Traditional software has a comfortable economic shape: build it once, serve it a million times for almost nothing. Pricing per seat works because an extra user costs you close to zero. AI features break that shape completely. Every summary generated, every document parsed, every agent run consumes compute you pay for — and your heaviest users consume the most.
Which is why so many teams shipped an AI feature in 2026, celebrated the adoption numbers, then discovered gross margin had quietly slipped several points. The feature worked. The pricing did not follow it.
Under flat per-seat pricing, cost scales with usage while revenue scales with headcount. The two used to move together; now they do not. Worst of all, the correlation inverts: your most engaged customers, the ones least likely to churn and most likely to expand, are the ones destroying your unit economics.
Before you price anything, measure. Instrument cost per AI action, then cost per active user per month, then gross margin for that feature in isolation. Most teams are surprised by the distribution — a small share of accounts typically drives the majority of spend.
Roughly 85% of SaaS companies now use some form of usage-based billing, but pure models are rare because pure models each have a fatal flaw.
Simplest to sell and easiest to forecast, and it works when usage per user is genuinely bounded — a feature someone touches a few times a week. It fails the moment power users appear or an agent starts running in a loop.
Margins are safe because cost and revenue move together. The problem is commercial: customers hate unpredictable bills, procurement struggles to approve them, and unpredictability suppresses the very usage you want to encourage.
The pragmatic default for most products. A predictable platform fee covers access and a generous included allowance; heavy usage draws from credits bought in blocks. Customers get a floor they can budget for, you get protection at the ceiling.
Charging per resolved ticket, per qualified lead, per completed workflow aligns beautifully with value as agents take over whole units of work. It also requires airtight definitions of success and mature instrumentation. Powerful when it fits, painful when the definition is arguable.
Bill shock is the leading churn driver in usage-priced products, and it is almost entirely preventable. The failure is never the price itself; it is the surprise. A customer who is warned at 75% of their allowance and offered an upgrade path stays. A customer who opens an invoice three times larger than expected cancels, and tells people why.
Build the usage meter into the interface from day one. Show what has been consumed, what remains, and what the next block costs. Transparency here is retention work disguised as billing work.
Customers will accept paying more for what they use. They will not accept finding out afterwards.
The industry's most resented move this year has been the mandatory AI tier — forcing existing customers onto a more expensive plan to keep functionality they already had. It generates a short-term revenue bump and long-term ill will.
Better: keep existing functionality where it is, introduce AI capability as an addition with its own clear value, and give current customers a genuine trial allowance before asking them to pay. If the feature earns its keep, adoption will do the upselling for you.
Measure cost per action before you price, pick the hybrid model unless you have a strong reason not to, meter in units humans understand, warn before the bill arrives, and never hold existing functionality hostage. AI features can be a healthy margin business — but only if the pricing model changes shape along with the product.
We design and build SaaS products — including the pricing, metering and usage interfaces that keep AI features profitable and customers calm.
Talk to us