AI unit economics measures what it costs to serve one unit of work versus what that work earns. For a startup founder, it answers whether growth builds profit or burns cash. A unit is one billable action, such as a resolved support chat, a generated report, or a month of monitored devices. Founders who track cost and revenue at that level can price with confidence.
Table of Contents
- What counts as your unit?
- Where does the money go?
- How should you price the product?
- When do margins improve?
- Which warning signs deserve action?
- Frequently Asked Questions
What counts as your unit?
Pick a unit that matches how customers get value and how you incur cost. A design tool might use finished exports. A support product might use resolved cases. A sales tool might use enriched leads. Avoid units that hide work.
Monthly logins hide heavy users and light users in the same average. Queries or tasks completed reflect effort more clearly. The right unit makes one sale and one service cost easy to compare. Revisit the choice as the product changes. New features often add hidden steps such as retrieval, reranking, or review. If those steps cost money, they belong inside the unit definition.
Where does the money go?
AI products carry two kinds of cost. Fixed costs include salaries, model evaluation, and base cloud services. Variable costs rise with use, including inference, data retrieval, storage, and human review. Inference is often the largest variable cost.
Longer inputs, larger models, repeated calls, and high accuracy settings all raise it. Caching, smaller models for simple tasks, and tighter retrieval limits can lower it without harming quality. Do not overlook support and failure costs. Regenerations, manual corrections, and customer help take staff time. Track them per unit so low prices do not turn popular features into losses.
How should you price the product?
Start from measured cost per unit, then add margin and delivery overhead. Delivery overhead means onboarding, monitoring, billing, and support. Many founders set prices from competitors first and discover later that heavy users lose money.
Choose a pricing model that moves with cost: Test pricing against real behavior, not ideal behavior. Look at median and heavy users separately. A plan that works for average use can still fail when a few accounts use ten times more.
- Per-use pricing fits variable workloads and protects margins on heavy use.
- Tiered plans fit predictable workloads and reward annual commitments.
- Hybrid plans pair a base fee with overage charges for power users.
When do margins improve?
Margins improve when cost per unit falls or revenue per unit rises. Costs often fall through better routing, shorter prompts, cached results, and fewer retries. Revenue per unit rises through higher-value use cases, add-ons, and better retention. Scale helps only if the unit is sound. More customers spread fixed costs, but they multiply variable losses.
Fix contribution margin first, meaning price minus variable cost stays positive. Then growth can carry the business. Watch retention by cohort and by use case. Customers who integrate the tool into daily work renew at higher rates. Low-retention segments often need simpler onboarding, narrower scope, or a different price.
Which warning signs deserve action?
Falling gross margin during growth means costs are rising faster than prices. Rising retry rates, longer outputs, and added human review often explain it. Review logs and cap the most expensive paths.
Concentration risk is another warning. One large customer or one costly feature can dominate inference spend. Set usage limits, overage fees, or custom pricing before that account renews. Pause free trials that attract automated abuse and measure cost per signup weekly.
Frequently Asked Questions
What is a healthy target for AI gross margin?
Targets vary by market and support load, so focus on positive contribution margin per unit first, then steady improvement as volume grows.
Should early startups charge per use or sell flat plans?
Per-use or hybrid pricing fits unpredictable AI costs better, while flat plans work only when you cap effort or know usage limits well.
How can founders cut inference costs safely?
Route simple tasks to smaller models, shorten inputs, cache repeats, limit retries, and measure quality after each change.