Chapter 1.12 — When traditional ML beats an LLM (and when to combine them)

🎯 Objective

Show that traditional ML, far from being obsolete, remains the correct choice for a huge class of problems — and that hybrid combinations are the most defensible enterprise pattern.

🧠 When traditional ML wins

• Repetitive, well-defined task.
• Latency must be low (p95 < 50 ms).
• Cost per call must be minimal.
• The output is classification or regression.
• There is enough labeled data.
• Behavior must be stable.

🤝 When to combine ML + LLM + agent

Typical hybrid pattern:

User input
   │
   ├─ 🧠 ML classifier (intent / triage)
   │     ├─ "simple" -> deterministic answer
   │     ├─ "documental" -> 🤖 LLM with RAG
   │     └─ "action" -> 🧩 Agent with tool calling + HITL
   │
   └─ 🔭 Shared observability

🚨 Failure modes

• Using an LLM where a classifier would solve it at a tenth of the cost.
• Using an agent where a deterministic workflow would suffice.
• Forcing a single path for all cases.

📌 Checklist

• [ ] Does the solution start with the simplest option?
• [ ] Are the LLM and agent reserved for cases that truly require them?
• [ ] Is there routing and fallback between the paths?

---