Here's a mistake I see founders make when they add AI to a product: they copy the free tier playbook from regular SaaS.
In regular SaaS, a free user is nearly free. They store a few rows in a database and load a few pages. You can carry thousands of them without noticing on your bill.
AI breaks that assumption. Every message a free user sends triggers a paid call to a language model. A free tier isn't a marketing gift anymore. It's a variable cost that grows with your success, and if you don't design it carefully, growth is exactly what bankrupts you.
I ran into this while building Chatbotz, a platform where businesses create an AI chatbot and put it on their website with one line of code. Here's the framework I use now.
Step 1: Decide who pays for the tokens
Before pricing, before limits, before the landing page, answer one question: who pays the model provider?
There are three honest options:
- You pay for everything. The simplest experience for the customer, and the riskiest for you. Your costs scale with usage, whether or not the customer ever pays.
- The customer brings their own API key. Your costs stay predictable, but setup is harder, and non-technical users may not know what an API key is.
- A hybrid. A small free allowance on your account, then the customer connects their own key or upgrades.
Chatbotz supports five providers (OpenAI, Groq, OpenRouter, Anthropic, and Gemini), with per-bot model selection and API key management. That flexibility was possible because the model layer was designed around it from the start.
My take: for most early-stage AI products, the hybrid wins. It keeps onboarding smooth without leaving your bill open-ended.
Step 2: Pick a free-tier model on purpose
Not every conversation needs the most powerful model. A chatbot that answers "what are your opening hours?" doesn't need a frontier model.
That's why the Chatbotz free tier uses Groq as its default provider. The reasoning is simple: a fast, low-cost default keeps the free experience snappy and the cost per conversation low. Premium models become something you unlock, not something you give away.
My take: treat the free-tier model as a pricing decision, not a technical one. It's the biggest lever you have on your unit economics.
Step 3: Set limits that protect you without frustrating users
Limits feel like a negative. They're actually how you keep the product alive. These are the five I'd put in place before launch:
- Messages per month. The hard ceiling per bot or per account.
- Context length. Models are stateless, so the whole conversation history is resent with every message. Long chats cost far more than short ones. Cap the history or summarize older turns.
- Response length. A maximum output size keeps one chatty answer from becoming an expensive one.
- Rate limiting. Limit requests per visitor and per origin, so one script can't drain a month of budget in an hour.
- Number of bots. Cap how many bots a free account can run.
The third and fourth matter most for products like Chatbotz. Each bot sits behind a public chat endpoint, which means anyone on the internet can talk to it. Without rate limits, an abusive visitor can burn through a customer's allowance, or yours.
Step 4: Do the napkin math before you ship
You don't need a spreadsheet. You need four numbers:
Monthly free-tier cost = cost per conversation × conversations per free user × number of free users
Then compare it to your conversion rate (how many free users become paying users) multiplied by what a paying user is worth.
Here's an illustration with made-up numbers, just to show the shape:
- Suppose a conversation costs $0.01.
- A free user has 50 conversations a month.
- You have 2,000 free users.
That's $0.01 × 50 × 2,000 = $1,000 a month to run the free tier. If 3% convert at $20 a month, that's 60 paying users, or $1,200 a month in revenue. It works, but barely, and it falls apart fast if costs double or conversion halves.
These numbers are illustrative, not real pricing. Plug in your own.
My take: if the free tier only works when everything goes right, the free tier is too generous.
Step 5: Know what a free tier is actually for
A free tier has two jobs: let people see value fast, and give them a reason to pay.
Give away enough to prove the product works: one bot, a limited number of messages, a real conversation on a real website. Then keep the things serious users need behind the upgrade: more bots, higher limits, premium models, and removing branding.
The test I use: can a user get a real "this works" moment in the first ten minutes? If yes, the free tier is doing its job. If they hit a wall before that moment, it's too stingy. If they never need to upgrade, it's too generous.
The checklist
Before you launch an AI free tier, make sure you can answer these:
- Who pays for the tokens: me, the customer, or both?
- Which model runs the free tier, and why?
- What are my limits on messages, context, response length, and requests?
- What does the free tier cost me per month at 100, 1,000, and 10,000 users?
- What makes a free user want to upgrade?
If you can't answer all five, you're not ready to launch. You're ready to run an expensive experiment.
The takeaway
An AI free tier isn't a gift. It's a bet, and like any bet, it needs a ceiling.
Pick the model on purpose, limit what can hurt you, do the math, and design the upgrade path before you design the pricing page. Done well, a free tier becomes your best acquisition channel. Done carelessly, it becomes your biggest line item.