Unit economics: CAC, LTV, contribution margin and AI costs

0:00 Unit economics

Here's a story that's become common. An AI startup launches, customers love it, usage explodes, and the founders celebrate. Then the model bill arrives, and it's bigger than revenue. They'd built a business that lost more money the more customers loved it. In this lecture, you'll learn the core unit economics every founder needs: contribution margin, customer acquisition cost, lifetime value and payback. And you'll learn how to treat AI model usage properly, as a cost of goods sold, so you can see your real margin per customer before it surprises you.

0:40 Why it matters

Why does this matter? Because growth multiplies whatever your unit economics already are. If each customer makes you money, growth makes you more. If each customer loses money, growth makes your losses bigger, faster. Unit economics are simply the revenue and costs attached to one unit: one customer, one order or one subscription. Here's the key idea. You can't fix negative unit economics with scale unless you know exactly which number scale will change, and why.

1:13 The lemonade stand

Let's build the concepts with an analogy. Imagine a lemonade stand. Each cup sells for one pound. The lemons, sugar and cup cost forty pence. That leaves sixty pence of contribution, which pays for the stand, the sign and eventually your profit. Now suppose you spend two pounds on a flyer campaign for every new regular customer. That's your customer acquisition cost, or CAC. If a regular buys two cups a week for ten weeks, they bring in twelve pounds of contribution. That's their lifetime value, or LTV, measured on contribution rather than revenue. LTV to CAC is twelve to two, or six. And payback, the time to earn back the two pounds, is under two weeks.

2:04 Definitions

Now the formal definitions. Contribution margin is price minus variable costs per unit. CAC is total sales and marketing spend divided by new customers won, including salaries and tools, not just ads. Monthly churn is the share of customers who leave each month. Average lifetime is roughly one divided by churn. So four percent churn means about twenty-five months. LTV is monthly contribution times lifetime. LTV to CAC tells you the return on acquisition spend, and CAC payback tells you how many months until a customer has paid back what it cost to win them. Rules of thumb, like three to one and about a year's payback, are prompts for questions, not laws.

2:53 Worked example 1: SaaS (illustrative)

A simple worked example, illustrative. A software tool for small retailers costs thirty dollars a month. Hosting, payment fees and support cost six dollars per customer, so contribution is twenty-four dollars, or eighty percent. Monthly churn is four percent, so average lifetime is about twenty-five months. LTV is twenty-four times twenty-five, which is six hundred dollars. CAC is two hundred dollars. So LTV to CAC is three, and payback is two hundred divided by twenty-four, a bit over eight months. That looks healthy. But notice how sensitive it is. If churn rises to eight percent, lifetime halves, LTV halves, and the ratio falls to one and a half.

3:40 AI usage is COGS

Now, AI products. When your product calls a language model, every customer action has a direct cost. Providers charge by tokens, which are small chunks of text, roughly three-quarters of an English word on average, and often more tokens per word in languages like Arabic or Urdu. Input tokens, what you send, and output tokens, what comes back, usually have different prices. That cost belongs in cost of goods sold, and so in contribution margin, exactly like the lemons in our lemonade. Treating it as a fixed IT overhead hides the most important risk: margins that shrink as customers use you more.

4:24 Cost per task

Here's the cost-per-task formula. Take input tokens times the input price, plus output tokens times the output price, with prices usually quoted per million tokens. Add an allowance for retries and regenerations. Add retrieval, search, tool or hosting costs per task. And if a person reviews the output, add their minutes times their loaded cost. Then multiply by tasks per month to get COGS per customer. One practical tip. Always write down the date you checked the provider's price page. Model prices change often, sometimes downwards, sometimes with new model versions that behave differently.

5:05 Worked example 2: AI support replies (illustrative)

Now the realistic scenario. All numbers are illustrative, not current prices. A startup sells an AI tool that drafts replies to customer emails for small online shops at forty-nine pounds a month. An average task sends three thousand input tokens and gets four hundred back. At illustrative prices of two pounds forty per million input tokens and twelve pounds per million output tokens, that's about one point two pence per task. Add retries and retrieval, and it's about one point eight pence. A typical shop sends twelve hundred tasks a month: about twenty-one pounds sixty of model cost. With six pounds of other variable costs, contribution is about twenty-one pounds forty, roughly forty-four percent. But a heavy shop sending five thousand tasks costs about ninety pounds. It loses forty-seven pounds a month.

6:02 Watch me: the sheet in action

Watch me do it. I open the unit economics sheet from the lesson. I create three columns: light, typical and heavy customers. I enter price, tasks per month, tokens per task and the provider's prices, with the date checked. The sheet calculates cost per task, COGS, contribution, lifetime, LTV and payback for each segment. Now I test levers. First, I route simple replies to a smaller model, after checking quality on fifty real examples. Cost per task drops. Next, I trim the context, sending only the relevant policy sections rather than the whole handbook. Then I add a usage allowance with overage to the pricing. Heavy customers turn positive. I haven't guessed. I've modelled each lever.

6:52 Levers

Let's gather the levers. For any business: raise price where value supports it, cut variable costs, improve retention, expand revenue from existing customers, and find cheaper acquisition channels such as referrals and content. For AI businesses, add these. Right-size the model for each task. Trim context. Use features many providers offer, like prompt caching and discounted batch processing, after checking the current docs. Cap output length. Add a usage dimension to pricing. And watch human review time. If every output needs five minutes of expert checking, that review, not the tokens, is probably your biggest variable cost.

7:34 Common mistakes

Now the common mistakes. Calculating LTV on revenue instead of contribution. Using hopeful churn numbers without data. Leaving salaries and tools out of CAC. Hiding expensive paid channels inside a blended CAC with free organic sign-ups. For AI products, treating model costs as fixed overhead, using only the average customer, and forgetting that languages, long documents and retries can multiply token counts. And finally, assuming that model prices will keep falling, so the problem will fix itself. Maybe they will. But plan with today's prices, and treat any fall as upside.

8:14 Recap and try this now

Let's recap. Unit economics tell you whether each customer makes money: contribution, CAC, churn, LTV, LTV to CAC and payback. For AI products, model usage is cost of goods sold. Calculate cost per task from tokens, retries, tools and human review, and model it per usage segment, not just on average. Then pull the levers: model choice, context, caching, pricing and review time. Your try this now: build the unit economics sheet for your idea with three usage segments, write down the date you checked your provider's prices, and run two scenarios: churn twenty percent lower, and CAC twenty percent higher.