Pricing AI Products: Seats, Usage, or Outcomes

Replit went from roughly $2M to $144M in ARR in about a year by 2025, and the growth ran on a pricing change: flat plans out, usage-based billing in¹. The margin story behind it is just as pointed. At one point in 2024 its gross margin was reportedly under 10%, and the company had disclosed operating at negative gross margins; after the pricing changes, margin climbed into the 20-30% range²¹. A product people clearly wanted was losing money on its heaviest users because the charge metric ignored what they cost to serve.

That gap is what AI features open up inside any per-seat product. Seat pricing assumes the next unit of usage costs you almost nothing, an assumption that held for two decades of SaaS and that every metered model call now breaks. The charge metric has become a product decision, and the PM is usually the person best placed to make it. Pick wrong and the failure shows up either as customers rationing their usage or as heavy users on a flat fee quietly consuming your margin.

Why per-seat pricing broke

AI features are compressing SaaS gross margins from the classic 80-90% down to 50-60%, because every model call burns real compute². The cost side of that story, including the cost-per-interaction model and the levers that cut it, is in the unit economics of AI features. This article covers the revenue side: given that the marginal cost exists, what do you charge for?

A seat collects the same fee whether its user triggers ten model calls a month or ten thousand, so revenue and cost stop moving together. Light users subsidize heavy ones, and the heavy ones are usually your best accounts. One 2026 analysis found that AI product companies that stuck rigidly to per-seat pricing carried gross margins about 40% lower on average than those that adopted usage- or outcome-based pricing¹. For reference, Salesforce, the canonical seat-priced business, runs gross margins around 77%¹.

AI-native companies have drawn the obvious conclusion. Bessemer finds them broadly abandoning seat-based pricing in favor of usage-, output-, and outcome-based models³. Seats still fit products where AI is a thin assist on work humans do. When the model does the work, the per-human unit stops mapping to anything, which is why agent products tend to price against tangible ROI, such as cost savings or output equivalent to a human's work, instead of per-seat fees³.

The five models, compared

Bessemer sorts AI charge metrics into three families: consumption (per API call, per token), workflow-based, and outcome-based (per successful outcome)³. On the ground, products ship five recognizable variants.

Three rows deserve a closer look.

Per-query pricing abstracts tokens into a unit the user can count: an image generated, a document summarized⁴. The buyer can predict their bill, which fixes the comprehension problem, but margins compress when usage skews toward complex tasks that burn more compute per unit than you priced for⁴.

Credits put a budget in the customer's hands while protecting your margin, because you set the exchange rate per task: a basic image costs 1 credit, a high-res video 50⁴. The unit stays stable for the buyer even as your underlying cost differs by task type.

Outcome-based pricing carries the highest possible margins and is the hardest to operationalize⁴. It gets its own section below.

The trade as you move toward outcomes

Bessemer puts it directly: "Your charge metric is a strategic statement, not just a billing decision."³ Moving from consumption toward outcome-based pricing is a deliberate trade: more value alignment, less cost predictability³. The customer pays when they get the result they bought the product for, and your revenue now depends on results you only partly control.

Consumption pricing fails in the opposite direction, and the failure shows up in behavior before it shows up in revenue. Leena AI initially charged on consumption, and customers became wary of using the product; the pricing model worked against the adoption it was supposed to fund³. That is the consumption model's standing risk. The more value a customer tries to extract, the larger their bill, so the meter trains them to hold back, and a customer who holds back gets less value and a harder renewal conversation.

Outcome pricing removes the anxiety, since extra usage costs the customer nothing until it produces a result. What it demands instead is operational. You need a measurable outcome (a resolved ticket, a filed claim, a collected invoice) and an attribution story for it. If the AI drafted the response and a human rewrote half of it, who earned the fee? Until both sides accept the answer, every invoice is a negotiation.

Bessemer's evidence points at verticals where the outcome is already quantified. Hybrid models work especially well in legal, where companies like EvenUp and Legora operate, because the customer measures the result in dollars before your product ever shows up³. The same logic underpins agent pricing: agents that take over a defined piece of work get priced against the cost savings or the human-equivalent output, because that ROI is tangible enough to put on an invoice³.

How to choose for your feature

Four questions sort most cases:

1. Is the outcome measurable and attributable? If the customer already tracks the result in their own reporting, outcome-based pricing is on the table. If not, it is off.
2. Does the buyer understand the unit? Engineers can budget in tokens. A support lead budgets in tickets. Price in the unit the budget owner thinks in.
3. Can you meter it reliably? Usage billing needs production-grade metering, with disputes, refunds, and edge cases handled. If your meter is an analytics event, you are not ready to bill on it.
4. Does usage track value? When the tenth query is worth as much as the first, usage pricing aligns well. When value front-loads (the first answer matters, the rest are refinements), a per-query meter punishes exploration.

If you cannot answer all four with confidence, default to a hybrid. A base subscription covers your predictable floor and the customer's budgeting needs, and a usage or outcome component on top scales with the cost you actually incur. Bessemer's finding is that hybrid models win when you are uncertain³, and most teams pricing their first AI feature are uncertain.

Whatever you choose also changes the shape of your revenue. Seat revenue is flat and forecastable. Usage and outcome components turn MRR into a range that moves with adoption, seasonality, and how successful your customers are. Before committing, model your current MRR against a scenario where 20-30% of revenue shifts to a variable component, and check what the swing does to your growth picture:

The charge metric also feeds your unit economics downstream. A model that lifts ARPU on heavy users changes LTV, which changes how much you can afford to spend acquiring them; the full chain is in CAC and LTV mastery.

Migrating a seat-priced product

Most readers are not pricing a new product. They are adding AI features to a seat-priced one, which is the harder version of the problem because existing customers anchored on the current bill.

Show the meter before you bill on it. Ship usage visibility (a dashboard showing each account's AI consumption) at least a quarter before any usage charge. Customers learn their number while it costs nothing, and you learn the real distribution. The accounts facing the biggest future bill are the conversations to have first, not the ones to surprise.

Start with new customers. Introduce the usage or hybrid plan for new signups while existing customers stay on current terms. This is the same test-with-new-customers discipline from the pricing strategy guide, and the grandfathering, notice-period, and lock-in mechanics there apply unchanged.

Cap before you charge. A generous included allowance with a paid tier above it monetizes the heaviest accounts, the ones costing you the most to serve, without touching the majority.

Watch usage, not just churn. Leena AI's lesson applies during migration too: if usage falls after the pricing change, the meter is creating anxiety even among customers whose bill did not move³. Usage that drops after a pricing change predicts churn months before the churn dashboard moves.

FAQ

Why are AI companies moving away from per-seat pricing? Per-seat pricing assumes serving an extra unit of usage costs almost nothing, and AI inference broke that assumption. Companies that held rigidly to seats saw gross margins about 40% lower on average than usage or outcome adopters, and AI-native companies are broadly abandoning seat-based models in favor of usage, output, and outcome pricing.

When does outcome-based pricing work for AI products? When the outcome is measurable in terms the customer already reports on, and attribution for mixed human-AI work is agreed before launch. It carries the highest possible margins and is the hardest model to operationalize. Verticals with dollar-quantified outcomes, like legal, suit it best.

What is credit-based pricing for AI features? The customer buys a budget of credits and each task draws against it at a rate you set: a basic image might cost 1 credit, a high-res video 50. Credits keep the unit stable for the buyer while letting you reflect the real compute cost of each task type.

Should I ever price in tokens? Only when your buyer is technical, typically for an LLM API or a developer product. Tokens give you high margin protection but score poorly on predictability and comprehension for everyone else. Most products should translate compute into queries, credits, or outcomes.

How do I move an existing product from seats to usage pricing? Show usage dashboards at least a quarter before billing on them, introduce the new model for new customers first, and keep existing customers grandfathered while you watch conversion and usage. The change-management mechanics are in the pricing strategy guide.

Sources

Footnotes

1. The Economics of AI-First B2B SaaS in 2026, Monetizely. Replit revenue and margin figures as reported by Monetizely, not from audited filings. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

2. AI Is Killing SaaS Margins. Outcome-Based Pricing Is How You Get Them Back, Fraction ↩ ↩² ↩³

3. The AI Pricing and Monetization Playbook, Bessemer Venture Partners ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

4. AI Pricing Models, Dodo Payments ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹