Software valuations have shifted. Recurring revenue alone no longer commands the multiples it did a few years ago, and the SaaS companies holding their ground are the ones building AI directly into the product – usually to surface insights that make the tool meaningfully smarter, rather than bolting a chatbot on the side.

It’s a sensible move. It’s also an expensive one.

In this article, we look at why AI inference costs are a structural challenge for SaaS margins, what that means for your numbers, and the three pricing approaches we see founders use to recover those costs without losing customers.

Why AI costs are different

Spend on Claude, ChatGPT and Gemini is climbing month on month, on two fronts: developers using AI to ship faster, and the models running inside the product itself. While AI token usage prices dropped sharply over the past year, total AI spend still rose several times over, because usage grew far faster than unit costs fell. The more useful your AI feature is, the more it gets used, and the more it costs you.

Here’s the part that matters for your numbers. AI inference – the cost of running the model every time a customer uses the feature – is now a variable cost of goods sold. That’s new for most SaaS businesses, which are used to near-zero marginal cost per user. Under a flat per-seat price, every extra bit of AI usage eats your gross margin, and your best, most engaged customers quietly become your least profitable.

It’s worth putting a number on the threat. As SaaS companies scale, the long-standing mantra is to target gross margins of 80-85%. That’s the bar for a clean, pure-software business, and it’s the profile investors reward with the highest multiples. AI bends that maths. ICONIQ’s early 2026 survey of around 300 software executives put average AI-native product gross margins at 52% for the year – a long way short of the old benchmark. The trajectory is improving, up from 41% two years earlier as teams get better at managing inference, but the gap to 80% is structural, not a bad quarter.

Most founders reading this are not building AI-native. They’re adding AI to an existing product, and the evidence there is sharper. Layering AI features onto a traditional SaaS product typically costs 12-17 points of gross margin. The simplest way to picture it: take an $80 seat running at an 80% margin, add about $15 of inference, model routing and supporting infrastructure, and that seat is suddenly closer to 65%. Worse, this cost doesn’t fade as you grow the way hosting did. Every query runs the model again, so inference scales with usage instead of shrinking to a rounding error at volume.

Left alone, a popular AI feature can drag an 80% business steadily towards 60%.

The three approaches to recovering AI costs

So the question we spend a lot of time on with founders is simple to ask and harder to answer: how do you recover these costs and pass them on in a way customers accept?

Broadly, there are three approaches.

1. Per-seat price increase

The simplest approach – the “now with AI” upgrade. You raise the price and fold AI in as added value. It’s the easiest to sell and it keeps revenue predictable, which customers and your board both like. The weakness is that it decouples price from cost. If usage is light and fairly even across your base, the maths works. If a handful of power users hammer the AI, you’re subsidising them.

2. Metered or credit-based usage

Pass the variable cost through directly, charging customers based on what they actually use. This protects your margin regardless of usage levels and aligns what you charge with what you spend. The downside is buyer friction – unpredictable bills are hard to budget for, and anything that looks like pay-as-you-go creates hesitation in procurement, particularly at enterprise level.

3. Hybrid model

A base subscription with an included usage allowance, then overage or credit packs beyond it. Or an AI tier sitting above your core plans. This gives customers a predictable floor and gives you margin protection on the upside. It’s quietly become the dominant approach for exactly this reason – it solves for both the customer’s need for predictability and the founder’s need to not get destroyed on margin by heavy users.

Price the value, not the cost

Whichever model you choose, the test isn’t your cost – it’s value. Customers won’t pay more because your token bill went up. They will pay when what you’re charging maps to value they can see: the insight surfaced, the hours saved, the work the AI did instead of them.

Price the value metric, not the raw cost underneath it. Get that right and usage-based pricing feels fair rather than punitive.

Practically, three things. As best you can, deconstruct your AI costs by customer and by feature. Model your gross margin at low, medium and heavy usage before you set a price, not after. And understand what value each customer segment is actually getting from the AI features they use most.

Oh, and moving to a usage-based model also changes how you track and report ARR – which is still what investors look at.

AI in the product is increasingly the price of staying relevant

Making sure it strengthens your margin rather than quietly eroding it is a pricing decision, and it’s one worth getting right.

Disclaimer: This article is for general informational purposes only and does not constitute financial, legal or tax advice. Please speak with a qualified adviser (hey, that’s us!) before making decisions based on your specific circumstances.