AI Pricing Models: Seat, Usage or Outcome, and Your Cash

AI companies price in three main ways: per seat, per unit of usage, or per outcome. Seat pricing bills a fixed fee up front, so cash arrives early. Usage and outcome pricing bill after the work is done, so cash arrives later and moves with volume. The most popular choice now is a blend of them.¹

I work with founders in the four industries I go real deep in: technology (AI and SaaS), healthcare and biotech, pharma, and pharmacy, and this question lands hardest on the AI companies in that group. I am not a CPA, so your auditor and your tax partner own the attest and tax questions.

Picture your next board meeting. Your lead investor asks what happens to cash if you move from seats to usage in January, and you have the pricing slide but not the cash slide. This is the walk-through I would give you over coffee so you can build that slide. Every cash and margin number below is made up (illustrative, not a client), so you can follow the math with a pencil.

Three ways to charge, and why most companies end up with a blend

Seat pricing is a model that charges a fixed fee per user, per month or per year. Usage pricing charges for what the customer consumes, like tokens, API calls or actions. Outcome pricing charges only when the product delivers a result you both defined in advance, like a resolved support conversation. Hybrid pricing is a base fee for access plus a usage or outcome charge on top.

If you are unsure which to pick, plan for the blend. Kyle Poyar's May 2026 survey of more than 230 B2B software and AI companies found 37% on hybrid pricing, up from 25% a year earlier.¹ Bain looked at more than 30 SaaS vendors adding generative AI and found about 65% layering an AI meter on top of seat fees, about 35% keeping seats and bundling AI into tiers, and none selling AI only as a separate add-on.² Solvimon calls hybrid the default rather than an experiment.³ Bessemer says it more bluntly: "Hybrid models (base subscription + usage/outcome tiers) win when you're uncertain."⁴

I would also assume you will change your model at least once, because your peers are doing it. ICONIQ found that 37% of about 300 executives building AI products plan to change their AI pricing in the next 12 months, and that 58% still carry a subscription or platform component, 35% charge by consumption and 18% by outcome.⁵ Salesforce Ventures heard the same from more than 300 executives.⁶ Cruxy's survey of 300 B2B SaaS chief executives has the line I keep repeating to founders: "Seat-based pricing was built for a world where humans did the work."⁷ Salesforce shows how fast this moves. It shipped three pricing models for Agentforce in roughly 18 months: $2 per conversation in October 2024, $0.10 per action in May 2025 and per-user licenses from $125 a month later that year.⁸⁹

Outcome pricing is the newest and the one everybody wants to talk about. Intercom charges $0.99 per outcome for its Fin agent on every plan.¹⁰ Zendesk says customers pay only for the issues the AI resolves on its own.¹¹ Sierra's Bret Taylor describes the same bargain: the customer pays when the agent resolves the issue and nothing when it hands the case to a person.¹² In my book that is the most honest model for the buyer and the hardest one on the seller's cash, which is the next thing to look at.

The same customer, four different cash balances

Here is the back of the napkin version, on illustrative numbers (not a client). Say your product is worth $120,000 a year to a customer, about $10,000 a month. The contract is signed on day 0, invoices go out at signing or at month end, and the customer pays 30 days after each invoice. Now look at day 90, when your board asks how the quarter went:

1. Seat: $120,000 billed annually in advance and paid on day 30. By day 90 you have collected $120,000, recognized $30,000 of revenue and hold $90,000 as deferred revenue.
2. Hybrid: a $36,000 platform fee billed annually in advance, plus about $7,000 a month of usage billed in arrears. By day 90 you have collected $50,000 ($36,000 plus the first two months of usage, $14,000) and recognized $30,000.
3. Usage: about $10,000 a month billed at month end and paid 30 days later. By day 90 you have collected $20,000 and recognized $30,000, so $10,000 sits in receivables.
4. Outcome: $1 per successful outcome, with 6,000, 8,000 and 10,000 outcomes in months 1 to 3 as the agent ramps. By day 90 you have collected $14,000 and recognized $24,000.

Same customer, same product and about the same annual value, and the day 90 cash runs from $120,000 down to $14,000. That gap is what I mean when I say each model moves your cash on a different day. Seat pricing funds your costs before you incur them. Usage and outcome pricing make you pay for the servers, the model calls and the people first, and collect two months later. Move a large customer base from seats to usage and you will feel that lag as a hole in one quarter, even though nothing is wrong with the business. That is why I never let a pricing change reach a board without a 13 week cash flow forecast sitting next to it.

What your auditor will ask about a usage contract

Seat pricing is the easy one. You bill ahead, hold the cash as deferred revenue and recognize it over the term. I cover the mechanics in what deferred revenue is in SaaS, and if your books are still on cash basis, cash vs accrual accounting is the place to start. Usage and outcome pricing are where I would call your auditor before signing the first big contract, because the treatment hangs on the contract wording.

With usage pricing, the first thing I check is whether the price is a plain fixed rate per unit. If it is not, revenue gets harder. Solvimon puts the starting point simply: "For usage-based revenue, the transaction price is mostly variable consideration, and you have to estimate it."¹³ A shortcut called the right-to-invoice practical expedient lets you recognize what you have the right to invoice when the rate tracks the value delivered, but Deloitte notes it is generally not appropriate with fixed fees, substantive minimums or varying prices.¹⁴ Declining tiers and minimum commitments with overage break it too.¹³ A minimum that covers a full year with overages bunched late likely has to be estimated and recognized over the term, and PwC separates that from a variable fee that is allocated to the day it relates to.¹⁵¹⁶

Outcome pricing adds a different question: what did you promise? If you promised continuous access, you recognize over time. If you promised a quantity of successful outcomes, each one is generally recognized as it occurs. A success rate measured across the whole term generally does not qualify for the allocation exception, while pricing that resets each period is more likely to.¹⁴ Prepaid credits add breakage when you expect some to go unused.¹³ My rule comes from a billing platform writer: "Constrain aggressively until usage history develops, especially for new AI pricing models."¹⁷

The public filings show where this ends up. Snowflake's 10-K says that unlike a subscription business it generally recognizes revenue on consumption, and it warns of limited visibility into future results.¹⁸ Twilio's 10-K for 2017 says customers are charged based on usage and most have no long-term obligation.¹⁹ I read both as forecasting warnings, not only accounting rules. I am not a CPA and your auditor decides the treatment, so bring them the contract language before it is signed.

Margin: who carries the compute

Margin is where seat pricing hurts AI companies, because a heavy user costs you real compute and pays the same fee. Bessemer's playbook puts AI gross margins at 50 to 60%, against 80 to 90% for SaaS.⁴ Here is the napkin version, on illustrative numbers (not a client). A customer pays $10,000 a month on seats, and at expected volume it costs you $4,500 to serve, which is a 55% gross margin. If that customer uses twice as much, your cost doubles to $9,000 and the margin falls to 10%. Under usage pricing the same doubling also doubles revenue to $20,000, and the margin stays at 55%.

Outcome pricing sits in between, and it can bite from the other side, because you pay for every attempt and bill only for the successes. Say the agent succeeds on 70% of attempts, so revenue is $10,000 against $4,500 of cost. If the success rate slips to 50% and the cost does not change, revenue falls to about $7,143 and the margin to 37%. That is why I would not offer outcome pricing until I had measured the success rate and the cost per attempt for a few months. Bessemer names the trade directly: outcome-based pricing gives the best value alignment but requires you to absorb the cost variability.⁴

I price against the cost line, not the competition. PostHog's team found its AI agent was improving almost every part of the product, so it decided to charge enough usage to cover its costs rather than maximize profit on the feature.²⁰ I cover the margin lines in AI gross margin: why yours is not 80% and the per-customer math in how to calculate unit economics.

What a bad pricing change looks like

The cleanest documented example is Cursor in June 2025. On June 16 it changed its $20 a month Pro plan from a request limit to $20 of usage at API rates, with the option to buy more.²¹ Users hit their limits fast, some were charged extra without understanding they needed a spend limit, and on July 4 the company apologized. In its own words: "we will refund any unexpected charges you may have incurred for usage over the past 3 weeks."²²

The pricing logic was not crazy. Simon Willison explained that one request can now consume a widely varying number of tokens, so Cursor passed those costs on.²³ What failed was cash and trust. Customers could not predict the bill, and the vendor ended up refunding money. GitHub began enforcing premium request limits in June 2025, with 300 a month on Pro and overages from $0.04 a request, and users complained.²⁴ It then announced a move to token-based AI Credits from June 1, 2026 because, in its words, "Today, a quick chat question and a multi-hour autonomous coding session can cost the user the same amount."²⁵ One Replit user said effort-based pricing "never ran me as much before but Agent 3 has been exceptionally high," about $1,000 in a week against $180 to $200 a month earlier.²⁶

My take is that none of these companies had the wrong model. They had a billing surprise: the customer found out what the new price meant when the invoice arrived. If you change pricing, tell customers what the bill will look like before it hits them, and give them a cap.

If you run a biotech, pharma, pharmacy or AI company

If you sell AI, everything above is your cash roll. If you buy AI tools, the usage model changes your close, because the bill arrives after the month ends and the cost has to be accrued before the invoice shows up. That is the same accrual discipline from my pharma work. Through a High Rock Accounting engagement, I served as the sole accountant on a clinical-stage biopharmaceutical client's month-end close, owning its preclinical, clinical, manufacturing and general accruals through its Nasdaq IPO. As a member of the client's accounting team, I spent 40 hours or more on accrual workpapers in the first week of every month. The lesson was simple: when cost follows volume and the bill comes later, the estimate is your job, not the vendor's.

I would carry that into any pharmacy, biotech or health tech company adopting AI: put each usage-priced vendor in a roll forward schedule, accrue the unbilled usage at month end and compare it to the invoice when it lands. For an AI company selling into health systems and pharmacies, slow procurement and annual budgets are a reason to keep a base fee in the contract. Those are my instincts, not data.

How I test a pricing change before I commit

I run it on paper first, then on a small group of customers. Build a 13 week cash flow forecast for the old price and the new price using your real invoice and collection timing, and find the week the balance bottoms out, as I walk through in how to build a 13 week cash flow forecast. Then run your lightest and heaviest customer through the new price, and pilot with one cohort under a spend cap. Keep a floor, because ICONIQ's own advice leans toward hybrid models with safeguards like annual commitments and tiered overages.⁵ Before you sign, call your auditor: a minimum with overage, a credit that expires or an outcome that resets only in aggregate can each change how revenue is recognized.

If you only do 4 things after reading this, do these:

1. Find the dip. Run one customer through each model at day 30, 60 and 90 and write down the cash.
2. Price against cost. Know your cost per unit and per attempt before you quote a price.
3. Keep a base fee. Unless you can show that cash and margin hold without one.
4. Warn customers early. Tell them what the bill will be before it arrives, and cap the downside.