The excitement around AI features has a rude awakening baked in: inference costs. Every time a customer uses your AI-powered feature, you're paying Anthropic, OpenAI, or Google for compute. At low usage, this is negligible. At scale, it becomes a margin problem that most SaaS CFOs are now grappling with for the first time.
The unit economics of AI-powered SaaS are genuinely different from classic SaaS. Traditional SaaS benefits from near-zero marginal cost — your 10,000th customer costs almost nothing more to serve than your 100th. AI features break this model. The 10,000th customer using your AI features costs as much as the 100th, or more.
This creates a gross margin compression that's showing up in company financials in 2026. Companies that launched AI features as a competitive necessity, without pricing them in, are now seeing blended gross margins drop from 80%+ to the low 70s or even high 60s.
The response requires discipline in three areas:
Usage controls and limits. Not every AI feature should be unlimited. Tiered limits based on plan level are fair, transparent, and margin-protective. If you're giving enterprise AI usage the same limits as SMB, you're subsidizing your largest customers.
Prompt and context optimization. The cost of an AI query scales with context length. Teams that invest in smart context truncation, caching repeated queries, and prompt optimization can cut inference costs by 40-60% without degrading output quality.
Pass-through pricing honesty. The premium tiers that include advanced AI features need to price in real inference costs, not historical compute benchmarks. If your enterprise tier was priced before you shipped AI features, the economics are probably wrong.
The CFOs asking hard questions about AI ROI aren't being backward-looking. They're being rigorous. The SaaS teams that answer those questions with real numbers — not just "AI is the future" — are the ones that survive the scrutiny.
Build the margin model before the CFO asks for it.