How to Price AI Features Without Losing Margin

The first AI bill is fine. The second one is fine. Then a customer rolls the feature out to their whole team, and finance asks why cost of goods sold jumped 40% in a month while revenue didn’t move at all.

This is the most common commercial mistake we see in AI product work: the feature shipped, the pricing didn’t change. Everything about the technical build was reasonable. The business model underneath it was still the one written for software that costs nothing to run.

Per-Seat Pricing Breaks on AI Features

A seat is a fixed price against a variable cost

Classic SaaS gets cheaper per user as it grows. Serving your ten-thousandth account costs roughly what the hundredth did. Inference doesn’t work that way — every summary, every agent run, every document parsed is a fresh charge on your card. There’s no economy of scale hiding at the top of the curve.

That means your power users, the ones you used to celebrate, are now your margin problem. In most AI features we’ve instrumented, the top 5% of accounts consume more than the bottom half combined.

The benchmark numbers have already moved

Through 2026, AI-first companies have been reporting gross margins in the 50–60% range, with inference alone accounting for a large slice of COGS. Mature SaaS sits at 75–90%. That gap isn’t a temporary inefficiency waiting for cheaper GPUs — it’s what happens when serving a customer has a real marginal cost.

Meanwhile most SaaS leadership teams say they’re moving off pure seat pricing within the next couple of years. If you’re building a SaaS platform with AI in it, you’re not early to this problem anymore. You’re on time.

Measure Cost Per Action Before You Pick a Price

The unit is the action, not the month

Pick the thing your feature actually does — one meeting summarised, one lead scored, one contract reviewed — and get the fully loaded cost of it. Model calls, retries, embeddings, retrieval, the second model you use to check the first one. Retries are where teams underestimate, sometimes by half.

You can’t price what you haven’t measured, and you can’t measure it without per-request cost logging in place. That’s observability work, and it needs to exist before pricing conversations, not after.

Price the tail, not the average

Averages will lie to you here. Take your cost distribution and look at the 95th percentile — the long document, the agent that looped four times, the customer whose data is messier than everyone else’s. Price so that account is still profitable, then use the margin from the median to fund growth.

Before you raise a price, check whether you can lower the cost instead. Prompt caching and routing cheap requests to smaller models often move unit economics further than a pricing change, and neither one requires a conversation with your customers.

Three Pricing Shapes That Actually Hold

Bundle into a higher tier

Simplest to sell, and fine when usage is predictable and light. You keep one price, AI becomes the reason to upgrade, and you cap exposure with a fair-use limit that’s written down and enforced in code. Skip the limit and you’ve just built an all-you-can-eat buffet with a variable food cost.

Sell credits

Credits are the model most teams land on, and hybrid pricing — a base subscription plus credits for AI — has become the default shape in the market. The advantage is flexibility: you decide what a credit buys. When a model gets 60% cheaper next quarter, you adjust the exchange rate quietly instead of renegotiating every contract.

Make credits legible. One credit should map to something a buyer can picture, like one document processed, not to a token count nobody can forecast.

Charge for outcomes

Charge per resolved ticket, per recovered lead, per approved application. It aligns beautifully with customer value, and it’s the hardest to execute — you need the outcome to be measurable inside your own system and undisputed by the customer. We built releadr’s AI lead-recovery engine around a countable event for exactly this reason. If the outcome needs a lawyer to define, use credits.

Build the Plumbing Before You Publish the Price

Metering, a ceiling, and a number finance reads

Whatever model you choose, three things need to exist in the product first: metering that attributes every call to an account, a hard ceiling per account that degrades gracefully rather than failing, and a margin dashboard your finance side actually reads. Without metering you’re guessing. Without a ceiling one customer can eat a quarter of profit. Without the dashboard nobody notices until the board meeting.

Pricing an AI feature is a product decision, not a spreadsheet exercise done afterwards. Decide the unit early, instrument it, then design the feature so the cost per unit is something you’d be happy to charge for.

If you’re adding AI to an existing product and want the unit economics worked out alongside the architecture, that’s how we scope AI builds — and it’s a much cheaper conversation before launch than after.

Yaniv Amrami is founder of quickdev. He has helped Israeli SaaS founders ship and price AI features across fintech, B2B, and marketplace products since 2017.