What unit economics actually means
Strip away the jargon and unit economics is the profit math on a single unit, which for most SaaS is one customer. You look at what that customer pays, subtract what it costs to serve and to acquire them, and see what is left.
The usual pieces are three. Revenue per customer, the cost to acquire them, and the cost to serve them over their lifetime. Get those three right and you know whether growth makes you richer or just busier.
The classic model, and where AI breaks it
Traditional SaaS has a lovely property: the cost to serve one more user is nearly zero. That is why software runs 80 to 90 percent gross margins, and why lifetime value can be approximated as price times how long people stay.
AI snaps that assumption. Inference is a direct variable cost that sits against every dollar you collect and grows with usage. Gross margins on AI first products land closer to 50 to 60 percent, and the harder problem is that the cost is no longer the same for every user. Two customers on the same plan can have completely different unit economics because one uses the AI feature ten times as much. The structural version of this is in why AI SaaS margins are structurally lower.
Building a model for one AI feature, step by step
You do not need a finance team for this. You need five numbers and the discipline to use real ones.
- Revenue per user. Start with the plan price, or average revenue per user if plans vary. Say $29 a month.
- AI cost per user. Sum each user tokens across features, times your model prices. This is the number nobody has handy. Say it averages $6 a month.
- Other cost to serve. Hosting, payment fees, support. Call it $3 a month.
- Gross margin per user. 29 minus 6 minus 3 is $20 a month, a 69 percent margin. That is your real per user profit before acquisition.
- Lifetime value against acquisition. At a 20 month average lifetime that is $400 of gross profit per user. Against a $150 cost to acquire, a ratio of about 2.7, which is healthy.
That is a complete unit economics model, and on the average it looks great. The trouble is the word average. The exact per user margin math, plan revenue minus AI cost, is worked through in how to calculate gross margin per user.
The number that trips everyone up
AI usage is not uniform, it is a long tail. A small group of heavy users drives most of the cost, so the average hides them completely.
Take the same model and look at the top decile. If your heaviest users run $40 a month in AI cost, their unit economics are 29 minus 40 minus 3, a loss of $14 every month. They are lifetime value negative, and your rosy 69 percent average never showed it. A model built on the mean can be green while real money leaks out of the tail, which is why you model the distribution and watch the p95, not just the average.
Where the AI cost data comes from
Every number above was easy except one. Revenue is in your billing system, hosting is on an invoice, acquisition cost is in your ad accounts. AI cost per user is the one nobody hands you.
The reason is that your provider reports a single organization total, with no idea which of your customers generated which slice. To get per user AI cost you have to attribute each call to a user id and roll it up, which is exactly what Weckr does. It records cost per user and per feature at the moment of each call, so your unit economics model runs on measured numbers instead of guesses.
import { Weckr } from '@weckr/sdk'
const wk = new Weckr({
apiKey: process.env.WECKR_API_KEY,
plans: { free: 0, pro: 29 }, // pass plan prices so margin is computed for you
})
const response = await wk.chat(openai, {
model: 'gpt-4o-mini',
messages,
userId: user.id,
feature: 'chat',
plan: user.plan,
})With cost and revenue side by side per user, the tail stops hiding. You can see the demo on seeded data at the demo, wire it up from the docs, and fit unit economics into the full picture in the guide to AI cost and margin management.
FAQ
What are unit economics for an AI product?
Unit economics is the profit math on a single unit, usually one customer: what you earn from them versus what it costs to serve them. For an AI product the twist is that inference is a real variable cost that scales with usage, so the cost side of the equation moves per customer instead of staying near zero the way classic software cost does.
How does AI inference cost break the classic SaaS model?
Traditional SaaS has near zero marginal cost, so gross margins sit at 80 to 90 percent and lifetime value is roughly price times lifetime. AI adds a direct per user cost that grows with how much someone uses the product, which drags gross margin down to 50 to 60 percent and makes it vary customer by customer. The simple lifetime value formula overstates reality because it ignores a cost that is no longer flat.
How do I calculate AI cost per user?
Sum the tokens each user consumed across your features, multiply by your model prices for input and output, and add any cache or tool costs. The catch is attribution: providers report one organization total, not a per user breakdown, so you have to tag each call with your user id and roll it up yourself, or use a tool that does it for you.
Why is average AI cost per user misleading?
Because AI usage follows a long tail. A few heavy users generate most of the cost, so the average looks healthy while your top decile is quietly unprofitable. A unit economics model built on the mean can show a green gross margin while real money leaks out of the tail. You have to model the distribution, not just the average.
Where does the AI cost data for a unit economics model come from?
The revenue side is easy, it is in your billing system. The AI cost side per user is the hard part because the provider bill is a single total. Weckr fills that gap by recording cost per user and per feature at the moment of each call, so you can put real per customer AI cost next to plan revenue instead of guessing.
Keep reading
Model the real number, not the flattering one
Unit economics on an AI product live or die on the cost per user, and that is the one number the average and the provider bill both hide from you. Build the model on the tail, not the mean.
Weckr is the data source for the AI cost side: per user, per feature, in real time, next to what each customer pays. Start at the demo.