Why freemium doesn't work with AI products

Freemium was the growth engine for nearly every relevant SaaS company of the last fifteen years. Dropbox, Slack, Notion, Figma. The logic was simple: let users try for free, let them get hooked, then charge when they need more.

It worked because the marginal cost of a free user was negligible. Servers, storage, bandwidth. Pennies. You could scale free users infinitely until some converted.

With AI, that math collapses.

The structural problem: marginal cost is no longer negligible

In traditional SaaS, serving one more user costs nearly nothing. The infrastructure is already running. The code is already written. A free user takes up a negligible fraction of resources.

In AI products, every user interaction consumes tokens from a language model. Each generated response has a direct cost, proportional to usage.

Nikhyl Singhal, former VP of Product at Facebook and Chief Product Officer at Notion, put it directly in an interview on Lenny’s Podcast: the cost of serving an AI user isn’t fixed. It’s variable. And that changes the entire freemium equation.

Think of an AI support agent that answers questions. Each question is an API call. Each detailed response consumes processing. A user who heavily uses the product—exactly the behavior you want in traditional freemium—is now the user burning through your budget the fastest.

The model that financed growth now finances losses.

What product companies are doing differently

Looking across the market, three patterns are emerging among companies navigating this transition more cleanly.

1. Usage limits as the primary gate

Instead of restricting features, restrict the volume of usage. Users get access to the full product, but with a cap on interactions or credits.

Notion implemented this with Notion AI: you can test the functionality, but ongoing use requires a subscription. It’s not a separate feature hidden behind a paywall. It’s the real product, with a meter.

This preserves the trial experience—users understand the value before paying—without creating an infinite cost hole.

2. Pricing based on value delivered, not seats

The traditional SaaS model charges per user. It makes sense when value scales with internal adoption. More people using Slack means more value for the company.

With AI, value doesn’t come from the number of people accessing it. It comes from the result the AI generates. An agent that automates contract review can serve one person and deliver value equivalent to an entire team.

Companies like Intercom (with Fin) and Stripe itself with its AI products are experimenting with models where pricing reflects outcomes: tickets resolved, tasks completed, documents processed.

The logic shifts from “how many seats do you need” to “how much value did you extract.”

3. Trials with intentional friction

In traditional SaaS, the mantra was reduce friction at all costs. Pure self-service. Signup in two clicks. No sales involved.

For AI products with high marginal costs, some companies are reintroducing friction strategically. Not to make it harder, but to qualify.

Trials that require a credit card. Onboarding with a human before unlocking full access. Aggressive limits on the free plan that force a conversation with sales to expand.

This filters out users without real intent to pay before they consume significant resources.

Traditional freemium

• Negligible marginal cost
• Feature-based gating
• Per-seat pricing
• Total self-service
• Conversion over time

Models adapted for AI

• Marginal cost proportional to usage
• Volume/credit-based gating
• Pricing based on value delivered
• Intentional friction to qualify
• Fast conversion or controlled costs

The risk of copying old playbooks

It’s tempting to look at freemium’s historical success and replicate it. After all, Notion, Figma, and Slack proved the model. Why wouldn’t it work with AI?

Because cost structure is different. And cost structure defines business model, not the other way around.

An AI startup offering unlimited free usage is betting it can capture enough money to finance the losses until conversion happens. In an abundant capital environment, that might work for a while. In a capital-scarce environment (where we are now), it’s a runway burn.

• Is your cost per active user proportional to usage, or fixed?
• Do you know the exact cost of each AI interaction in your product?
• Does your free plan have a clear usage limit?
• Does the value you deliver scale with seats or with outcomes?
• Are you qualifying users before giving them unrestricted access?

What this means for builders right now

If you’re building a product with integrated AI, monetization isn’t a secondary decision. It defines business viability.

As I see it, the safest path is:

Understand real costs first. Before designing plans, know what each type of usage actually costs. Not estimates. Real numbers.

Design the free plan with a ceiling. If you’re offering free access, make the usage limit clear. Enough to demonstrate value, not enough to sustain continuous use without payment.

Consider outcome-based pricing. If your product delivers measurable results—documents processed, tasks automated, tickets resolved—pricing can reflect that. It aligns incentives: you only win when the customer wins.

Don’t fear friction. Qualifying users before granting full access isn’t hostility. It’s intent filtering. Anyone with a real problem to solve accepts one extra step.

Freemium didn’t die. It evolved.

The core insight of freemium—let users experience value before paying—remains valid. What no longer works is the traditional implementation: unlimited free usage funded by zero marginal cost.

With AI, freemium survives as a limited trial, as starter credits, as controlled demonstration. Not as a growth model at any cost.