How to Price AI Features in SaaS Without Destroying Your Margin (2026)
Per-seat pricing breaks the moment inference costs scale with usage. Five pricing models, with the margin math for each.
Classic SaaS had near-zero marginal cost per user. AI features do not. Ship an unlimited AI assistant on a $29 seat and your top 5% of users will quietly consume more inference than the plan is worth.
Know your unit cost first
Before pricing, instrument cost per action: tokens in, tokens out, model tier, retries, and embeddings. Then compute cost per active user per month, not per request. Typical 2026 ranges we see:
| Feature type | Cost / active user / month |
|---|---|
| Inline suggestions, short completions | $0.10–$0.80 |
| Chat copilot over your data (RAG) | $0.90–$4.50 |
| Multi-step agents with tool calls | $4–$40 |
| Long-document or video processing | $8–$120 |
The five models
- Bundled into the seat. Simplest, best for low-cost features. Only safe when your p99 user costs less than 20% of seat price. Add a fair-use ceiling.
- AI add-on per seat. A separate $15–$40 AI seat. Clean margin protection, easy to sell, but you now have an adoption problem: buyers only add it for users they are sure will use it.
- Credits. Universal currency across features. Flexible and margin-safe, but customers hate opaque credits — publish a plain-English table of what each action costs.
- Usage passthrough with markup. Honest and scalable for developer-facing products, hopeless for non-technical buyers who need budget predictability.
- Outcome pricing. Charge per resolved ticket, per qualified lead, per processed document. Highest willingness to pay, hardest to measure, and it demands airtight attribution.
The hybrid that usually wins
Seat price for access, an included credit allowance sized to the 70th-percentile user, and metered overage at a visible rate. This protects margin against heavy users, keeps the bill predictable for typical ones, and gives your expansion revenue a natural mechanism.
Margin rules of thumb
- Target 70%+ gross margin on AI features, not the 85–90% of classic SaaS.
- Cap the worst case: hard limits per account per day, with a clear upgrade path when hit.
- Route by difficulty. Sending every request to your most expensive model is the single biggest margin leak — a cheap model handles 60–80% of traffic with no quality loss on easy inputs.
- Cache aggressively. Prompt and response caching typically cuts 25–45% off a mature workload.
Announcing a price change
Grandfather existing customers for a full billing cycle, show them their actual usage against the new allowance before it takes effect, and never introduce credits and a price rise in the same email. The churn from AI repricing is almost always a communication failure, not a value one.
Bottom line
Price on value, defend on cost, and instrument before you decide. You cannot pick a pricing model for a feature whose unit economics you have never measured.
FAQ
Frequently asked questions
Should AI features be included in the base plan?+
Only when your 99th-percentile user costs less than about 20% of the seat price. Above that, use an add-on or credit allowance.
Do customers accept credit-based pricing?+
Yes, if the conversion is transparent. Publish a table of what each action costs in credits and show real-time consumption in the product.
What gross margin should I target on AI features?+
Around 70% is healthy in 2026. Model routing and response caching are the two levers that move it most.
Is outcome-based pricing worth it?+
It commands the highest willingness to pay but requires airtight attribution and dispute handling. Start with hybrid credits and move to outcomes once measurement is trusted.
Building something similar?
Let's talk in 30 minutes.