How to Price AI Features in Your SaaS Product (We Got It Wrong Twice Before Getting It Right)
The first time we priced our AI feature, we added $10/month to the Pro plan. The reasoning: AI is valuable, $10 seems fair, let's see what conversion looks like. We never checked what the feature cost to serve. It was averaging $8.40 per Pro user per month. We were making $1.60 in margin on our most-used feature. We celebrated when it drove upgrades.
The mistake almost every team makes first
Pricing a feature on perceived value without knowing the cost to deliver it. For most software features, this is fine — the marginal cost is near zero. AI is different. Every request costs money. A heavy user might cost 20× more than a light user for the same plan. If your pricing doesn't account for that variance, the users who love the feature most will be your least profitable customers.
The data you need before pricing any AI feature: not the average cost per user, but the distribution. The average lies. You need to know that 15% of your users drive 70% of your AI costs — because those are the users your pricing decision is actually about. Flat pricing is a subsidy from your light users to your heavy ones. You need to know how large that subsidy is before you decide if it's worth it.
The second mistake: identical AI access across all tiers
After we found the margin problem and raised the Pro price, it helped — but didn't fully fix it. Our enterprise customers were using 6× more AI per session than Pro customers, on contracts that were only 3× the price. We had priced for the revenue uplift of moving to Enterprise, not for the cost increase that came with the heavier usage.
The fix was straightforward once we had the data: rebuild the enterprise tier with a monthly AI usage allowance and per-unit overages above it. Enterprise customers generating $5k in AI costs on a $3k/year contract is not a segment — it's a subsidy. We rebuilt the pricing model. Churn was zero. Several accounts even upgraded to higher allowances voluntarily, because the usage data finally made the value visible to them too.
What you actually need to price AI correctly
Three numbers per customer segment: average AI cost, the distribution of that cost (especially the high-usage tail), and the revenue each segment generates. Pricing AI features without those three things is guessing. The answers almost always surprise in the same direction — your highest-engagement users are cheapest to acquire and most expensive to serve.
Once you have that data, the decisions follow naturally: which features to gate at higher tiers, where usage-based overages make sense, which segments need per-customer caps, and where you have healthy margin to lean into. None of this is obvious from a total spend number. All of it becomes obvious when you can see cost by customer and feature, matched against the revenue each generates.
That is the view PerUnit provides — AI cost by customer, feature, and pricing tier, connected to your Stripe revenue so you can see margin alongside spend. We built it after making the mistakes above so other teams don't have to. Our free AI margin calculator shows you cost per customer and AI as a percentage of revenue as a starting point.