Cost Management & ROI

AI features can be expensive to run. Understanding the economics helps you make better product decisions.

AI Cost Components

API Costs (Pay-Per-Use)

Component	What It Means	Typical Cost
Input tokens	Text you send to the model	$1-15 per 1M tokens
Output tokens	Text the model generates	$3-60 per 1M tokens
Fine-tuning	Custom model training	$8-25 per 1M training tokens
Storage	Storing fine-tuned models	$0-5 per model/day

Token estimation rule of thumb:

1 token ≈ 4 characters (English)
100 words ≈ 75-100 tokens
1 page of text ≈ 500-800 tokens

Self-Hosted Costs

Component	Monthly Range
GPU infrastructure	$2,000-50,000+
ML engineering time	$15,000-40,000
MLOps tools	$500-5,000
Monitoring/observability	$200-2,000

Calculating Cost Per User Action

Step 1: Estimate Token Usage

Average input: 500 tokens
Average output: 200 tokens
Calls per user action: 2 (e.g., generate + refine)

Total tokens per action: (500 + 200) × 2 = 1,400 tokens

Step 2: Apply Pricing

Using GPT-4o pricing (Dec 2025):

Input: 500 × 2 = 1,000 tokens × $0.0025/1K = $0.0025
Output: 200 × 2 = 400 tokens × $0.01/1K = $0.004
Total per action: $0.0065

Step 3: Scale to Usage

Monthly active users: 100,000
Actions per user/month: 20
Total actions: 2,000,000

Monthly AI cost: 2,000,000 × $0.0065 = $13,000
Cost per user: $0.13/month

Cost Optimization Strategies

Strategy 1: Choose the Right Model

Task Complexity	Model Tier	Example
Simple classification	Small/Fast	GPT-4o-mini, Claude Haiku
Standard generation	Medium	GPT-4o, Claude Sonnet
Complex reasoning	Large	GPT-4, Claude Opus

Impact: Switching from GPT-4 to GPT-4o-mini can reduce costs by 90%+ for simple tasks.

Strategy 2: Optimize Prompts

Technique	Savings
Shorter system prompts	10-30%
Remove unnecessary context	20-40%
Use structured outputs	Varies

Strategy 3: Cache Responses

If same input → Return cached response
Else → Call API → Cache result

Cache hit rate of 30% = 30% cost reduction

Strategy 4: Batch Processing

Real-time vs batch trade-off:

Approach	Cost	Latency
Real-time API	Higher	<1 second
Batch API	50% lower	Hours

Use batch for: Reports, analytics, background processing

Strategy 5: Fallback to Cheaper Models

Try: Small model (fast, cheap)
If confidence < threshold:
    Escalate to: Large model (slow, expensive)

Typical result: 70-80% handled by cheap model = major savings.

Calculating ROI

ROI Formula for AI Features

ROI = (Value Generated - Total Cost) / Total Cost × 100%

Value Sources

Value Type	How to Measure
Revenue increase	Conversion lift × Average order value
Cost reduction	Hours saved × Labor cost
Efficiency gain	Tasks automated × Previous cost
Risk reduction	Prevented losses

Example: AI Customer Support

Costs:

AI API costs: $5,000/month
Implementation: $20,000 (one-time)
Maintenance: $2,000/month
Year 1 total: $20,000 + ($7,000 × 12) = $104,000

Value:

Support tickets resolved by AI: 40%
Previous cost per ticket: $15
Monthly tickets: 10,000
Savings: 10,000 × 0.4 × $15 = $60,000/month
Year 1 total: $720,000

ROI:

ROI = ($720,000 - $104,000) / $104,000 × 100%
ROI = 592%

Presenting AI Business Cases

What Executives Want to Know

Question	Your Answer Should Include
"What does it cost?"	Upfront + ongoing, per-user breakdown
"What's the return?"	Specific metrics, conservative estimates
"What's the risk?"	Failure scenarios, mitigation plans
"When do we break even?"	Timeline to positive ROI
"What if it doesn't work?"	Rollback plan, sunk costs limited

Business Case Template

AI Feature: [Name]

Investment Required:
- Development: $X (one-time)
- Operations: $Y/month
- Year 1 Total: $Z

Expected Value:
- [Metric 1]: $A/month (conservative)
- [Metric 2]: $B/month (conservative)
- Year 1 Total: $C

ROI: X%
Break-even: Month N

Risks & Mitigations:
- Risk 1: [Mitigation]
- Risk 2: [Mitigation]

Decision Needed: [Specific ask]

AI ROI depends on matching the right model to the task, optimizing usage, and clearly connecting costs to business value. Always start with the business outcome, work backward to the AI investment needed.

Up next: Test your understanding with Module 3 Quiz. :::