The default state: AI is a tax
Most SaaS teams shipped AI features on a flat plan, eating provider costs to ride the hype. It worked while usage was light. It stops working the moment a power user starts running agents — one customer can wipe out the margin of fifty others, and there's no signal in your P&L until the quarter closes.
Flip the script in four moves
- Meter every token — by user, workspace, and model.
- Bring your own provider key — direct OpenAI/Anthropic pricing, not gateway markups.
- Add a metered line to the invoice — Stripe Meters or equivalent.
- Show the meter to the customer — caps they control, no surprise bills.
One thing this playbook is not about
Command+K itself does not support BYOK: every call the widget makes runs through our managed AI Gateway on prepaid credits, which is what keeps per-model cost reporting exact. The advice above is about how you price your own product, not about how you pay us.
The shape of a healthy AI P&L
Direct cost ≤ 30% of metered revenue. Power users pay more, light users pay less, and the median customer never thinks about the bill.
Pricing psychology that works
- Include a generous free allowance in every plan — covers the median user.
- Show usage live, not at end-of-cycle — predictability beats price.
- Let customers set their own caps — they self-select into upgrades.
- Frame overage as "extra credits," not "overage charges."
Ship it in a week
- Day 1-2: token metering + per-user aggregation.
- Day 3: Stripe meter + line item on subscription.
- Day 4-5: in-product usage bar + cap controls.
- Day 6-7: cohort margin dashboard.
The Monetize layer bundles the metering, Stripe wiring, and caps so you skip straight to the dashboard.
FAQ
- How do I know if my AI feature is a cost center?
- If your gross margin on AI usage is negative or unmeasured, it's a cost center. Pull the last 30 days of token spend, divide by the count of paying customers using the feature, and compare to incremental revenue. If you can't pull those numbers in an hour, that's the first problem to fix.
- Does usage-based pricing scare customers off?
- Only when it's opaque. Hybrid pricing — a flat seat fee plus metered AI usage with a visible cap — converts the same as flat pricing in B2B SaaS, and grows ARPU 20-40% on power users. Show the meter, ship the cap, and let customers raise it themselves.
- What margin should I target on AI features?
- 60-75% gross margin is the SaaS benchmark. With direct provider keys and metered billing, you can hit that. With pass-through gateway pricing and flat plans, you'll grind toward 0% the moment a power user shows up.
- What's the minimum instrumentation to flip the script?
- Token-level usage records keyed by user and workspace, a Stripe meter that ingests them, a per-user cap with a visible bar in your product, and a dashboard showing margin per cohort. That's the whole MVP.