Track 2 — AI Product Economics

Module 2.1: AI COGS Analysis

Master the economics of AI features: inference costs, token economics, and pricing architecture. Build AI features that protect — not destroy — your gross margins.

3 Lessons~50 minIntermediate-Advanced

🎯 What You'll Learn

  • How AI COGS differs from traditional SaaS COGS
  • How to analyze token economics (input/output, model selection, caching)
  • How to design pricing models that account for AI variable costs
  • How to project AI margin at 10x and 100x scale
1

Lesson 1: The AI COGS Equation

Unlike traditional software where COGS is near-zero after development, AI features have per-request costs that scale linearly with usage. This fundamentally changes the margin equation.

Traditional SaaS COGS

Hosting + bandwidth + support = 15-25% of revenue. Marginal cost of serving one more user is nearly zero.

Traditional SaaS gross margin: 75-85%
AI-Enhanced SaaS COGS

Traditional COGS + inference costs + embedding storage + model fine-tuning + guardrail processing. Each user interaction has a real cost.

AI SaaS gross margin: varies 40-75% depending on AI intensity
The Margin Collapse Risk

If AI feature cost per user > revenue per user, you lose money on every interaction. More users = more losses. This is AI margin collapse.

Critical if AI COGS > 30% of feature revenue
📝 Exercise

Calculate your AI feature's COGS: (tokens consumed × cost per token) + (embedding storage) + (guardrail processing). Express as cost per user per month.

2

Lesson 2: Token Economics Deep Dive

Tokens are the fundamental unit of AI cost. Understanding token economics — input vs. output, model selection, caching — determines whether your AI feature is profitable.

Input vs. Output Tokens

Output tokens cost 2-4x more than input tokens. A chatbot that generates long responses costs dramatically more than one that gives concise answers.

GPT-4o: $2.50/1M input, $10/1M output (4x ratio)
Model Selection Impact

Choosing GPT-4o vs. GPT-4o-mini can reduce costs 15-20x. Most features don't need frontier models. Matching model capability to task complexity is the #1 cost lever.

Test: can a smaller model handle 80%+ of your queries adequately?
Prompt Engineering as Cost Control

Shorter, more precise prompts = fewer tokens = lower cost. System prompts repeated on every request are the biggest hidden cost multiplier.

Target: system prompt < 500 tokens. Each 100 tokens × 1M requests = real money.
Caching Strategies

Semantic caching (storing responses for similar queries) can reduce inference calls 30-70%. The cache hit rate directly reduces your AI COGS.

Target cache hit rate: > 40% for FAQ-style queries, > 20% for general.
📝 Exercise

Use the AUEB calculator at /tools/aueb to model your AI feature's token economics at current volume, 10x volume, and 100x volume. At what scale does margin collapse occur?

3

Lesson 3: API Pricing Architecture

How you price AI features determines whether they're profit centers or cost centers. The pricing model must account for the variable cost nature of AI.

Usage-Based Pricing

Charge per API call, per query, or per action. Aligns costs with revenue. Risk: usage spikes can overwhelm infrastructure.

Best for: developer tools, APIs, B2B platforms
Tiered Pricing

Free tier (limited queries) → Pro tier (more queries + features) → Enterprise (unlimited + SLAs). The free tier is your PLG acquisition engine.

Critical: free tier cost must be < CAC of alternative acquisition channels
Seat-Based with AI Budget

Per-seat pricing plus an AI "credits" budget per seat. When credits run out, user upgrades. Combines predictability with usage correlation.

Most common for AI-enhanced SaaS products in 2025-2026
📝 Exercise

Design a pricing model for an AI feature with: 1) known cost per query, 2) variable usage patterns, 3) a free tier for PLG. Calculate break-even at each tier.

📊 Module Assessment

Complete to demonstrate mastery of Module 2.1: