normal)
- Project success
Cost Savings: $1,880 (85%)
5.2 Harness Introduction Effects
| Metric | Without Harness | With Harness | Improvement |
|---|---|---|---|
| Runtime failure rate | 15/week | 3/week | 80% ↓ |
| Incident MTTR | 4 hours | 45 min | 81% ↓ |
| Rollback ratio | 12% | 2% | 83% ↓ |
| Emergency patch frequency | 8/month | 1/month | 87% ↓ |
Project Conversion:
- Medium project runtime failure response cost: Average ₩200M (5% of total cost)
- 80% failure reduction with harness: ₩160M savings
5.3 Harness Investment Costs
| Item | Effort | Cost |
|---|---|---|
| Initial setup (CI/CD integration) | 1 week (2 DevOps) | ₩20M |
| Implement runtime verification logic | 1 week (1 DevOps + 1 developer) | ₩20M |
| Ongoing operations (6 months) | 8 hours/week | ₩60M |
| Total | ₩100M |
Net ROI: (₩160M - ₩100M) / ₩100M = 60%
6. Open Weight Model TCO Comparison
AIDLC supports both cloud APIs like Claude/GPT and open weight models like GLM/Qwen. TCO varies significantly based on choice.
6.1 Cloud API Costs
| Model | Input Token Price | Output Token Price | Context Size |
|---|---|---|---|
| Claude Sonnet 4.0 | $3/MTok | $15/MTok | 200K |
| GPT-4.1 | $5/MTok | $15/MTok | 128K |
| GPT-3.5 Turbo | $0.5/MTok | $1.5/MTok | 16K |
Medium Project Usage Estimate (6 months, 10 developers):
- Monthly requests: 10 people × 100/day × 20 days = 20,000
- Average input tokens: 2,000 (context + prompt)
- Average output tokens: 1,000 (code generation)
- Monthly tokens: (20,000 × 2,000) + (20,000 × 1,000) = 60 MTok
Monthly Cost:
- Claude Sonnet 4.0: (40 MTok × $3) + (20 MTok × $15) = $420 (approximately ₩5.5M)
- GPT-4.1: (40 MTok × $5) + (20 MTok × $15) = $500 (approximately ₩6.5M)
- 6-month total: $2,520-$3,000 (approximately ₩33-39M)
Advantages:
- No initial cost
- No infrastructure management needed
- Immediate use of latest models
Disadvantages:
- Cost explosion with increased usage
- External data transmission (unsuitable for confidential projects)
- Network latency (400-800ms)
6.2 Self-Hosting Costs
Infrastructure: EKS + GPU nodes (vLLM-based)
| Component | Spec | Monthly Cost |
|---|---|---|
| GPU instance | p5.48xlarge (H200×8) Spot | $20,000 (approximately ₩26M) |
| Storage | 1TB EBS gp3 | $100 (approximately ₩1.3M) |
| Network | Data transfer | $200 (approximately ₩2.6M) |
| Operations personnel | 0.5 MLOps engineer | $5,000 (approximately ₩6.5M) |
| Total | $25,300 (approximately ₩33M/month) |
6-month total cost: $151,800 (approximately ₩200M)
Break-Even Point Calculation:
Point where cloud API cost exceeds self-hosting:
Monthly requests = $25,300 / (average cost per request)
Claude Sonnet 4.0: $0.09/request
→ 281,111 requests/month (140 developer scale)
GPT-3.5 Turbo: $0.006/request
→ 4,216,667 requests/month (2,100 developer scale)
Conclusion:
- Small projects (10 developers): Cloud API advantageous
- Medium projects (50 developers): Depends on situation
- Large projects (100+ developers): Self-hosting advantageous
- Confidential projects: Self-hosting mandatory
6.3 Hybrid Strategy
In practice, cloud API + self-hosting hybrid is optimal:
| Task Type | Model Selection | Reason |
|---|---|---|
| Complex architecture design | Claude Sonnet 4.0 (cloud) | High-quality reasoning needed |
| Repetitive code generation | GLM-5/Qwen (self-hosted) | High volume requests, low latency |
| Code review | GPT-4.1 (cloud) | Deep analysis needed |
| Unit test generation | GLM-4 (self-hosted) | Mass generation |
Cost Optimization Effect:
- 60% cloud API usage reduction
- 40% self-hosting efficiency improvement
- Total cost 30-40% savings
7. Productivity Metrics
Core metrics for measuring AIDLC adoption effects.
7.1 AIDLC Productivity Metrics
📈 AIDLC Productivity Metrics
Before and After AI Adoption
7.2 Detailed Measurement Items and DORA Mapping
📊 Metrics
Measuring AIDLC Adoption Impact
Key Metrics
DORA Metrics Mapping
7.3 Metric Collection Methods
Development Productivity
Data Sources:
- GitHub/GitLab metrics: Commit count, PR review time, deployment frequency
- JIRA/Linear: Ticket processing speed, backlog burn-down
- AI tool logs: LLM call count, generated code LOC, acceptance rate
Measurement Cycle: Weekly (aligned with Bolt cycle)
Operational Stability
Data Sources:
- AMP/AMG: MTTR, error rate, SLO achievement rate
- Argo CD: Deployment success rate, rollback frequency
- PagerDuty: Incident count, escalation rate
Measurement Cycle: Daily (real-time dashboard)
Cost Efficiency
Data Sources:
- AWS Cost Explorer: Infrastructure costs
- LLM provider dashboard: API costs
- Timesheets: Actual effort invested
Measurement Cycle: Monthly
7.4 Benchmark Data
Based on McKinsey "The economic potential of generative AI" (2023):
| Industry | Productivity Improvement | ROI Period | Note |
|---|---|---|---|
| Software Development | 35-45% | 3-6 months | Code generation focus |
| Financial Services | 20-30% | 6-12 months | Compliance burden |
| Manufacturing | 15-25% | 12-18 months | Legacy system integration |
Korean SI market sets 20-30% range as realistic goal similar to financial services.
8. RFP Proposal Writing Guide
8.1 Cost Reduction Statement Strategy
Wrong Approach:
"We will improve development productivity by introducing AI."
This expression is vague and unmeasurable. Issuers want specific numbers.
Correct Approach:
"We optimize costs as follows by applying AIDLC methodology:
- 44% Inception Phase Reduction: 9 weeks → 5 weeks with AI requirements analysis
- 40% Construction Phase Acceleration: 400% developer productivity improvement with ontology-based code auto-generation
- 50% Operations Cost Reduction: 73% MTTR reduction with AI Agent autonomous incident response
Total proposal: ₩4.5B (10% reduction vs traditional method)
However, ₩150M initial AIDLC investment is included, which will create ₩285M error reduction effect throughout the project (90% net ROI)."
8.2 Risk Mitigation Strategy
AIDLC adoption entails new risks. Specify mitigation strategies in proposal.
| Risk | Mitigation Strategy | Responsibility |
|---|---|---|
| AI-generated code quality variation | Ontology-based auto-verification + 2-stage review | Bidder |
| Harness integration complexity | Pre-analyze legacy environment + phased application | Bidder |
| Organizational transition resistance | 4-week training program + pilot project | Issuer+Bidder |
| Data leakage concerns | Self-hosted LLM + network isolation | Bidder |
8.3 Phased Application Roadmap
Applying AIDLC to all phases at once is risky. Present phased approach.
Phase 1: Pilot (First Bolt, 2-4 weeks)
- Goal: Verify AIDLC feasibility
- Scope: 1 core module
- Ontology: Lightweight design
- Harness: Basic verification only
- Expected effect: 15-20% productivity improvement
Phase 2: Scale-up (2-3 months)
- Goal: Expand to entire team
- Scope: 5 major modules
- Ontology: Cover entire domain
- Harness: Integration test automation
- Expected effect: 25-35% productivity improvement
Phase 3: Optimization (Late project)
- Goal: Maximize ROI
- Scope: Entire system
- Ontology: Continuous improvement
- Harness: AI Agent autonomous operations
- Expected effect: 30-40% productivity improvement
8.4 Proposal Checklist
- Specify concrete savings rate vs traditional method (%)
- Specify initial investment cost and present ROI period
- Quantify ontology/harness investment vs long-term effects
- Specify break-even point by project scale
- Present rationale for cloud API vs self-hosting choice
- Attach phased application roadmap
- Detail risk mitigation strategies
- Specify measurement metrics and collection methods
- Include reference project cases (if available)
- Include issuer training plan
9. Next Steps
After understanding cost effectiveness framework, proceed to next documents:
- Adoption Strategy: Strategy for phased AIDLC organizational adoption
- Role Composition: How PM, architect, developer roles change
- Governance Framework: Methods for enterprise-wide AI-generated code quality management
Also reference methodology guides to prepare for actual implementation:
- Ontology Engineering: Practical guide for structuring domain knowledge
- Harness Engineering: Methods for automating runtime feedback loops
- Open Weight Models: Self-hosted LLM deployment guide
References
- McKinsey Global Institute. "The economic potential of generative AI: The next productivity frontier." 2023.
- Hamel, Jeremy & Patil, DJ. "The Strawberry Manifesto: How to Build AI Products That Work in the Real World." 2024.
- Forsgren, Nicole et al. "Accelerate: The Science of Lean Software and DevOps." 2018. (DORA metrics source)
- AWS. "Building Generative AI applications using Amazon Bedrock." 2024.
- DeepLearning.AI. "Building and Evaluating Advanced RAG Applications." Andrew Ng. 2024.