AI unit economics and pricing
AI features cost money every time they run. This module teaches you to estimate those costs, cut them safely and price AI features without wrecking your margin.
Learning objectives
- Break down the cost drivers of an AI feature
- Calculate cost per request, per task and per active user
- Apply cost levers guarded by evals
- Choose a pricing model and fair-use limits
2.1 The anatomy of AI cost
Traditional software costs almost nothing extra per click. AI features have a real marginal cost every time they run. A PM who can't estimate it can't price the feature, set usage limits or spot a margin problem before finance does.
| Cost driver | What drives it |
|---|---|
| Input tokens | Instructions, conversation history, retrieved documents, tool results — often the biggest part |
| Output tokens | Length of answers; usually priced higher per token than input |
| Number of calls | One call per answer, or dozens per task for an agent |
| Retrieval and tools | Search, vector databases, third-party APIs |
| Human review | Staff time to check or approve outputs — often forgotten |
2.2 A worked cost model
The prices below are illustrative only — check your provider's current price list — but the method works with any numbers. Assume $3 per million input tokens and $15 per million output tokens.
2,000 input tokens (question + retrieved articles) = $0.006 · 300 output tokens = $0.0045 · ≈ $0.01 per answer. At 5 answers per user per working day (110 a month): ≈ $1.16 per active user per month.
Context grows as the agent works: average 8,000 input tokens per step × 20 = 160,000 = $0.48 · 500 output tokens × 20 = 10,000 = $0.15 · ≈ $0.63 per task — about 60 times Feature A. At 3 tasks per user per day, that's over $40 per user per month.
The lesson: agents and long contexts change the economics. Model costs at expected and heavy usage before you set a price.
2.3 Levers that cut cost without cutting quality
- Prompt caching — reuse long, repeated prompt prefixes (instructions, documents); providers charge cached reads at a fraction of the normal input price
- Model routing — send simple requests to a smaller, cheaper model and hard ones to a larger model
- Batch processing — run non-urgent jobs (overnight summaries, back-catalogue tagging) through discounted batch APIs
- Trim the context — retrieve fewer, better chunks; summarise long histories
- Limit output — ask for the length the user actually needs
- Cache whole answers — for frequently repeated questions
Every lever should be guarded by evals: a cheaper model that fails your must-never conditions is not a saving. AWS's Generative AI Lens has a full cost-optimisation pillar.
2.4 Gross margin and the heavy-user problem
Gross margin is revenue minus the direct cost of delivering the product, as a share of revenue. Software has traditionally enjoyed very high margins; AI features add variable cost that can erode them. Usage is usually skewed: a small share of users generates most requests. On a flat-rate plan, your heaviest users can cost more than they pay. Track cost per active user by plan and by percentile (median, 90th, 99th), not only the average.
2.5 Pricing models for AI
| Model | How it works | Strengths | Watch out for |
|---|---|---|---|
| Included in the seat | AI features part of the existing subscription | Drives adoption; simple | Margin erosion from heavy users |
| Seat + AI add-on | Extra fee per user for AI features | Clear value signal; recovers cost | Low uptake if value isn't obvious |
| Usage or credits | Pay per request, document or credit bundle | Cost and revenue move together | Buyers dislike unpredictable bills |
| Outcome-based | Pay per resolved case, qualified lead or completed task | Aligns price with value | Needs agreed, measurable outcomes; you carry the risk of failures |
Most products combine models: AI included up to a fair-use limit, then credits or a higher tier. Whatever you choose, state limits honestly and show users how much they've used.
For an AI feature you know, estimate tokens per request, requests per user per month and cost per active user at the median and the 99th percentile. Then choose a pricing model and fair-use limit, and calculate the gross margin for a typical and a heavy user.
See a worked example
Example — AI drafting in a CRM, $40 per seat: 1,500 input + 400 output tokens ≈ $0.0105 per draft. Median user: 60 drafts a month ≈ $0.63. 99th percentile: 2,000 drafts ≈ $21. Decision: include 500 drafts a month in the seat; above that, a $10 'AI Pro' add-on with routing to a smaller model for short drafts. Heavy-user margin stays positive.
Key takeaways
- AI has a real marginal cost per request.
- Agents and long contexts can cost many times more than single calls.
- Caching, routing, batching and trimming cut cost — check with evals.
- Track cost per user by percentile, not just average.
- Most products combine pricing models with honest limits.