Part III — The Practitioner's Toolkit

Pricing and Packaging AI Features

Pricing AI features is one of the most consequential and least discussed decisions in AI product management. Traditional software features have near-zero marginal cost at scale — serving one more user costs almost nothing once the feature is built. AI features are different: every inference call costs money, and those costs scale directly with usage. A pricing model that ignores this reality can turn a successful AI feature into a margin problem.

What You Will Learn

• How AI feature economics differ from traditional software economics
• The four pricing models for AI features and their tradeoffs
• How to calculate the unit economics of AI feature pricing
• Packaging decisions that affect both monetization and user adoption

31.1 How AI Economics Differ from Traditional Software

Traditional SaaS products have high fixed costs (engineering, infrastructure) and near-zero variable costs at scale. The marginal cost of serving one additional user approaches zero once the product is built. This economic structure supports subscription pricing: charge a fixed monthly fee, serve unlimited usage, profit from scale.

AI features break this model. Every LLM inference has a real cost: input tokens cost money, output tokens cost money, embedding lookups cost money. The marginal cost of serving one additional AI feature use is not zero — it is proportional to the token volume of that use.

This creates several problems for teams that apply traditional pricing models to AI features:

Power users can destroy margins. A user who triggers the AI summarization feature 200 times per month costs 200x more to serve than a user who uses it once. In a flat subscription model, the high-volume user is cross-subsidized by low-volume users — which may be fine, or may be catastrophically uneconomic depending on the usage distribution.

Generous free tiers are expensive. Free AI feature tiers that are meant to drive adoption can become significant costs if the free usage volume is high. Budget caps must be built into free tiers from the start.

Cost scales with capability. More capable models are more expensive. A feature that uses GPT-4 costs more than the same feature using GPT-3.5, which costs more than a fine-tuned smaller model. Model selection is a cost decision as much as a capability decision.

31.2 Four Pricing Models for AI Features

Included in subscription. AI features are included in the product subscription at no additional charge, subject to a fair use limit. This model maximizes adoption and simplifies the pricing story, but requires accurate cost modeling to ensure that average usage does not destroy margins. Best for AI features where typical usage is low and variable, and where the AI capability is a differentiator that drives subscription conversion.

Usage-based add-on. AI features are priced separately from the base subscription, on a per-use or per-credit basis. Users pay for what they consume. This model aligns revenue with cost and prevents the margin destruction that flat pricing enables. The downside is friction: users who must think about per-use charges use features less, which reduces the adoption that AI features need to improve through the data flywheel.

Credit bundles. Users purchase a block of credits that are consumed by AI feature usage. Credits reset monthly or carry over. This model provides the cost predictability of subscriptions for users while maintaining a closer alignment between revenue and cost than pure subscriptions. It is the most common model for AI-heavy products because it balances user experience against economic sustainability.

AI tiers. Different subscription tiers include different AI feature access — lower tiers get limited AI (lower usage caps, less capable models), higher tiers get full AI access. This model segments users by willingness to pay for AI capability and allows the product to capture value from power users who need high-volume AI access. It works best when AI capability is a genuine differentiator between tiers, not a feature that users feel entitled to at all tiers.

31.3 Calculating Unit Economics

Before choosing a pricing model, calculate the unit economics of the AI feature.

Step 1: Measure cost per AI feature use. For LLM-based features: average input token count × input token price + average output token count × output token price + any embedding or retrieval costs. Do this measurement on a representative sample of real user inputs, not on an idealized example.

Step 2: Measure usage frequency distribution. What is the distribution of usage across users — median, 75th percentile, 90th percentile, 99th percentile? The distribution matters more than the average for pricing decisions, because high-volume tail users drive cost.

Step 3: Calculate cost at different pricing models. For subscription inclusion: what is the average monthly AI feature cost per user? What is it at the 90th percentile? At what point does the feature become uneconomic at the planned subscription price? For usage-based: what usage volume would a user need to justify the premium? Is that volume realistic for the target customer?

Step 4: Model the impact of improvement. AI inference costs decline over time as models improve and competition increases. A pricing model that is uneconomic today may be economic in 12 months. Include this trajectory in the analysis, but do not make pricing decisions based on projected cost reductions that have not materialized.

31.4 Packaging Decisions

Pricing is how you charge; packaging is what you charge for. Packaging decisions affect both monetization and user behavior.

Granularity of the AI unit. Is the chargeable unit a query, a document processed, an inference call, a session? The choice affects user behavior. Per-query pricing discourages exploratory use. Per-document pricing discourages processing large corpora. Choose the unit that aligns with how users think about value.

Bundle vs. à la carte. Bundling multiple AI features into a single tier simplifies the pricing story and increases perceived value. À la carte pricing allows precise capture of willingness to pay for each feature but creates complexity. Bundling is almost always the right choice for early-stage AI features.

Free tier design. If the product has a free tier, decide what AI access free users get. Options: no AI access (maximum monetization incentive, maximum adoption friction); limited AI access with a monthly cap (balances adoption and cost); AI access with clear quality differentiation (free tier gets a less capable model, paid tier gets the full model). The third option is viable when model quality differences are perceptible to users; it is not viable when users cannot tell the difference.

Enterprise pricing. Enterprise customers with high AI usage volumes require custom pricing negotiations. Build the capability for custom usage limits and enterprise-specific pricing from the product architecture — volume caps that are hardcoded rather than configurable will create implementation debt when the first enterprise deal requires negotiated terms.

Pricing and packaging decisions made at launch are hard to change without affecting existing customers. Invest in the analysis before launch rather than inheriting a model that requires a painful migration.