Deliverable 4.6 — AI Usage Declaration
Requirement: AI tools used, sections AI-assisted, what was manually validated, how to prevent blind AI usage
Source: AI Strategy.md, Analysis v2.md, Strategy.md, Project Management.md
Note: This is the deliverable with the highest strategic value — demonstrating AI leadership, not just AI usage
1. AI Tool Used During This Exercise
| Tool |
Model |
Usage |
Scope |
| GitHub Copilot |
Claude Opus 4.6 (VS Code Agent Mode) |
Sole AI tool — conversational agent driving all document creation, analysis, cross-referencing, website development |
Entire assessment |
No other AI tools were used. All work happened inside a single VS Code workspace through sequential conversational interactions with GitHub Copilot in Agent Mode.
2. How We Actually Worked — Step by Step
Rather than claiming vague "AI-assisted drafting", here is the exact workflow from start to finish:
| Step |
What Happened |
Human Role |
AI Role |
| 1 |
Requirements extraction |
Provided assessment brief PDF, directed scope |
Parsed requirements into structured Requirement.md |
| 2 |
Strategy & analysis |
Guided focus areas, set project constraints |
Generated Strategy.md, Analysis.md with frameworks |
| 3 |
Deliverable documents (4.1–4.6) |
Directed each doc creation, reviewed output, requested corrections |
Drafted all 6 deliverables with architecture, timelines, failure models |
| 4 |
Supporting analysis docs |
Identified what supporting analysis was needed, prioritized |
Created 15+ supporting docs (tech stack, planning, security, testing, cost, observability, etc.) |
| 5 |
Cross-document consistency |
Spotted capacity math mismatch (different numbers across files), directed sync |
Audited all 33 files, found and fixed inconsistencies across 7 documents |
| 6 |
Gap analysis |
Asked "what's missing for a complete submission?" |
Identified 5 gaps, created Cost Analysis, Security Strategy, Testing Strategy, Observability, API Design docs |
| 7 |
Language conversion (VI→EN) |
Decided all documents should be in professional English |
Converted all 33 files from Vietnamese/mixed to consistent English |
| 8 |
Website creation |
Directed "build a website to present this assessment" |
Built complete Next.js + SQLite website with flat document viewer |
| 9 |
Website restructure |
Said "make it entity-based with proper objects, use Mermaid.js for diagrams" |
Rebuilt into structured pages: Dashboard, Architecture, Phases, Services, Risks, Tech Stack, Team, Documents |
| 10 |
Accuracy review |
Said "this AI declaration isn't honest, fix it" |
Rewrote this section to reflect actual workflow |
3. AI-Assisted vs Human — Per Section (Honest Numbers)
| Section |
AI/Human |
What The Human Actually Did |
| 4.1 Architecture Diagram |
85% AI / 15% Human |
Directed: "use Strangler Fig, YARP, per-service DB". Reviewed output, corrected event flow. AI generated all diagrams, service boundary details, communication rules |
| 4.1 Service Boundaries |
85% AI / 15% Human |
Approved 6 bounded contexts. AI proposed them from requirement analysis, wrote all details |
| 4.1 Communication Model |
80% AI / 20% Human |
Set the 3 rules (sync/async/CDC). AI designed the full model and schema |
| 4.2 Phased Approach |
80% AI / 20% Human |
Set constraints: "5 engineers, 9 months, 4 phases". Validated capacity math. AI generated phase structure, deliverables, cutover procedures |
| 4.2 Zero-Downtime |
85% AI / 15% Human |
Approved Strangler Fig + canary approach. AI generated all cutover details and rollback guarantees |
| 4.3 Failure Modeling |
85% AI / 15% Human |
Reviewed 8 AI-generated scenarios, approved final 5, adjusted likelihood ratings. AI generated all scenarios, mitigations |
| 4.4 Trade-Off Log |
75% AI / 25% Human |
Engineering judgment on each decision (YARP over Ocelot, Container Apps over AKS, Bicep over Terraform). AI structured and wrote justifications |
| 4.5 Assumptions |
80% AI / 20% Human |
Identified key gaps from brief reading (payment frozen, tool procurement). AI organized into 12 assumptions with impact analysis |
| 4.6 AI Declaration |
90% AI / 10% Human |
Said "this section is wrong, be honest about how we worked". AI rewrote everything |
| Capacity Math |
80% AI / 20% Human |
Directed phase-by-phase calculation approach. AI computed all numbers, human verified final total |
| Website + Presentation |
95% AI / 5% Human |
Directed design decisions ("entity-based", "use Mermaid", "make it polished"). AI wrote all code |
Summary
Overall AI contribution: ~85% (content generation, code, analysis, formatting, cross-referencing)
Overall Human contribution: ~15% (direction, decisions, constraint setting, review, correction)
But that 15% is what MATTERS:
- Which architecture pattern to use
- Which trade-offs to accept
- What constraints are non-negotiable
- When the AI output is wrong
- When to say "this is missing" or "this is enough"
AI generates. The Tech Lead DECIDES.
4. What Was Manually Validated
| Area |
Validation Method |
| Service boundaries |
Reviewed AI-proposed 6 bounded contexts. Approved after checking coupling (Travel↔Event ambiguity resolved by lifecycle analysis) |
| Capacity math (~44 MM) |
AI calculated phase-by-phase: P0(2.0, ×1.0) + P1(18.0, ×2.0) + P2(19.0, ×2.0) + P3(6.5, ×1.0) = 45.5 ≈ 44 effective MM. Human directed recalculation when cross-document inconsistency was found (originally different numbers in 7 files) |
| Phase timeline |
Verified each phase's scope feasibility. Checked dependencies: ACL must exist before Travel go-live |
| Trade-offs |
Each trade-off verified against constraint: "feasible for 5 engineers?" YARP vs Ocelot, Container Apps vs AKS — human judgment based on team reality |
| Failure scenarios |
Reviewed AI-generated list, filtered 8→5, adjusted likelihood/impact ratings |
| Cross-document consistency |
After AI generated all 33 docs, human spotted capacity numbers differed between files. Directed AI to audit and fix all 7 affected documents |
| Website accuracy |
Reviewed generated website against source documents. Directed restructure when flat doc viewer wasn't sufficient |
5. Preventing Blind AI Usage in the Backend Team
5.1 Governance Framework
┌──────────────────────────────────────────────────────────────────┐
│ AI GOVERNANCE PYRAMID │
│ │
│ ┌──────────┐ │
│ │ SECURITY │ Payment code: 2 human reviewers│
│ │ GATE │ Zero AI-only merge │
│ └────┬─────┘ │
│ │ │
│ ┌───────┴────────┐ │
│ │ HUMAN REVIEW │ All business logic: human │
│ │ GATE │ validates requirement trace │
│ └───────┬────────┘ │
│ │ │
│ ┌───────────┴──────────┐ │
│ │ AI AUTO-REVIEW GATE │ CodeRabbit first-pass │
│ │ │ Flag issues for human │
│ └───────────┬──────────┘ │
│ │ │
│ ┌───────────────┴────────────────┐ │
│ │ CI PIPELINE GATE │ Lint + tests + │
│ │ (automated, mandatory) │ SAST + contract │
│ │ No --no-verify bypass │ tests. Must pass. │
│ └────────────────────────────────┘ │
│ │
│ EVERY LINE OF AI CODE passes through ALL 4 gates. │
│ Skip = blocked merge. No exceptions. │
└──────────────────────────────────────────────────────────────────┘
5.2 Specific Practices
| Layer |
Practice |
Enforcement |
| Code Generation |
AI output MUST pass full CI pipeline (lint, unit tests, contract tests, SAST). No --no-verify bypass |
Branch protection rules: CI green required for merge |
| Business Logic |
All AI-migrated business rules require human validation checklist: trace requirement → implementation → test |
PR template includes "Business rule validation" checkbox. Reviewer must check |
| Architecture |
AI drafts ADRs. Tech Lead signs off. AI cannot autonomously decide service boundaries or DB strategy |
ADR template requires "Approved by: [name]" field |
| Code Review |
2-gate review: CodeRabbit auto-review (first pass) → human review (final gate, mandatory). AI flags suspicious patterns → human focuses effort |
GitHub branch protection: 1 required reviewer + CodeRabbit check |
| Security |
Payment-related code: 2 human reviewers, zero AI-only merge. All services: SAST/DAST mandatory in CI |
CODEOWNERS file: payment paths require 2 approvals |
| Knowledge Transfer |
Weekly "explain this code" sessions — team must understand what AI wrote. Rotate presenters |
Calendar recurring event. Retro tracks participation |
| Prompt Governance |
Prompt library versioned in git. Prompt changes require PR review. Test on sample before team-wide rollout |
/prompts/ directory in repo with CODEOWNERS |
| Metrics Tracking |
Dashboard tracks: AI-generated LOC vs human, AI PR rejection rate, AI code bug rate vs human |
Weekly AI Workflow Check (15 min, led by Tech Lead) |
5.3 Anti-Pattern Detection
RED FLAGS (Tech Lead must watch for):
❌ "AI wrote it, it works, ship it"
→ Fix: Require explanation. If engineer can't explain → doesn't merge.
❌ AI-generated tests that only test happy path
→ Fix: Test review focuses on edge cases. "What's NOT tested?"
❌ Copy-paste from AI without understanding
→ Fix: Random "explain this code" quizzes in standup.
❌ AI hallucinating method names that don't exist
→ Fix: CI catches (build fails). But also: code review checks imports.
❌ Prompt used once, result shipped, prompt discarded
→ Fix: Effective prompts → add to prompt library. Version control.
❌ "AI is always right" mindset
→ Fix: Culture setting from Day 1. Show examples of AI being wrong.
Celebrate catching AI errors in review.
5.4 AI Effectiveness Measurement
Weekly Dashboard (reviewed in AI Workflow Check):
┌────────────────────────────────────────────────────┐
│ AI Metrics (Week of ______) │
│ │
│ Code Generation: │
│ Lines written by AI: ____ │
│ Lines written by human: ____ │
│ AI-written code merged: ____% (target >60%) │
│ │
│ Quality: │
│ AI PR rejection rate: ____% (target <20%) │
│ AI code bug rate: ____/1000 LOC │
│ Human code bug rate: ____/1000 LOC │
│ │
│ Velocity: │
│ Tasks completed this week: ____ │
│ AI-assisted tasks: ____% (target >70%) │
│ Avg time per migration task: ____h (trending ↓?) │
│ │
│ Review: │
│ CodeRabbit issues flagged: ____ │
│ Issues confirmed by human: ____% (precision) │
│ Human-only issues found: ____ (AI missed) │
│ │
│ Verdict: AI multiplier this week = ____x │
│ On track for 2x? [YES/NO/ADJUST] │
└────────────────────────────────────────────────────┘
6. What Makes This Declaration Different
MOST CANDIDATES SAY:
"I used AI to help write this document"
"AI assisted with initial drafts"
"I validated everything manually"
WE SAY:
1. ONE tool: GitHub Copilot (VS Code Agent Mode, Claude Opus 4.6)
2. AI did ~85% of the work — we don't hide this
3. Exact step-by-step workflow: what happened, in what order
4. Honest per-section split (not "AI helped" — actual percentages)
5. The human's 15% was ALL the judgment: what to build, which constraints matter, when AI is wrong
6. Governance FRAMEWORK — not just how WE used AI, but how the TEAM will use AI
7. Measurable practices (dashboards, metrics, weekly reviews)
8. Anti-pattern detection (what BLIND usage looks like)
This is TECH LEAD thinking:
Not "how I use AI" → "how I ensure 5 engineers use AI responsibly at scale"
Not "AI helped me" → "here's exactly what AI did and what I did, with receipts"