What Should You Charge For When AI Does the Work?

This information is correct as of November 2026. Growth Method has made every attempt to ensure the accuracy and reliability of the information provided.

For two decades, SaaS pricing rested on two assumptions: serving one more customer cost you almost nothing, and the value a customer got grew with the number of people using your product. AI breaks both. Every task now carries a real compute cost, and when an AI agent does the work, a customer can get more value from fewer people.

This guide is for the marketers who own packaging, the pricing page and the story around them. It covers what you charge for: seats, usage, credits, outputs or outcomes. For how customers get in (freemium, free trials, reverse trials), see our guide to SaaS pricing models. For how buyers perceive a price once they see it, see price perception.

What a pricing model decides

Your value metric is the unit a customer pays for: a seat, a token, a credit, a piece of work or a business result. Choose a metric that grows as the customer gets more value and that your costs can live with, and pricing mostly takes care of itself.

Here is how the options compare.

The state of play

Pricing is changing fast, and three data sets show how:

• More frequent pricing changes. PricingSaaS, which tracks the top 500 SaaS and AI companies with public pricing, logged more than 1,800 pricing changes in 2025, about 3.6 per company. The number offering credits rose from 35 to 79, a 126% increase. Its co-founder Rob Litterst summed up the year: “2025 will be remembered as the year when seemingly everybody lost confidence in their pricing.”
• Hybrid is now the most common model. In Kyle Poyar’s 2026 State of B2B Monetization report, a survey of about 230 software companies, 37% had hybrid pricing, up from 25% a year earlier. Early-stage companies under $5M ARR leaned most on flat fees, and companies over $150M ARR held on to per-seat pricing.
• AI-native companies price differently. In the same data, 76% of AI-native companies had variable or hybrid pricing, against 49% of SaaS companies. Asked which pricing model investors prefer, only 15% of respondents said flat-fee or seat-based. The rest said hybrid (35%), outcome-based (26%) or usage-based (24%).
• Margins are thinner. In Bessemer’s State of AI 2025, the fastest-growing AI startups it studied averaged about 25% gross margins. A second group with steadier growth averaged about 60%, which Bessemer describes as slightly below SaaS peers. The sample was 20 companies, so treat the figures as directional.

Model costs are falling. Guido Appenzeller of a16z found that for a model of the same capability, inference cost falls about 10x a year. But the amount of work per task is rising. Cursor explained its 2025 pricing change by noting that “the hardest requests cost an order of magnitude more than simple ones.” Cheaper tokens do not mean a cheaper customer.

Why AI breaks seat pricing

Seat pricing worked because headcount was a fair proxy for value. Ten people using a tool got more from it than two people. When an AI agent drafts the content, answers the support ticket or runs the report, that link breaks. The customer who automates the most gets the most value, and under seat pricing pays you the least.

The other assumption fails on cost. A seat used lightly and a seat that triggers thousands of model calls cost you very different amounts but pay the same. Under a flat price, your heaviest users can become your least profitable.

Seats are not finished. They still fit products where people do the work, and many AI pricing models keep a seat or platform fee as the base. But seats are no longer a safe default for a product where AI does a meaningful share of the work.

The models, one by one

Seats

Seats are predictable for buyers, easy to forecast and well understood by procurement. They still suit collaboration tools and products where AI assists a person rather than replacing one. Many vendors have folded AI into their seat price; PricingSaaS reports that Notion, Slack and Loom bundled former AI add-ons into their plans, often with a price increase.

Usage (tokens)

Charging for raw consumption matches your costs perfectly and suits developer tools whose buyers think in tokens. For everyone else it is a poor customer-facing price. Most buyers cannot predict how many tokens a task will use, so they cannot predict the bill, and every invoice shows them exactly what the model provider charges you.

Credits

Credits give a vendor one meter for many different AI features: an image costs one number of credits, a prototype another. That flexibility explains the growth. Figma, HubSpot and Salesforce all added credit models in 2025.

The problem is that a credit can mean two things at once: a share of your costs, or a share of the value delivered. Figma’s model, where an image costs between 5 and 25 credits depending on the model chosen, leaves customers to work out which trade-off is worth it.

Credits work better as a wrapper than as the unit itself. HubSpot prices each AI capability on its own metric (support on resolutions, prospecting on leads), then converts every action into credits that customers buy in a pack and draw down across products. Poyar compares it to a retainer with a law firm: you estimate the overall size up front and true up later.

“There’s an inherent tension between credits as cost-based or value-based. A cleaner approach would be dual-track monetization: delineate value (the platform) and cost (the tokens) into separate buckets.”

Kyle Poyar, Growth Unhinged

Outputs

An output is a countable unit of work: an image generated, an email sent, an action an agent takes. Customers can see it, check it and compare it with what the same work would cost done by hand. Salesforce introduced Flex Credits for Agentforce in May 2025. Instead of $2 per conversation, customers pay $0.10 for each action the agent performs.

The risk with outputs is rationing. If every unit costs money, customers use fewer of them, even when using more is how they would get the most value. Elena Verna, who leads growth at Lovable, puts it sharply: “Every time you monetize an input, you give customers a reason to ration the exact behavior that leads them toward value.” So price the unit closest to the result, not each step on the way to it, and use generous tiers so a normal month never hits the cap.

Outcomes

Outcome pricing charges for a business result. It is the most talked-about model and the hardest to run. The best-known examples come from customer support, where the result is easy to define. Zendesk launched outcome-based pricing for its AI agents in August 2024, charging only for issues the AI resolves on its own. Intercom charges $0.99 per outcome for its Fin agent, where a resolution counts if the customer confirms the answer or leaves without asking for more help.

Outside support, few results are that clean. Revenue, pipeline and retention all have many causes, and the vendor and the customer will each credit their own side. The experiments that do exist are at the very top of the market: Poyar reports that OpenAI began testing outcome-based pricing with select enterprise customers, and that Cognition offered to fund up to $10 million of usage until its AI delivers real engineering value.

“For everyone else, outcome-based pricing is, for all intents and purposes, a myth. There’s a telemetry gap around what counts as the outcome, whether it occurred, and whether your product caused it.”

Scott Woody, CEO and founder, Metronome (now part of Stripe)

Woody is talking about vendors without monopoly-like power or very large contracts. His recommendation for the rest is outputs. If you still want to test outcome pricing, Poyar’s CAMP framework is a useful filter: the outcome needs consistency, attribution, measurability and predictability.

Hybrid

Most AI products end up here: a platform fee for access and value, plus a usage component that covers variable AI cost. Clay’s March 2026 pricing split its single credit into two meters: Actions for using the platform, and Data Credits for the data Clay buys on the customer’s behalf. After its 2025 change, Cursor’s Pro plan included $20 of frontier model usage a month, with more available at API rates. Gumloop passes model and compute costs through at cost and adds an 8% orchestration fee.

Hybrid has one weakness. Poyar found that “companies with hybrid pricing are generally the happiest, yet struggle to explain their pricing to customers.” Two meters need twice the explanation on your pricing page.

AI cost and your margin

Whatever you charge for, decide who carries the model bill. There are three common answers:

1. Include it in the price. Simplest for buyers, but your margin moves with every model price change and every power user. It suits light, predictable AI use.
2. Pass it through. Charge AI usage at cost, with or without a small fee. The platform price stays honest and model price swings stop hurting your margin. Customers see a variable line on the bill, so give them spend limits and clear usage reporting.
3. Bring your own key. Customers connect their own model provider account and pay it directly. Technical buyers like the control. Many marketers will find setting up an API key a hurdle, so offer it alongside pass-through, not instead of it.

The cheapest margin lever is the model itself. Much of the AI work in a typical SaaS product is drafting, summarising and classifying, which rarely needs a frontier reasoning model. Route each job to the cheapest model that does it well: proprietary models where quality or agent reliability earns the premium, and hosted open-weight models for the rest. An AI gateway makes that routing a configuration change rather than a rebuild.

How to choose your value metric

Five questions, in order:

1. Who does the work, people or AI? If people, seats still fit. If AI does a growing share, move part of the price to what the AI produces.
2. Can you count a unit the customer recognises? Resolutions, launches, documents and actions are units a buyer can picture. Tokens and compute minutes usually are not. If you can count it and the customer values it, price on the output.
3. How much does your cost vary between customers? If your heaviest customer costs 10 times your lightest, a flat price will hurt your margin. Add a usage component or pass AI costs through.
4. Can the buyer predict the bill? Finance teams buy predictability. Pair any usage meter with tiers, caps, included allowances or spend limits.
5. Can you prove the outcome? Only price on results if you can measure them, show you caused them and agree the definition with the customer up front. Otherwise price the output that leads to the result.

Make a usage price feel predictable

“You only pay for what you use” sounds good until the first overage bill. If you add any usage component, Poyar suggests four structures that keep the bill predictable:

• Hard budget caps at the account level, and ideally per user too.
• An annual drawdown. The customer commits to a level of usage for the year and spends it flexibly, like a gift card.
• A platform fee with usage included, plus a charge beyond it. The included allowance gets customers using the product before they think about the meter.
• An adaptive flat rate. The customer commits to a tier and uses as much as they like with no overages, then trues up at renewal. You carry the cost of any extra usage in the meantime.

Buyers want proof before they commit, too. HubSpot added trials for its AI agents, which lengthened sales cycles. On the Q2 2026 earnings call, CEO Yamini Rangan explained why: “Customers adopting AI want proof of value before they commit and predictability in what it costs.”

Write down your principles

Before you pick any numbers, write the answers down as a short set of principles. Verna argues every company needs a pricing constitution: one document that sets how everyone makes packaging decisions. Faster shipping means more features, and every feature raises the same questions about which plan it belongs in. Without agreed principles, each one becomes its own negotiation.

Her core principles, with the wording Lovable uses:

• Charge for value, not for friction you created. “Never manufacture pain to monetize it.”
• Price the outcome, not the inputs. “Monetize the outcome. Not the steps required to get there.”
• Leave some value uncaptured. If something spreads the product or brings people back, keep it free: “Don’t tax your own distribution.”
• Keep it simple. One primary pricing metric, two if you must, and no more than four plans (Free, Good, Better, Best), each for a different customer need.

A small group from product, engineering, marketing, sales and finance then rules on the edge cases. Verna’s test of whether the document works: it should cost you something, such as a feature you could have charged for and chose not to.

Make your pricing readable by AI

Before a buyer reaches your site, they may already have asked ChatGPT, Gemini or an AI agent what you cost. Two studies show how often the vendor’s own page loses that conversation:

• Pricing pages are rarely the first source. Kyle Poyar and Nikolas Laskaris analysed 7,675 AI responses about the pricing of the Cloud 100 companies, using data from Profound (which sponsors the newsletter and employs Laskaris). Company pricing pages appeared in 46% of answers but came first in only 12%. Vendr, Reddit and G2 filled the gap. Plaid did best, and its billing docs were cited more often than its pricing page.
• Pricing is where agents get stuck. Kevin Indig and Siteline tested agents on 100 B2B products, asking each to answer from the vendor’s own site. Agents managed this 92 to 93% of the time for security and integrations questions but 79% of the time for pricing. Hidden prices, calculators, toggles and prices that load only with JavaScript all sent agents to third-party sites instead.

The fixes are cheap, and marketing owns most of them:

• Publish real prices as plain text in server-rendered HTML. For custom tiers, explain what drives the price instead of only saying “contact sales”.
• Keep every plan, price and limit on one canonical pricing URL, and mark legacy plans clearly so old prices stop circulating.
• If you use a slider or calculator, show the same tiers as a text table.
• Back the pricing page with billing docs and an FAQ covering overages, minimums and cancellation.
• Add schema.org Product and Offer markup with the price and currency. In Indig’s readiness test, this alone raised one page’s score from 73 to 93.
• Check robots.txt does not block AI crawlers such as GPTBot, ClaudeBot and PerplexityBot.

The same work helps your AEO: an answer engine can only quote the price it can read.

Change your pricing without a backlash

At 3.6 changes per company per year, you will change your pricing again. How you roll it out matters as much as the new price.

In June 2025, Cursor moved its Pro plan from request limits to included usage at API rates. Customers were surprised by the charges, and within three weeks CEO Michael Truell published an apology: “We recognize that we didn’t handle this pricing rollout well, and we’re sorry.” Cursor refunded unexpected charges from the previous three weeks.

Clay did the opposite in March 2026. It let existing customers keep their legacy plans, promised to contact anyone the new pricing would save money, and said up front what the change would cost the company.

“We expect this update to cost us roughly 10% in revenue. But we’re betting this sets our ecosystem up to grow even faster. What customers pay should reflect how they actually use the product, and for a long time it hasn’t.”

Varun Anand and Karan Parekh, Clay

A checklist for your next change:

1. Let existing customers stay on their plan, or choose when to move.
2. Explain the change before the first new bill, not after it.
3. Show usage in the product, in the same unit you charge for.
4. Offer spend limits and alerts on any variable component.
5. Contact customers the new pricing would save money, before they find out themselves.
6. Update the pricing page, billing docs, FAQ and schema markup on the same day, then ask AI assistants what you cost a week later.
7. If you move to usage pricing, check that sales compensation still rewards what you want. Poyar warns that reps paid only on bookings will oversell the first deal.

What to ignore (for now)

• Outcome pricing outside clear cases. Unless the result is as countable as a resolved support ticket, you will spend more time arguing about attribution than you gain from alignment.
• Tokens on the invoice. Meter tokens internally. Woody puts it plainly: token billing is “useful infrastructure and, for almost every company, a bad customer-facing pricing model.”
• Credits for everything. One meter across many features is convenient for the vendor and confusing for the buyer. Litterst expects the pendulum to swing “back toward simplicity and predictability.”
• Copying an AI-native company’s price page. Their costs, buyers and margins differ from yours. Copy the reasoning, not the numbers.
• Investor fashion. Usage pricing is popular with investors right now. Poyar’s advice: “Don’t change your pricing because of your investors. Change it because of your customers.”

Where this is heading

Outputs will become the default unit. They are the one metric a buyer can picture, a vendor can count and nobody can dispute, and agents make more work countable every year. The likely standard is a modest platform fee for access, plus a price per unit of work, with AI costs passed through or included up to a limit.

Pricing pages will be read by machines first. An AI agent shortlisting vendors will skip a price it cannot read and take one from G2 instead. The vendors with clear, plain-text, well-documented pricing will win more of those shortlists than vendors with clever calculators.

And pricing will change continuously. That makes the principles more important than any single price. Write them down, keep the rollout fair, and treat each pricing change like a marketing campaign: plan it, launch it, analyse what it did to conversion and revenue, then adjust.

The bottom line

AI makes every task cost money and lets customers get value without adding people, so seat-only pricing gets weaker as AI does more of the work. Charge for something the customer can count and values, cover your variable AI cost explicitly, keep the bill predictable, and publish all of it in a form both people and agents can read.

About Growth Method

Pricing for AI work is a live question for us too. Growth Method is free to start, and Pro is $99 per workspace a month with the AI included and no overage.

Growth Method is the practical way to use AI in marketing. Our agents suggest campaigns from your own data, then help you plan, launch and analyse them, with no AI vendor lock-in, security taken care of and humans in the loop. It isn’t tied to one AI provider, and every campaign shows its real AI cost from recorded usage.

We are on-track to deliver a 43% increase in inbound leads this year. There is no doubt the adoption of Growth Method is the primary driver behind these results.

Laura Perrott, Colt Technology Services

If you want to treat your next pricing change as a campaign you plan, launch and analyse, get started here.

Frequently asked questions

Is per-seat pricing dead for AI products?

No, but it is weakening. Seats still work when people do the work and value grows with headcount. They break when AI agents do the work, because a customer who gets more done with fewer people pays you less. Most AI-era pricing keeps a platform fee, sometimes per seat, and adds a usage-based component on top.

What is the difference between output-based and outcome-based pricing?

An output is a countable unit of work the product produces, such as a generated image, a resolved conversation or an action an agent takes. An outcome is a business result, such as revenue generated or cost saved. Outputs are easy to count and hard to dispute. Outcomes are harder to measure and attribute, which is why Metronome CEO Scott Woody calls outcome-based pricing a myth for most companies.

Should AI usage be passed through at cost?

Passing AI usage through at cost, or letting customers bring their own API key, keeps your platform price honest and protects you from model cost swings. The trade-off is that customers see a variable line on the bill and you earn no margin on AI. Some vendors add a small fee on top of pass-through costs instead; Gumloop adds 8%.

What are credits in SaaS pricing?

Credits are a prepaid currency that customers spend across different product actions, each priced at a set number of credits. They let a vendor charge for varied AI features with one meter. The risk is that credits blur cost and value, so customers struggle to know what anything costs. The PricingSaaS tracker found the number of top SaaS and AI companies offering credits rose from 35 to 79 during 2025.

How do I make my pricing page readable by AI agents?

Publish real prices as plain text in server-rendered HTML, keep every plan and limit on one canonical pricing URL, explain usage-based pricing in words rather than only in a calculator, add schema.org Product and Offer markup, back the page with billing docs and an FAQ, and check that robots.txt does not block AI crawlers.

How do you change SaaS pricing without upsetting existing customers?

Let existing customers keep their plan or choose to move, tell them before the first changed bill rather than after, show usage clearly in the product, add spend limits, and contact customers proactively if the new pricing would save them money. Cursor’s 2025 pricing change shows the cost of skipping these steps: it ended in a public apology and refunds.