How to Price AI Features in Your SaaS: Usage-Based vs. Tier-Included

You built an AI feature. Users love it. Now you need to price it. This decision will determine whether your AI feature is a growth engine or a margin destroyer, and most SaaS teams get it wrong on the first attempt.

The core tension: AI features have variable costs that scale with usage, but SaaS customers expect predictable pricing. Every pricing model you choose is a bet on how you solve that tension.

The Four Pricing Models

There are exactly four ways to package AI features in a SaaS product. Each has different margin profiles, different customer expectations, and different failure modes.

Model 1: Included in Plan (Bundled)

The AI feature ships as part of your existing plan tiers. No separate line item. No usage limits (or very generous ones).

Who does this: Notion AI was initially a separate add-on at $10/user/month, then bundled it into all paid plans. Linear includes AI features in every plan.

Economics:

When it works: When AI usage per user is predictable and bounded. If your AI feature generates 50-200 tokens per interaction and the average user triggers it 5-10 times per day, you can model the cost reliably.

When it breaks: When power users consume 50x the median. One customer with 500 users generating 2,000 AI requests per user per day will eat your entire margin on that account.

The math: At GPT-4o pricing (~$2.50/1M input tokens, $10/1M output tokens), a typical bundled AI feature costs:

• Median user: 150 requests/month x 800 tokens = 120K tokens = ~$0.42/month
• Power user: 2,000 requests/month x 1,200 tokens = 2.4M tokens = ~$8.40/month
• If your plan is $20/user/month, the median user costs you 2.1% of revenue. The power user costs 42%.

Model 2: Usage-Based Add-On

Customers pay per AI interaction, per credit, or per token. Often sold as "AI credits" that map to some unit of work.

Who does this: Intercom charges separately for its AI agent (Fin) on a per-resolution basis. Jasper uses a credit system tied to word output.

Economics:

When it works: When the AI feature delivers clear, measurable value per interaction. If your AI resolves a support ticket (saving $5-15 in agent time), charging $0.99 per resolution is easy to justify.

When it breaks: When usage is exploratory or habitual. Charging per AI search query makes users hesitate before searching. Charging per AI suggestion makes users ignore suggestions. You train customers to avoid the feature you built.

The friction tax: Intercom reported that Fin's resolution rate is 50% higher on accounts with unlimited plans vs. per-resolution billing. Usage-based pricing suppresses the behavior you want to encourage.

Model 3: Separate AI Tier

A dedicated pricing tier that unlocks AI features. Often called "Plus," "Pro AI," or "Enterprise AI."

Who does this: GitHub Copilot is a separate $10-19/month subscription. Grammarly has a distinct "Premium" tier with AI features.

Economics:

When it works: When AI is a distinct capability that some users need and others do not. Developer tools, writing assistants, and analytics platforms where AI is a "mode" rather than a feature.

When it breaks: When the AI feature is integral to the core product experience. If your product feels incomplete without AI, gating it behind a tier creates a worse free experience rather than a compelling upgrade path.

Model 4: Freemium with AI Upgrade

Free users get limited AI access (5-20 interactions per month). Paid users get more or unlimited access.

Who does this: ChatGPT itself uses this model. Canva offers limited AI image generation on free plans.

Economics:

When it works: When AI is the primary reason users upgrade. The free limit must be high enough to demonstrate value but low enough to create real friction at the cap.

When it breaks: When free users consume significant AI resources without converting. If your conversion rate from free-with-AI to paid is below 3-5%, you are subsidizing usage that never monetizes.

The Margin Problem No One Talks About

All four models share the same underlying problem: if you are calling an external AI API, your AI feature margin is structurally capped.

Here is the math that most SaaS founders do not run:

Typical SaaS unit economics:

• Revenue per user: $25/month
• Non-AI COGS: $3/month (hosting, bandwidth, support)
• Gross margin before AI: 88%

After adding API-based AI features:

• Average AI cost per user: $2-8/month (depending on usage)
• New gross margin: 68-80%

That is a 8-20 percentage point margin compression. For a company doing $5M ARR, that is $400K-$1M in annual margin erosion. For a company raising Series B, that margin compression directly impacts your valuation multiple.

The problem compounds at scale:

Revenue scales with users. AI API costs also scale with users. Your margin percentage stays compressed forever.

How Fine-Tuning Flips the Margin Equation

Fine-tuning a smaller model (3B-7B parameters) and running it locally or on dedicated infrastructure changes the cost structure from variable to fixed.

API model: Cost = (tokens consumed) x (price per token). Linear scaling. No ceiling.

Fine-tuned model: Cost = (server cost) + (one-time training cost). Fixed monthly cost regardless of usage.

The crossover numbers:

At 50,000 queries/month — roughly 5,000 active users with moderate usage — the fine-tuned model costs 89% less. At 1M queries, it costs 98% less.

This is not marginal. This is the difference between a SaaS with 68% gross margin and one with 92% gross margin.

The Pricing Decision Tree

Use this framework to choose your pricing model:

Step 1: Is AI usage predictable per user?

• Yes (bounded interactions, consistent patterns) → Consider bundling (Model 1)
• No (highly variable, power users exist) → Continue to Step 2

Step 2: Does each AI interaction have clear, measurable ROI?

• Yes (resolves ticket, generates document, completes analysis) → Usage-based (Model 2)
• No (suggestions, search, assistance) → Continue to Step 3

Step 3: Is AI a distinct capability or integrated into core UX?

• Distinct (separate mode, optional workflow) → Separate tier (Model 3)
• Integrated (embedded in every interaction) → Freemium with limits (Model 4)

Step 4: Regardless of model chosen, what is your cost structure?

• Under 10,000 monthly AI queries → API is fine, pricing model is more important than infrastructure
• 10,000-100,000 queries → Run the crossover analysis. Fine-tuning likely saves 50-90%
• Over 100,000 queries → Fine-tuning is almost certainly the right infrastructure choice

What Notion, Linear, and Intercom Actually Did

Notion started with AI as a $10/user/month add-on. Adoption was moderate. They then bundled it into all paid plans and raised base prices by $2-3/user. Result: higher adoption, better retention, and roughly neutral revenue impact per user — but significantly better positioning as an "AI-native" workspace.

The lesson: bundling works when AI increases the core product's stickiness enough to offset the cost.

Linear included AI features from day one without a separate price. Their AI features (issue summarization, writing assistance, triage suggestions) are lightweight — each interaction consumes 200-500 tokens. At their price point ($8-10/user/month), the AI cost per user is roughly $0.30-0.80/month. Manageable.

The lesson: if your AI features are lightweight and bounded, bundling costs are trivial.

Intercom charges per AI resolution for Fin, their AI support agent. Each resolution replaces a human agent interaction worth $5-15. Charging $0.99 per resolution delivers clear ROI to the customer while maintaining healthy margins.

The lesson: usage-based pricing works when each interaction has obvious, quantifiable value that exceeds the price.

The Hidden Variable: Cost Infrastructure

Your pricing model and your cost infrastructure are two separate decisions, but they interact:

• API + bundled pricing = margin risk. You absorb variable costs at a fixed price.
• API + usage-based pricing = margin safe, adoption risk. Revenue and cost scale together, but friction reduces usage.
• Fine-tuned + bundled pricing = the ideal state. Fixed costs, fixed pricing, maximum adoption, predictable margins.
• Fine-tuned + usage-based pricing = pure profit scaling. Fixed costs, revenue grows with usage.

The companies that win long-term are the ones that decouple their pricing model from their cost structure. You can charge however you want — per seat, per usage, per tier — as long as your underlying cost is fixed and predictable.

The Playbook: Ship, Learn, Optimize

Here is the sequence that works:

1. Month 1-3: Ship AI features using an API. Price them however gets adoption fastest (usually bundled or freemium). Do not optimize costs yet. Learn how users actually use the feature.
2. Month 3-6: Measure actual usage patterns. Calculate your AI COGS per user. Identify whether you have a power user problem.
3. Month 6-12: If AI costs exceed 10% of per-user revenue, start the migration to fine-tuned models. Adjust pricing based on actual usage data, not assumptions.
4. Month 12+: Fine-tuned models running, costs are fixed, pricing model is validated by data. Now you can be aggressive — offer unlimited AI, bundle it everywhere, use it as a competitive moat.

The SaaS companies that treat AI pricing as a one-time decision will get it wrong. The ones that treat it as a data-driven optimization loop — start with APIs, measure, migrate to fine-tuned, adjust pricing — will build durable margin advantages.

Further Reading

• Building AI Features in Your SaaS: When to Stop Calling the OpenAI API — the full cost analysis for API-based AI features at scale
• The Hidden Cost of Per-Token AI Pricing — why per-token pricing is more expensive than it appears
• Shipping AI Features in Your SaaS Without an ML Team — how product teams can own AI features end-to-end