Seats, Usage or Outcomes: Pricing an AI Product Without Losing Your Margin

Per-seat pricing was the best business model software ever had. Each new user cost almost nothing to serve, so every seat sold was nearly pure margin. AI breaks that arithmetic. Every question, generation and agent step costs real money in inference, and the customers who love your product most are the ones who cost you the most.

The market is reacting. In Tremont's 2025 State of Monetization report, the share of SaaS companies using seats as their primary model fell from 21% to 15% in twelve months, while hybrid models (a subscription plus usage) rose from 27% to 41%. Kyle Poyar's survey for the same report found usage-based pricing at about 20% and outcome-based pricing at about 3%, with both expected to grow.

Why the seat breaks

Model a hundred-seat customer paying $30 a seat. If each seat runs a few hundred AI actions a month at two cents each, you keep most of the revenue. If the customer adopts your product the way you hoped and usage triples, your margin collapses, and the seat price has not moved. Your best customer has become your least profitable one.

Try it: push usage to 2,000 actions and model cost to 3 cents. The seat model goes negative. The usage model keeps its margin because revenue follows cost. The hybrid model sits in between, which is exactly why buyers and sellers have settled on it: the customer gets a predictable budget line, and the vendor is protected from heavy users.

The four shapes, and when each is honest

Seats still fit when the value comes from people working together and AI is a small part of the cost. Charging for AI as an add-on per seat is the usual compromise.

Usage or credits fit agents, APIs and generation tools, where the work done is the value. The risk is bill shock. Hard caps, alerts and a clear dashboard are part of the product, not an afterthought.

Hybrid fits most B2B products in 2026: a platform fee, a generous included allowance and a published overage rate.

Outcomes, such as charging per resolved support ticket, are the most appealing and the hardest to run. They are honest only when both sides agree on what counts as a result and who caused it. If you would argue with a customer about attribution, do not price on it yet.

Pricing is a product decision

The teams I have seen get this right treat pricing like a feature with its own discovery work. They interview buyers about how they budget, not just what they would pay. They instrument cost per action from day one, because you cannot price what you cannot measure. And they pick a value metric customers already understand: conversations, documents, minutes or seats, rather than "tokens", which nobody outside engineering can forecast.

This connects to two arguments I have made before. In When Code Is Cheap, What Still Protects a SaaS? I argued that features no longer protect a business. Pricing that lines up with customer value is one of the few things competitors cannot copy over a weekend, because it depends on knowing your customers' economics. And in When Building Gets Cheap, Product Judgment Gets Expensive, the point is that choosing what to build is now the scarce skill. Choosing how to charge for it belongs to the same job.

A pre-launch checklist

• Know your cost per action at today's model prices, and at half and double them.
• Model margin for your median, heavy and extreme customers, not the average.
• Pick one value metric and test it in five sales conversations before you build billing.
• Ship spend caps and usage alerts in the first version.
• Plan a price review for when model costs fall. They have kept falling, and customers will notice even if you do not.

Treat any change as a smaller bet: grandfather existing customers, try the new model on new accounts first, and decide in advance what result would make you reverse it.

Sources

• Tremont / Kyle Poyar: 2025 State of Monetization report
• Schematic: Kyle Poyar on the state of AI pricing in 2025
• Futurum: enterprise buyers prefer consumption and outcome-based pricing

Across the series

• The European Accessibility Act Is Live. Here’s What It Asks of Your Product.
• Your Visitors Already Said No. Is Your Site Listening?
• Your AI Chat Box Is Probably Unusable With a Screen Reader
• Can You Train AI on Your Customers’ Data? Ask Them First
• Roadmapping When the Model Gets Better Every Seven Months
• Accessibility Testing: A Green Scan Is Not a Finished Journey
• AI Data Retention: ‘Not Used for Training’ Does Not Mean Deleted
• Agent Evaluation: The Demo Worked. Can You Ship It?