# Pricing and packaging AI in a per-seat world
A customer pays you $50 per seat per month. That customer's power user runs 400 AI queries a day, each one calling a large language modellarge language modelA Large Language Model is an AI system trained on vast text data to predict and generate language, enabling tasks like writing, summarizing, and answering questions.View full definition → that costs you real money in compute. Your gross margingross marginGross margin is the share of revenue left after subtracting the direct cost of producing goods or services, expressed as a percentage of revenue.View full definition → on that seat just went negative, and you did not even notice until the cloud bill arrived.
This is the quiet crisis inside SaaS today. The seat-based model that built the industry assumes a customer's cost to serve is roughly flat. AI breaks that assumption completely.
For twenty years, SaaS pricing was elegant. You paid per user (a "seat"), and each seat cost the vendor almost nothing to serve. A Slack user who sent 10 messages cost about the same as one who sent 1,000. Marginal cost was near zero, so seat pricing printed money.
Inference (the compute cost of running an AI model to generate an answer) changes the math. Every AI feature you ship has a variable cost per use, usually measured in tokenstokensA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → (the chunks of text a model reads and writes, roughly 3/4 of a word each).
Two customers on the same plan can now have wildly different costs to serve:
Under flat seat pricing, the heavy user is subsidized by the light user. When AI adoption is low, you absorb it. When adoption scales, and it always scales faster than you expect, your margins bleed.
Say your blended AI feature costs roughly $0.02 per action in model fees (an illustrative figure). At 200 actions per user per month, that is $4 of cost on a seat you may be selling for $30. If your target gross margingross marginGross margin is the share of revenue left after subtracting the direct cost of producing goods or services, expressed as a percentage of revenue.View full definition → is 75 percent or higher (typical for healthy SaaS), a single feature can eat a meaningful slice of it before you count support, hosting, or the rest of your product.
The trap: usage is not evenly distributed. A small fraction of "whale" users often drive the majority of consumption. Flat pricing means you cannot charge those whales more, but you still pay for them.
There is no single right answer. There is a spectrum, and each point trades off simplicity against margin protection.
Simple, predictable for the buyer, easy to forecast. But it exposes you fully to variable AI cost. Only safe when AI usage is light or capped hard.
Charge per unit consumed: per tokentokenA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition →, per query, per document processed. This is how the underlying model providers (OpenAI, Anthropic, Google) charge you, so it passes cost through cleanly.
The problem: buyers hate unpredictable bills. A procurement team cannot approve a line item that might be $2,000 one month and $9,000 the next. Usage-only pricing also punishes adoption, exactly the behavior you want to encourage.
This is where most successful AI SaaS is landing in 2026. The structure:
The buyer gets a predictable floor. You get a mechanism to charge heavy users more. Margins are protected because usage above the allowance is priced to cover incremental inference cost plus markup.
Notion, GitHub Copilot, and many CRMCRMCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.View full definition → vendors have shipped versions of this: a base subscription plus AI capabilities gated by credits, tiers, or add-on seats.
You cannot price what you cannot measure. Instrument every AI call. Log tokenstokensA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → in, tokenstokensA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → out, model used, and mapmapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → it to a customer. This is the single most important operational change AI forces on a SaaS finance team.
A simplified logging record looks like this:
{
"customer_id": "acme_corp",
"feature": "email_draft",
"model": "gpt-class-medium",
"tokens_in": 850,
"tokens_out": 320,
"est_cost_usd": 0.014
}Aggregate this and you can see your true gross margingross marginGross margin is the share of revenue left after subtracting the direct cost of producing goods or services, expressed as a percentage of revenue.View full definition → per customer, per feature, and spot the whales before they become a problem.
Do not bill customers in raw tokenstokensA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition →. TokensTokensA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → are an internal engineering unit; buyers cannot reason about them. Translate to a value metric the customer already cares about:
The best value metric scales with the value the customer receives, not just your cost. If more usage means more value to them, they will accept paying more.
Bundle enough included usage that the typical customer never thinks about limits. This keeps the experience feeling like flat SaaS for the 80 percent. Then set overage pricing that comfortably covers your marginal inference cost with margin on top.
Watch the gross margin floor. Price overage so that even at unlimited consumption, you never sell AI below your cost to serve. This is your insurance policy against runaway usage.
For a deeper primer on how model providers price the inputs you resell, see OpenAI's pricing documentation, which shows the per-tokentokenA token is the basic unit of text that language models process, often a word fragment, whole word, or punctuation mark rather than a single character.View full definition → structure your costs inherit.
🎬 [VIDEO: "How to Price AI Products" — https://www.youtube.com/results?search_query=how+to+price+ai+products — Overview of usage-based and hybrid pricing strategies for AI features]
Pricing is the number. Packaging is how you bundle and name it. Two moves matter most.
Do not make AI a separate line item customers can decline. If AI is a checkbox add-on, budget owners will cut it in a downturn. Weave AI into the core tiers so it becomes part of why customers stay. Reserve add-on packs for genuine overage, not baseline access.
Use tiers to segment by usage intensity. A "Pro" tier with a modest AI allowance and an "Enterprise" tier with a large allowance lets customers self-select into the cost band that matches their behavior. Heavy users naturally migrate up.
Knowledge check
1. Why did per-seat pricing work so well for traditional SaaS but break down with AI features?
2. Under a flat per-seat model with AI features, what best describes the relationship between light and heavy users?
3. A product team notices their AI feature has low adoption today and healthy margins, so they conclude flat seat pricing is fine. What is the flaw in this reasoning?
4. Select ALL correct answers about why AI inference costs are difficult to manage under seat-based pricing.
Select all the correct answers.
5. Select ALL correct answers describing what the '$4 of cost on a $30 seat' example is meant to illustrate.
Select all the correct answers.
If you already sell flat per-seat and now need to add usage components, this is a sensitive migration. Customers reasonably fear a price increase disguised as "innovation."
Practical guidance:
The companies handling this well treat usage transparency as a feature, not a threat. A customer who can see and manage their consumption trusts the pricing more than one who is kept in the dark.
One underappreciated consequence: your cost of goods sold now moves with customer behavior, not just your headcount or infrastructure choices. Finance teams used to forecasting flat hosting costs must now model variable inference as a function of adoption curves. Build scenarios for low, medium, and high AI adoption, and stress-test your margins in each. If a viral spike in usage would break your unit economics, you have a pricing problem, not a demand problem.