Traditional software has a cost structure pricing teams could almost ignore. Once a SaaS product is built, serving one more customer costs very little, so gross margins of 75% to 85% are common and pricing decisions focus almost entirely on value and competition. AI products don’t work that way. Every summary, analysis, answer, or automated action consumes compute that the vendor pays for, often through a model provider whose prices can change with a single announcement.
That shifts the job of pricing. It is no longer enough to ask what customers will pay. You also have to ask what each customer will cost to serve, how widely that cost varies, and what happens to your margin when the underlying costs move, up or down. Companies that skip these questions often discover the problem through their gross margin line, usually after they have locked customers into plans that are hard to change.
This article explains where AI costs really come from, why flat pricing hides serious margin risk, six guardrails that protect profitability, and how to decide what to do when model costs fall. The graphics throughout are designed to be saved and used as reference in your own pricing reviews.
Where the Cost of an AI Product Actually Comes From
When teams estimate the cost of an AI feature, they usually start and stop with model inference: the price per token or per call charged by the model provider. Inference is often the largest single cost, but it rarely tells the whole story. A production AI workflow typically also pays for retrieving and preparing data, calling external tools or APIs, orchestrating multi-step tasks, monitoring quality, and running the infrastructure that holds it all together.
The most underestimated cost is human review. Many AI products route a share of tasks to people, whether to check low-confidence outputs, handle escalations, or meet compliance requirements in regulated industries. Even when only one task in ten needs a person, the labor cost of that review, spread across every task, can rival the cost of the model itself.
In the example above, a document-review task priced at $1.00 costs $0.40 to deliver once every component is counted, producing a 60% gross margin. That is healthy for an AI product, but it sits well below what investors and finance teams expect from traditional software. It also means small changes in cost matter a great deal. If the review rate doubles, or the product moves to a more capable and more expensive model, the margin can fall by ten points or more without any change in price.
The practical takeaway is to calculate cost to serve per unit of value, meaning per task, document, conversation, or whatever unit your customers recognize, and to include every cost that scales with usage. That number becomes the foundation for every pricing decision that follows.
Why Flat Pricing Hides Margin Risk
Averages are dangerous in AI pricing. A plan that looks comfortably profitable across the customer base can be deeply unprofitable for the customers who use the product most, and those customers are often the most engaged and most visible accounts you have.
Consider a $500 per month plan with unlimited AI usage, where each task costs $0.10 to serve. A light customer running 300 tasks a month costs $30 to serve, a 94% gross margin. A typical customer running 1,000 tasks costs $100, which is still a strong 80%. A heavy customer at 3,000 tasks costs $300, and the margin drops to 40%. A power user running 6,000 tasks costs $600 to serve on a $500 plan, and the vendor loses money on every month of that relationship.
This spread is far wider than most SaaS teams are used to. In traditional software, a heavy user and a light user cost roughly the same to serve. In AI, usage can vary by a factor of 20 or more within a single plan, and cost varies right along with it. The problem compounds over time, because engaged customers tend to automate more work as they learn the product, pushing their usage steadily upward while their price stays fixed.
The fix is not to punish heavy users. They are often your best advocates and your strongest expansion opportunities. The fix is to design pricing so that revenue rises with the usage that drives cost, while keeping the experience predictable enough that customers are comfortable using the product fully.
Six Guardrails That Protect AI Gross Margin
No single tactic solves margin risk, but a combination of guardrails can keep profitability stable even as usage grows and model costs shift. The six below work together, and most successful AI pricing models use at least three or four of them.
Fair-use limits. Every plan should include a clear allowance, set generously enough that typical customers never think about it. The allowance protects you from the extreme tail of usage without making the product feel metered to the majority. If your median customer uses 1,000 tasks a month, an allowance of 2,500 covers nearly everyone while capping your exposure.
Overage pricing. Allowances only work if there is a clear path beyond them. Publish a per-unit overage rate upfront, show customers their usage inside the product, and alert them before they reach their limit. Overage revenue then converts your heaviest users from a margin problem into an expansion opportunity. For guidance on explaining usage units clearly, see How to Communicate AI Credit Models to Customers.
Model routing. Not every task needs the most capable and most expensive model. Routing simple requests to smaller, cheaper models and reserving premium models for complex work can cut inference costs substantially without customers noticing any drop in quality. This is an engineering decision with direct pricing consequences, so pricing and product teams should plan it together.
Committed volume. Offering lower unit prices in exchange for annual commitments improves revenue predictability and helps you plan capacity. Committed customers also tend to be more deliberate about adoption, which makes their usage easier to forecast.
A cost-linked pricing metric. The unit you charge for should rise with your real cost driver. If cost scales with the number of pages processed, pricing per user leaves you exposed, while pricing per document or per page keeps revenue and cost moving together. Choose a unit that customers understand and that tracks your cost reasonably closely.
Quarterly cost reviews. AI costs change faster than almost any other input in software. Recalculate cost to serve every quarter, and track gross margin by customer segment rather than as a single average. The average will hide the problem until it becomes large.
When Model Costs Fall: Cut the Price or Keep the Margin?
For much of the AI market, the cost of running a given level of model capability has fallen sharply over time. That is good news for margins, but it creates a pricing question every AI company eventually faces: when your costs drop, should your prices drop too?
There is no universal answer. The right move depends on two factors: how differentiated your product is, and how quickly competitors are cutting their own prices.
Hold the price when your product is well differentiated and competitors aren’t racing to the bottom. Customers pay for the outcome your product delivers, not the tokens behind it. If the contract review, support resolution, or sales insight is just as valuable as it was last quarter, there is no reason for the price to fall simply because your costs did. The savings flow to margin, which funds product investment.
Reinvest the savings when you are differentiated but competitors are cutting prices quickly. Rather than reducing your headline price, keep it steady and give customers more for it: larger usage allowances, access to more capable models, or new features. This protects your price anchor while making your offer stronger in comparison.
Share the savings when your product is less differentiated but the market is moving slowly. A modest price reduction keeps you competitive and signals good faith, while the remaining savings improve your margin.
Pass the savings through when your product is close to a commodity and competitors are cutting aggressively. In a price war over an undifferentiated offering, holding price mostly means losing share. Compete on cost, reliability, and service, and invest in building the differentiation that will let you escape this quadrant.
Whatever you choose, avoid tying your published price directly to your model costs. Customers who learn to expect a price cut every time a model provider lowers its rates will treat your pricing as a cost pass-through rather than a reflection of value, and that makes future increases much harder.
Putting It Together
Protecting AI margins starts with honest accounting: know your full cost to serve per unit of value, including human review, and track it by customer segment rather than on average. From there, build guardrails into your pricing structure so revenue grows with the usage that drives cost. A generous allowance, clear overage rates, and a cost-linked pricing unit do most of the work, and model routing and committed volume add further protection.
Finally, treat cost changes as strategic decisions rather than automatic price adjustments. Falling costs are an opportunity to widen margins, strengthen your offer, or win share. Which one you choose should depend on your competitive position, not simply on your supplier’s latest price list. For more on structuring AI pricing models, see Transitioning to Consumption Pricing for AI and How to Price AI Agents: Seats, Usage, or Outcomes?
Final Thought
In traditional software, pricing was mostly a conversation about value. In AI, it is a conversation about value and cost at the same time. The companies that keep healthy margins won’t be the ones with the cheapest models. They’ll be the ones that know exactly what each customer costs to serve and have built pricing that stays profitable as that number moves.
