Pricing AI features is a strategy problem, not a pricing problem. The mechanics — per-seat, usage-based, tiered — are well-understood. The harder question is which pricing model aligns incentives with the actual value the AI delivers, without breaking the buyer's expectations or your unit economics. This post is the decision framework, the model-by-model tradeoffs, and what we've seen work.
The core tension
Your cost (model API, infrastructure) is per-call or per-token. Your customer's value is usually per-task or per-outcome. Pricing needs to bridge these without either of you taking outsized risk.
Get it wrong in one direction: you eat the cost when customers are heavy users, unit economics collapse. Get it wrong in the other: your customers dislike the variable bill, your sales cycle stretches, larger customers negotiate you back into predictable pricing anyway.
The main pricing models
Included in premium tier
AI features included with the highest tier of the existing SaaS product. Push customers toward upgrading; don't meter the AI itself. Simple, sales-friendly.
When it works: AI is a differentiator for the tier, not the full product. The cost distribution is predictable (not too many heavy users). Examples: GitHub Copilot included with Enterprise tier; many 'AI add-ons' that are really 'premium tier' by another name.
When it breaks: heavy users generate runaway costs you can't cap without customer friction. Mitigation: internal fair-use limits with escalation paths, often implicit at first and explicit if costs get out of hand.
Usage-based (per call / per task)
Directly meter the AI work. Cost aligned with usage. Transparent. Common in developer-tool AI, where the audience understands and accepts usage pricing.
When it works: technical audiences, spiky usage patterns, enterprise buyers who can budget for variability. Strong for developer platforms.
When it breaks: non-technical buyers fear the bill. Procurement teams reject variable costs. Unit economics on your side depend on customer concentration — one heavy user can eat most of your margin.
Credit packages
Buyer purchases a credit pack ($X for N credits); credits are consumed by AI actions. Predictable for buyer; cost-aligned for seller.
When it works: consumer and prosumer products, seat-based products where AI usage needs to be separately metered. Creative software, content generation.
When it breaks: customers with unused credits feel cheated. Customers who run out mid-workflow are frustrated. Both are CSAT hits if credit-to-action mapping isn't obvious.
Per-seat AI add-on
Existing SaaS product adds an AI-specific add-on priced per seat ($X per user per month). Familiar to SaaS buyers; easy to forecast.
When it works: workflow-integrated AI that benefits all users roughly equally. Sales productivity tools, customer support agent assistance.
When it breaks: usage is highly skewed (20% of users generate 80% of AI load). You either over-price for light users or under-price for heavy ones. Consider combining per-seat with usage caps per seat, escalating for heavier plans.
What works in practice
Most successful AI pricing is hybrid: per-seat base price, with usage tiers layered in. A standard seat gets N AI actions per month; heavy users upgrade to a higher tier with M actions; extreme users pay for additional usage. This aligns incentives across customer types while keeping the core pricing conversation familiar.
Second-best pattern: free trial with clear 'upgrade moment' driven by hitting limits. Users explore AI without friction; paid upgrade happens when they're already hooked. Critical: the limits must feel fair (enough to feel valuable, not enough to frustrate).
Common mistakes
Pricing the model cost, not the customer value. A 3-minute task that would take a human 30 minutes isn't worth the model cost; it's worth the saved time.
Free AI features to drive adoption. Works for consumer products; breaks unit economics in B2B unless funded by existing paid tier.
Over-complicating the pricing page. A pricing page with 6 axes and N tiers loses customers. Keep the base structure simple, handle edge cases in enterprise contracts.
Under-pricing based on projected model cost reductions that don't materialize as fast as expected. Build in margin for the cost curve to flatten temporarily.
Not having a value story. 'Our AI costs $X' is weak. 'Customers saving Y hours per week' is strong. Every salesperson needs the calculator. See our sales enablement post.