For AI startups, free trials are not free and revenue is not profit
Free credits belong in the cost of finding customers, and computing bills have to come out before growth counts.
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch · 20Growth: How to Build a $100M Growth Engine: Lessons from Wispr Flow and Superhuman | Why You Should Do Paid Ads Today and How To Do Them | How to Build the Best Referral Programs and How to Crush UGC with Matt SwulinskiSelling software used to be simple math. You added up what you spent on ads to win a customer, called it customer acquisition cost, and compared it with lifetime value, the revenue that customer would bring in over time. If the second number was comfortably bigger than the first, you could grow.
For AI products, that math breaks, said Matt Swulinski, the growth lead at the voice dictation startup Wispr Flow and the AI employee startup Victor, who previously led growth at the email startup Superhuman. In the beginning, he said, pick the single event that matters most to the business and optimize only for the cost of getting it. For many software companies that is starting a subscription or a trial. At Wispr Flow it was the download. At Victor it is getting a company to add the app inside Slack or Microsoft Teams.
Then change what you count. Every free trial credit, every token allotment you give away to get someone to try the product, has to be added to your ad spending. Swulinski calls this fully loaded acquisition cost. If you leave the free usage out, you are not really calculating what it cost to win the customer.
Then change the other side of the equation too. Lifetime value is not enough, because revenue is not profit when each answer costs you money. An AI company pays for the large language models it uses, for computing platforms, and for the work of running those models for each user. Swulinski said the token bill, including providers and inference costs, is often the biggest cost item, and founders need a finance leader to catch the hidden items they miss.
That matters because usage varies wildly. One new user might turn into a 50 dollar account and the next into a 50,000 dollar a year contract, which confuses ad systems like Meta and Google that were built for simpler, more predictable purchases. The takeaway is to compare fully loaded acquisition cost with lifetime value after all those usage costs, or lifetime value to gross profit. At a very high priced product like Victor, Swulinski said, you can pay a couple of thousand dollars to win one customer and the economics still work.
Questions people ask
Why do Meta and Google struggle with AI software?
Their ad systems learn from the purchase value you report back to them, and they look for more people like your best buyers. With usage based AI tools, one signup can be worth almost nothing and the next can be worth tens of thousands of dollars a year, so the average tells the system very little. That makes automated bidding and targeting much noisier than with a fixed price product.
What exactly are token and inference costs?
When a user prompts an AI product, the company sends that request to a large language model and gets text, voice, or another output back. Providers charge by the amount of text processed, measured in tokens, plus the computing needed to run the model, often called inference. High use customers can run up large bills on services from model makers and cloud computing providers.
How much should a startup pay to acquire a customer?
Swulinski said most companies aim over time for lifetime value worth about three times acquisition cost. A well funded startup may accept a one to one ratio early to win distribution, because distribution is how you outcompete clones. At a high priced product like Victor, he said a couple of thousand dollars in acquisition cost can still work.