AI unit economics only matter when AI cost is a direct input to revenue. Internal tooling? Skip the complexity. Charge for AI? You need it.
This is for founders, engineering directors, and FinOps or platform leads at AI product companies. The board is asking "what does it cost per customer?" or you're building pricing and need to know your AI cost of goods sold. The question is whether you need unit economics now, what to measure if you do, and how to take the first step without over-building.
By the end you'll know: when cost per customer matters vs when to skip it, the minimum dimensions to track, and how to decide cost per customer vs cost per feature vs cost per request.
When unit economics matter vs when to skip
Not every AI spend needs unit economics. Use this to decide:
| Scenario | Need cost per customer? | What to track instead |
|---|---|---|
| You charge per AI usage (usage-based pricing) | Yes | Cost per customer, cost per request, margin by plan |
| You bundle AI in a fixed-price product | Yes | Cost per customer or per seat, COGS as % of revenue |
| Internal tooling (coding assistant, internal chat) | No | Total spend by team or project; revisit if you productize |
| Pre-revenue or experimental | No | Provider and model totals; add attribution when AI becomes a revenue driver |
| Freemium with usage caps | Maybe | Cost per user for paid tier; total for free tier |
The nuance: cost per request is an engineering metric—useful for debugging and optimization. Cost per customer is a pricing and margin metric—useful when AI spend varies by customer and you need to know if you're profitable. If you're not charging for AI directly, cost per customer is often overkill. If you are, it's essential.
What to measure: minimum dimensions
If you need AI unit economics, track these dimensions so you can compute cost per customer, cost per feature, and cost per request:
| Dimension | Why it matters | Example |
|---|---|---|
| Provider | Compare OpenAI, Anthropic, Bedrock, Vertex, Azure OpenAI | openai, anthropic |
| Model | Explains cost-per-request changes | gpt-4o, claude-sonnet |
| Feature or workflow | Connects spend to product surfaces | chat, summarization, support-bot |
| Team or owner | Cost ownership and review | platform, product-ops |
| Customer or account | Enables cost per customer, margin analysis | org_12345, enterprise-plan |
You need all five to compute meaningful AI cost per customer. Provider and model alone give you cost per request. Add feature and customer to get cost per customer by product area. See how to attribute AI costs by feature, team, and customer for implementation.
First steps
If you need unit economics: Start with attribution. Define and capture provider, model, feature, team, and customer on every meaningful request. Use provider-native grouping (OpenAI projects, Anthropic workspaces) plus application metadata. Then add a reporting layer that joins provider usage with your product data. How to attribute AI costs by feature, team, and customer walks through the structure. For model-level cost per project and customer, AI cost monitoring gives you one place to see it.
If you don't need it yet: Skip the complexity. Track total AI spend by provider and model. Revisit when AI becomes a direct cost of revenue—when you're pricing a product that uses AI, when the board starts asking, or when you're building a unit-economics model for fundraising.
Related reading
- How to Attribute AI Costs by Feature, Team, and Customer
- AI Cost Monitoring
- How to Track LLM API Spend
- AI Cost Overruns: Why They Happen and How to Prevent Them
FAQ
What is AI cost per customer?
AI cost per customer is the total AI spend (API calls, model usage, embeddings, etc.) attributable to a single customer or account. You need it when you charge for AI usage or bundle AI in a product—to know margin, set pricing, and understand which customers are profitable.
When do I need unit economics for AI?
When AI cost is a direct input to revenue: usage-based AI products, fixed-price products that bundle AI, or freemium with paid AI tiers. If AI is internal tooling or pre-revenue experimentation, total spend by provider and team is usually enough. Add unit economics when you're pricing or when the board asks.
What is AI COGS?
AI cost of goods sold (COGS) is the direct cost of delivering AI to your customers—API costs, model costs, inference costs. It's the variable cost that scales with usage. Track it when AI is a revenue driver so you can compute gross margin and set pricing.
How do I track LLM cost per request?
Capture provider, model, and request metadata (feature, customer, team) on each call. Join usage or billing data with that metadata. Cost per request = total cost / request count, optionally broken down by model, feature, or customer. See how to attribute AI costs for the attribution structure.
What is AI cost attribution?
AI cost attribution is assigning each dollar of AI spend to a business dimension—provider, model, feature, team, or customer. It's the foundation for unit economics. Without it, you have a bill total but no way to compute cost per customer, cost per feature, or cost per request.