AI Unit Economics 101: Will Your AI Costs Eat Your Margins?

Traditional SaaS cost-of-goods-sold is mostly fixed: hosting, a database, maybe some third-party API fees that barely move with usage. The moment your product calls an AI model at runtime, that changes — you now have a cost line that scales directly with how many people use your product and how much they use it. If you charge for the product, that line interacts with your pricing in ways that are easy to get wrong.

Cost per active user vs cost per paying user

These two numbers are often confused and they tell you very different things.

• Cost per active user includes everyone who triggers an AI call — free trial users, free-tier users, churned-but-still-logged-in users. This is your total runtime AI spend divided by total active users.
• Cost per paying useris the number that actually needs to be covered by revenue. If only 5% of active users pay, your cost per paying user can be 20x your cost per active user — and that's the number that determines whether your pricing works.

Break-even price vs required price for target margin

• Break-even price is the price at which your AI cost per paying user is fully covered, with zero margin left over for hosting, support, or profit.
• Required price for target marginis higher — it's the price needed so that, after AI cost, you still hit a target gross margin (commonly 70–80% for software).

The gap between these two numbers and your planned price is the headline result. If your planned price is below break-even, every paying user actively loses you money on AI cost alone — no amount of volume fixes that; it makes it worse.

The free-tier trap

Free tiers are a deliberate, often correct choice — but every free-tier user who calls your AI feature costs money with no offsetting revenue. If your conversion rate from free to paid is low (which it usually is, in the low single digits), the AI cost generated by your entire free-tier population can rival or exceed the AI cost generated by your paying population. Two things help: a hard usage cap on free-tier AI features, and routing free-tier traffic to a cheaper model than paid traffic. Who actually absorbs this cost — the company, passed through to the user, or some mix — is worth deciding explicitly rather than discovering by accident.

Business model changes everything

• Subscription. AI cost is effectively a fixed COGS per seat, regardless of how much a given user actually uses the AI feature. Heavy users are subsidized by light users — until a few outliers blow up the average. Usage caps per plan tier are the usual fix.
• Usage-based.AI cost can be passed through more directly, since price already scales with usage. The risk shifts from "margin erosion" to "price shock" for your users if usage spikes.
• One-time purchase.There's no recurring revenue to offset ongoing AI cost, so any runtime AI feature needs to either be cheap enough to fund from the one-time price or be scoped out of the product entirely.

A worked example

A subscription product priced at $19/month expects 5% of its 2,000 active users to pay — 100 paying users. If total monthly runtime AI cost across all 2,000 active users is $400, that's $0.20 per active user, but $4.00 per paying user. At $19/month with a 70% target gross margin (i.e. non-AI costs need to leave at least $13.30), AI cost has plenty of room. But if the free tier grows to 20,000 active users at the same per-user AI cost without a proportional increase in paying users, total AI spend grows 10x while revenue doesn't move — and the same $4.00-per-paying-user number becomes $40. That's the calculation worth running before you remove a usage cap, not after.

Want to check your own pricing against your projected AI costs, with break-even and target-margin price points calculated automatically?