Stage 1 · AI Use-Case ROI Framing & Technical Feasibility Triage Pattern
First-stage AI feasibility pattern preventing over-engineering: systematically evaluates error tolerance, latency SLAs, unit economics, and HITL verification cost to choose between deterministic code, classical ML, or GenAI agents.
Best For
Project scoping reviews, build-vs-buy TCO modeling, and the crucial first 5 minutes of senior AI System Design interviews.
Human-AI Boundary
Business owners and engineering leads jointly lock North Star ROI KPIs and acceptable error blast-radius thresholds before kickoff.
Anti-Patterns to Avoid
- Deploying an autonomous multi-agent LLM loop where a 50-line SQL query or GBDT classifier achieves higher accuracy at 1/1000th the cost
- Measuring POC success by qualitative demo vibes while ignoring the cost of human QA verifying edge-case hallucinations
Latency/Cost Profile & Eval Metrics
Design-time analytical gate (~1-2 weeks discovery sprint before full build) · Inference COGS + HITL review labor vs Baseline manual process savings
- Net Task Time Reduction & Payback Period: >= 40% verified operator time saved and < 6 months engineering payback (Compare Baseline Cycle Time vs AI+HITL Total Cycle Time on 50 real historical tickets)
- Cost of Failure (Blast Radius) Bound: Zero unreviewed high-severity external actions (Classify errors into reversible UX friction vs irreversible financial/legal liability)
💰 Business ROI Model & Evidence Level
production_practiceMonthly Net ROI = Volume × (Labor Saved/Task - Inference COGS - HITL QA Cost) - Amortized Build Cost, stress-tested against P90 error-remediation cost.
🏗️ End-to-End Architecture Pipeline
Quantify current human handle time and worst-case error liability.
Route deterministic rules to code, structured scoring to ML, and unstructured synthesis to LLMs.
⚖️ Engineering Trade-Off & Decision Matrix
🚨 Production Failure Modes & Signatures
- Solving an inherently human organizational alignment gap with a RAG chatbot
- Underestimating long-tail input formatting variance (scanned tables, corrupted encodings) in POC ROI projections
🎤 Interview Defense & System Design Deep-Dive Q&A
- When a stakeholder asks for an autonomous multi-agent workflow, what 4 diagnostic questions prove whether a simpler workflow or classical ML is superior?
Bidirectional Taxonomy Links
Internal Tools & Primary Sources
- Anthropic — Building Effective Agents (When to Use Workflows vs Agents) (engineering_blog, 2026-11-15)
- Chip Huyen — AI Engineering Architecture & Feasibility (engineering_blog, 2026-11-15)