--- title: "Variable Cost Has Entered the Chat: Pricing AI Features Without Killing Margin" description: "AI features turn pricing from a packaging exercise into an economics problem. If you ship before you model individual AI usage costs, your profit margins evaporate on contact with real usage." author: "Kody Everson" url: "https://theipp.org/insights/the-margin-mirage-why-ai-features-break-your-pricing-model-on-day-one" date: "2026-06-06T12:17:02.728Z" --- # Variable Cost Has Entered the Chat: Pricing AI Features Without Killing Margin ## Summary AI features turn pricing from a packaging exercise into an economics problem. If you ship before you model individual AI usage costs, your profit margins evaporate on contact with real usage. ## Main content For two decades, SaaS product managers have been able to treat gross margin as someone else's problem. Hosting was cheap, marginal cost per user approached zero, and pricing debates were really packaging debates: which features go in which tier, how to nudge expansion, where to set the seat price. The economics took care of themselves. AI features have ended that era, and most product organisations have not noticed yet. The moment you ship a feature that calls a frontier model, you have introduced genuine variable cost into a business model that was never designed to absorb it. Your margin story, the one in the executive deck, the one underpinning your CAC payback math, was built for a world where shipping more usage was free. It is not free anymore.. This is not a finance problem to be punted to the CFO. It is a product problem, because the levers that determine cost-to-serve, model choice, context length, retrieval strategy, caching, fallback logic, usage caps, are all product decisions. Pretending otherwise is how you end up with a hit feature and a collapsing gross margin in the same quarter. ## The Quiet Death of Zero Marginal Cost Classic SaaS unit economics rested on a comfortable asymmetry. A new user cost you a few cents in storage and compute. A power user cost you slightly more cents. The distribution of usage did not really matter, because the tail was cheap. You could offer unlimited plans, generous free tiers and flat per-seat pricing because the worst case was still profitable. The world of AI token costs completely changes this. A single complex query against a large model with a long context window can cost more than a user's entire monthly subscription. I have heard of internal data from multiple companies where the top 5% of users on an AI feature consume 40-70% of total AI usage spend. This is not an edge case, it is the predictable consequence of pairing flat-rate pricing with variable-cost delivery. ## Why Product Owns This Problem There is a temptation to frame this as a pricing team issue, or an infrastructure optimisation issue. Both framings let product off the hook, and both are wrong. Consider the levers that actually move cost-to-serve on an AI feature: - **Model selection.** Routing to 'Thinking' models versus smaller open-weight alternatives can change unit cost by 30x. This is a product decision, because it affects quality, latency and user trust. - **Context window discipline.** Whether you stuff 100k tokens into every call or retrieve precisely is a retrieval design decision that lives in the product spec. - **Caching and memoisation.** Which responses can be cached, for how long, with what invalidation rules, is a product policy question about freshness and personalisation. - **Fallback behaviour.** What happens when a user hits a cap? Degrade silently, prompt to upgrade, queue, refuse? Each option has revenue and retention consequences. - **Streaming versus batching.** Affects perceived performance and cost in opposite directions. None of these are infrastructure choices in the old sense. They are product choices that determine whether the feature has a viable business model. If your PMs cannot articulate the cost-per-call of the features they own, they are not managing those products. They are managing the user-facing half of them. ## The Three Pricing Failures Across AI product launches I have seen, three patterns dominate. ### 1\. Flat-rate pricing on variable-cost features The most common failure. A team adds an AI assistant to an existing seat-based product and leaves the seat price alone, or bumps it by 20%. The CFO sees expansion revenue. Six months later, finance discovers that the AI cohort has 30 points lower gross margin than the non-AI cohort, and the heaviest users are loss-making at any plausible retention curve. ### 2\. Usage caps without product design Teams retrofit caps after the cost shock, typically as a hard limit announced via email. The result is predictable: power users churn loudly, mid-tier users feel cheated, and the brand absorbs the reputational hit. Caps are a product feature. They need design, communication, graceful degradation and an upgrade path that feels like value, not punishment. ### 3\. Premium AI tiers priced on willingness-to-pay alone The opposite mistake. Teams price the new AI tier based on what they think users will pay, never reconciling that number with cost-to-serve at projected usage. This works until usage grows. Then the tier that looked like 80% gross margin in the pitch deck turns out to be 35% in production, because users did not just adopt the feature, they integrated it into daily workflow. ## What Good Looks Like The product organisations getting this right share a small number of practices. None of them are exotic. All of them require the discipline to treat cost-to-serve as a first-class product metric. **Instrument cost per call, per feature, per cohort, from day one.** If you cannot answer the question "what does this feature cost us per active user in the 90th percentile of users" within 30 seconds, you do not have observability, you have wishful thinking. **Build pricing into discovery.** When a team validates an AI feature, the evidence standard should include a unit economics hypothesis alongside the user value hypothesis. "Users want this and we can deliver it for less than $X per interaction" is the minimum bar. If you cannot answer this clearly, the discovery is incomplete. **Assign decision rights for AI usage budgets.** Someone needs to own the question of how much the company is willing to spend per user per month on usage for a given feature, and that someone should be the PM. Not the head of infrastructure, not finance. The PM, with finance as a partner. Without this, model upgrades happen because they are technically better, not because they pencil. **Design caps and tiers as product surfaces.** Hitting a limit should be a moment of value reinforcement, not a wall. The best AI products show users their consumption, explain what drove it, and offer a credible upgrade narrative. This is core product craft applied to a new constraint. **Hold yourself to outcome accountability that includes margin.** If your AI feature ships, the outcomes you commit to should include cost-to-serve targets, not just engagement and retention. A feature that drives engagement while destroying margin is not a success, it is a deferred problem. ## The Strategic Implication Zoom out, and this is bigger than any one feature. The companies that will win the next decade of AI-enabled software are not the ones with the best models. They will be the ones whose product organisations can operate fluently in a world where every shipped capability has a marginal cost, where pricing and product design are inseparable, and where the gross margin curve is something PMs read every week. This is a return, in some ways, to older disciplines. Hardware product managers have always lived this way. Telecoms product managers have always lived this way. Cloud infrastructure product managers have always lived this way. SaaS product management got a twenty-year holiday from variable costs, and that holiday is over. The teams still acting as if model economics are an engineering concern will find that their margin story does not survive contact with production. The teams that internalise this shift, that rebuild their discovery, prioritisation and pricing practices around cost-to-serve as a first-class concern, will be the ones with a defensible business when the AI feature gold rush settles into something that has to actually make money. Your pricing page is a product decision. It always was. AI just made the consequences impossible to hide. ## Related pages - [Insights](https://theipp.org/insights.md) - [Product Profile](https://theipp.org/tools/product-profile.md) - [Standards](https://theipp.org/standards.md)