Most teams bolt an AI feature onto an existing plan and call it a differentiator. Then the inference bill arrives and the feature that was supposed to drive expansion is dragging down gross margin on every account that uses it.
The framework
- Measure the unit cost. Know the fully-loaded cost per AI action — tokens, retries, retrieval, the lot. You cannot price what you have not metered.
- Set a floor, not just a price. Establish the minimum you must charge per action to stay margin-positive, then price above it with headroom.
- Cap or meter heavy users. A flat add-on invites abuse. Either meter usage or cap it per tier so a handful of power users don't eat the margin of the whole cohort.
- Make the value legible. Tie the charge to a visible outcome — drafts produced, tickets resolved — so the price reads as fair, not punitive.
Where it breaks
The trap is "AI included free in every plan." It feels generous and demos well, but it converts your best margin line into your worst. Generosity is a pricing decision with a cost.
Bottom line: treat AI features as metered infrastructure with a margin floor — never as a free sweetener.