Old software had a beautiful secret. Once you built it, the next customer cost you almost nothing. That is why normal SaaS companies keep 75 to 85 paise of every rupee they earn.

AI broke that secret.

Now every time a user clicks your AI button, you pay the model provider. Real money, every single time. Recent industry surveys put average gross margins on AI products near 52% — far below what SaaS founders are used to.

So the question is no longer "should we add AI?" It is "what do we charge so we don't lose money on our best customers?"

The problem is not the average user. It is the heavy one.

Imagine you charge ₹1,500 per month, flat.

Most users click your AI feature five times a month. It costs you ₹20. Wonderful.

But one user runs it 400 times a month. That user costs you ₹1,600. You are now paying for the privilege of serving them — and they are usually your happiest, most loyal customer.

Flat pricing hides this. It spreads one expensive user's cost across everyone else. That works until you have twenty of them.

What real companies did about it

Intercom built an AI support agent called Fin. Instead of charging per user, they charge $0.99 every time the AI fully solves a customer's problem. If it fails and a human steps in, the customer pays nothing.

The result: Fin went from about $1M to over $100M in yearly revenue on this model.

Here is the honest side of that math. A company handling 30,000 chats a month, where the AI solves 60%, pays around $17,820 a month. The bill grows as the AI gets better. Customers accept it because they can see what they bought.

Others followed. Zendesk charges around $1.50 per solved ticket. Salesforce charges about $2 per conversation. HubSpot actually cut its price to $0.50 in April 2026 to win customers.

Meanwhile, per-seat pricing is shrinking fast — from 21% to 15% of SaaS companies in twelve months. Atlassian even reported its first-ever drop in enterprise seat count, because AI agents were replacing the people who needed those seats.

So what should you actually do?

Most small SaaS teams should not copy Intercom. Outcome pricing needs you to prove the outcome, measure it, and argue about it when a customer disagrees. That is expensive.

Start with hybrid. A fixed monthly fee, plus usage after a limit. This is now the most common model in the industry, used by roughly 41% of AI vendors, up from 27% a year ago.

In practice:

• Base plan covers your fixed costs and gives customers a predictable bill
• Included quota — say 500 AI actions per month
• Overage — a small charge above that, or an upgrade to the next plan

Two rules that matter more than the model you pick:

First, never sell unlimited AI. Agentic features can use 5 to 30 times more tokens than a simple chatbot reply. One customer can quietly destroy a month's profit.

Second, expect to change it. Around 92% of AI companies that charge by usage have already changed their pricing at least once. Build your billing so you can adjust the number without rewriting the system.

The part most founders get wrong

They pick a price first, then check the cost later.

Do it the other way. Measure what one completed task actually costs you — including retries, long documents and failed attempts — then price above it with room to breathe.

If you are not sure what your AI feature costs per use, that is the first thing to fix. We build AI features into existing SaaS products, and we work out this number before writing any code.

Talk to us about your AI feature → or WhatsApp +91-8902-969639.

Sources: Intercom, Zendesk, Salesforce and HubSpot published pricing (2026); Bessemer Venture Partners AI Pricing Playbook; ICONIQ 2026 margin survey; Gartner token-consumption estimates. Rupee figures are illustrative.