Chain Multiple AI Calls -- Classify Cheap, Generate Expensive

To cut LLM costs, chain two AI calls: a fast, cheap model classifies or triages each input, and only inputs that need sophisticated generation are routed to a larger, costlier model. Simple cases get a template answer. This classify-then-route pattern can reduce spend by 60-80% in high-volume workflows without lowering quality on the hard questions.

What is the classify-cheap, generate-expensive pattern?

Not every input needs a large, expensive model. Use a two-stage pattern: a fast, cheap model classifies or triages the input, and only inputs that require sophisticated generation get routed to the more powerful (and costly) model. This can cut LLM costs by 60-80% in high-volume workflows.

Real-world example: Incoming customer questions are classified by gpt-4o-mini (fast, cheap). Only complex questions are sent to claude-sonnet-4-20250514 for a detailed response. Simple FAQs get a template answer.

How do you route between a cheap and an expensive model?

Workflow structure:

[Webhook] → [Classify: gpt-4o-mini] → [IF: is_complex] →── true ──→ [Generate: Claude Sonnet]
                                                          └─ false ─→ [Template Response]

Classification node (OpenAI, gpt-4o-mini):

Classify the following customer question into exactly one category:
- "faq" — answerable from standard documentation
- "complex" — requires detailed, contextual reasoning
- "escalate" — needs human agent

Return JSON: {"type": "faq" | "complex" | "escalate", "faq_topic": "string or null"}

Question: {{ $json.question }}

IF node condition:

{{ JSON.parse($json.message.content).type === "complex" }}

Generation node (Anthropic, Claude Sonnet) -- only runs for complex questions:

You are a senior customer success agent. Provide a thorough, empathetic
response to this customer question. Reference specific product features
and include next steps.

Question: {{ $json.question }}
Customer tier: {{ $json.account_tier }}

How much does classify-then-route save?

Cost comparison for 1,000 daily questions (hypothetical):

Approach	Model	Cost/day (approx)
All to Sonnet	Claude Sonnet	~$15.00
Classify + Route	Mini + Sonnet	~$4.50

The classify-then-route pattern is one of the highest-ROI optimizations for AI workflows at scale.

Related: Use Manual Trigger During Development Instead of Webhook or Schedule · Flatten Deeply Nested API Responses