How to Price AI Features
How to price AI features when inference cost varies by customer: credit systems, outcome pricing, margin floors, and what to do when token prices fall again.
On this page 7 sections
- Why per-seat pricing breaks when the agent does the work
- Four pricing models, with the margin maths
- Setting a margin floor and a cost per account ceiling
- How to design a credit system buyers can actually budget
- Handling heavy users without punishing adoption
- Repricing when model costs drop 60 percent
- When to fold AI back into your core tiers
- Frequently asked questions
The short answer
AI features turn a fixed-cost software business into a variable-cost one, which breaks per-seat pricing whenever the agent does work a seat used to do. Four models are available: bundle into existing tiers, sell an AI add-on, sell credits or a token pool, or charge per outcome such as a resolved support ticket. Launch as a metered add-on with generous included credits, set a margin floor and a cost-per-account ceiling, then fold it into tiers once cost per account is predictable.
Key points before you start
Software companies have spent twenty years enjoying a beautiful property: the marginal cost of one more user was close to zero. AI removes that. Every summary, every agent run, every retrieval call has a real cost that lands in your cost of revenue, and it varies by a factor of fifty between your lightest and heaviest accounts. Pricing has to absorb that variance, and per-seat pricing cannot.
Why per-seat pricing breaks when the agent does the work
Because your revenue metric shrinks as your product improves. That is the whole problem in one sentence.
Take a support tool priced at 100 dollars per agent per month with 50 agents at an account, so 5,000 dollars monthly. Ship an AI that resolves 40 percent of tickets. The customer now needs roughly 35 agents. Your revenue drops to 3,500 dollars, and it dropped because you delivered value. No amount of clever packaging fixes a metric that moves in the wrong direction.
This is why the support category moved first and fastest. Intercom priced Fin per resolution rather than per seat. Zendesk restructured around automated resolutions. Salesforce introduced conversation-based pricing for Agentforce alongside its per-user licences. In each case the company chose a metric that grows when the AI does more work, not less.
Not every product has this problem. GitHub Copilot remained a per-seat product because it makes an existing developer faster rather than replacing developers, so the seat count does not fall. The test is simple: does successful AI adoption at your customer reduce the number of units you charge for. If yes, change the metric. If no, per-seat with an add-on is fine.
The test most teams skip
Model your top ten accounts at 100 percent AI adoption. If revenue falls in more than three of them, your pricing metric is broken and no discounting policy will save it.
Four pricing models, with the margin maths
Here is the honest comparison. Costs assume a mid-market product where a typical AI action costs you somewhere between 2 and 15 cents in inference at current model prices.
| Model | How it works | Gross margin behaviour | Best for | Main risk |
|---|---|---|---|---|
| Bundle into tiers | AI included, no separate line | Unbounded downside. Heavy accounts can run negative | Mature features with predictable usage | One power user destroys account margin silently |
| AI add-on | Fixed monthly fee per seat or per account | Capped upside, variable margin per account | First launch. Clean usage data | Adoption gap. Many buy it and never use it |
| Credits or token pool | Pool of credits consumed by actions | Margin protected by design if credits priced above cost | Multi-action products with varied workloads | Buyers cannot forecast spend and stall |
| Outcome pricing | Charge per resolution, per task, per qualified result | Highest alignment, highest variance | Products with a countable, auditable result | Disputes over what counts as an outcome |
Run the maths on a specific case. Suppose your AI add-on is 20 dollars per seat per month and the average seat performs 150 AI actions monthly at 6 cents each. Cost is 9 dollars, revenue is 20, gross margin on the add-on is 55 percent. Acceptable, but the average hides everything. The top decile of users runs 900 actions, costing 54 dollars against 20 in revenue. That account is losing you 34 dollars per heavy seat per month.
This is why per-seat AI add-ons quietly fail at scale. The pricing works on the mean and breaks on the distribution. If you price on averages, model the 90th percentile before you launch.
50% to 70%
Typical gross margin band for AI-heavy SaaS features at launch, against a 75 to 85 percent traditional SaaS benchmark
Aggregated practitioner reports, saas-marketing.net estimate
Setting a margin floor and a cost per account ceiling
Two numbers, decided before launch, written down where finance can see them.
The margin floor is a blended gross margin below which the AI line is not allowed to sit. Sixty percent is a defensible starting point for a product whose core software margin is in the low 80s, because it blends to something the board can live with. Pick yours and commit.
The ceiling is a hard inference cost per active account per month. It exists so that an account consuming 40 times the mean triggers an alert rather than an end-of-quarter surprise. Set it at roughly three times your modelled mean cost, and route breaches to a named person rather than a dashboard.
Instrumenting the guardrails
- Log cost per action at the event level
Token counts and model used, tagged with account id, stored in your warehouse. You know it works when you can produce a cost per account for last month in one query.
- Model the distribution, not the average
Plot actions per account at the 50th, 90th and 99th percentile. Price against the 90th, not the mean.
- Set the floor and the ceiling with finance in the room
Written into the pricing doc with a named owner. If nobody in finance signed it, it is not a guardrail.
- Alert at 80 percent of the included pool
In-app and by email to the admin. The goal is a conversation, never a surprise invoice.
- Review cost per active account weekly for the first two quarters
Weekly, not monthly. Model behaviour changes and prompt changes both move cost faster than a monthly cadence can catch.
- Define the repricing trigger in advance
For example, a sustained 50 percent drop in unit cost converts to a 50 percent increase in included credits at the same price.
Editable CSV worksheet
Save your marketing measurement plan
Keep a worksheet for your inputs, assumptions and next actions. You can also print the calculation directly from your browser.
How to design a credit system buyers can actually budget
Credits fail for one reason: nobody can picture what a credit is. A pool of 50,000 credits is not a number a procurement lead can defend in a budget request.
Fix it by anchoring credits to recognisable units. One credit equals one document summarised. Five credits equal one agent run. Twenty credits equal one full report generated. Then publish a table showing what a typical month looks like for a small, medium and large team, with an example.
| Plan | Included credits | Example monthly workload | Overage rate |
|---|---|---|---|
| Starter | 2,000 | 400 document summaries, 100 agent runs | $0.02 per credit |
| Growth | 12,000 | 2,000 summaries, 800 agent runs, 100 reports | $0.015 per credit |
| Scale | 60,000 | 8,000 summaries, 5,000 agent runs, 700 reports | $0.01 per credit |
Three rules make credit systems survivable. Publish the overage rate before anyone hits it, because a surprise overage is a churn event and a support ticket in the same message. Let unused credits roll over for one period only, which reduces the fear of over-buying without creating an unbounded liability. And never change what a credit buys without notice, because that is a price rise wearing a technical costume and customers read it exactly that way.
The mechanics overlap heavily with conventional metered pricing, so if you have not built one before, work through the usage based pricing guide and model the tiers in the usage based pricing simulator before you publish anything.
The credit anti-pattern
Pricing credits directly in tokens. Buyers do not know how many tokens a task takes, you will change models next quarter, and the number is meaningless to anyone outside engineering. Abstract to tasks and keep the token maths internal.
Handling heavy users without punishing adoption
Your heaviest AI users are usually your most valuable accounts and your best references. Cutting them off mid-month to protect margin is the worst available trade.
A better sequence. Soft alert at 80 percent of the included pool, in-app and to the admin. Automatic overage at a published rate, so usage continues and the invoice is predictable. For the top 2 percent of consumers, a commercial conversation at the account level within the month, usually resulting in a committed higher tier at a better unit rate. Nobody gets switched off.
There is one case for a hard stop: free trials and free tiers, where abuse is real and the account has no commercial relationship to protect. Cap those firmly and make the cap visible from the first screen.
Where you place the AI capability across plans is a packaging decision as much as a pricing one, and the usual rules in feature gating still apply. Gate on volume, not on capability. A Starter customer who cannot use the AI at all never becomes a Growth customer, because they never experienced the thing you want them to pay for.
Newsletter launch list
The Friday SaaS Marketing Brief
Join the list for the upcoming SaaS Marketing Brief. Get the marketing planning worksheet immediately.
Repricing when model costs drop 60 percent
They will drop. Inference costs per unit of capability have fallen dramatically and repeatedly since 2023, and pricing built on the assumption that today’s cost is permanent has been wrong every single year so far.
You have three options when your unit cost halves, and only two are good.
Cut the headline price. This feels generous and is usually a mistake. Prices are sticky downward, competitors match instantly, and you have permanently reset the anchor for a category where value is still rising.
Increase what is included. Same price, twice the credits. Customers experience this as a gift, your margin improves on accounts that do not increase usage, and you keep the price anchor. This is the option I would take in almost every case.
Hold everything and bank the margin. Defensible for a quarter or two while you fund the next capability. Indefensible once a competitor announces a doubled allowance, which they will.
Whichever you choose, decide the trigger in advance and tell the account team what it is. Repricing decided reactively, in the week a competitor announces, is how products end up with four incompatible legacy plans and a billing system nobody understands.
When to fold AI back into your core tiers
Launch as a metered add-on. Fold into tiers once three conditions hold: cost per active account is predictable within roughly plus or minus 30 percent, attach rate on the add-on is above about 40 percent, and the feature is table stakes in your category rather than a differentiator.
The last condition is the one people get wrong. While AI is a differentiator, a separate line item makes the value visible and gives sales something to sell. Once every competitor has it, the separate line item becomes friction in a deal and a reason for procurement to ask what they are paying extra for. At that point bundling it into your Growth and Scale tiers, and raising those tiers modestly, converts a defensive feature into a pricing increase nobody argues about.
The broader question of what metric you should charge on at all is worth settling before any of this, and the two comparisons that matter most are seat based vs usage based pricing and value based vs cost plus pricing. Cost plus is particularly tempting with AI because you finally have a real marginal cost to point at, and it is still the wrong anchor. Charge for the resolved ticket, not for the tokens it took.
If you are also building the go to market around the AI capability, the positioning side is covered in marketing an AI native SaaS product and the launch-specific version in marketing an AI native SaaS product. Adoption data from those launches is what tells you when folding into tiers is safe. Sector-level context on how fast metered models are spreading sits in the usage based pricing adoption research, and the fundamentals live in the main SaaS pricing strategy hub.
Next step: pull last month’s inference cost by account, sort descending, and look at the top ten. If any of them cost you more than they pay, you have a pricing decision to make this quarter rather than next year.
Editable CSV worksheet
SaaS Pricing planning worksheet
A practical pricing planning worksheet: decisions, owners, evidence and next actions.
Frequently asked questions
Why does per-seat pricing break with AI features?
Because the agent does the work the seat used to do. If your AI resolves 40 percent of support tickets, the customer needs fewer support agents, so they buy fewer seats. Your revenue falls precisely because the product got better. Any pricing metric that shrinks as your value grows is misaligned, which is why support and SDR tools moved to outcome or usage metrics first.
What is credit based AI pricing?
Customers buy a pool of credits that get consumed by AI actions, with different actions costing different amounts. It decouples price from raw token counts, which buyers cannot forecast, and gives you room to change models underneath without repricing. It works when one credit maps to a unit the buyer recognises, and fails when credits are an abstraction nobody can budget against.
What gross margin should an AI SaaS feature target?
Traditional SaaS runs 75 to 85 percent gross margin. AI-heavy products commonly land in the 50 to 70 percent range at launch, and public company disclosures have shown AI-driven cost of revenue pressure across the sector. Set a floor you will not cross, typically 60 percent blended, and treat any account below it as a pricing bug to fix rather than a customer to keep quietly.
Should you bundle AI into existing tiers or sell it separately?
Sell it separately first. A separate add-on gives you usage data, a clean margin read per account, and the ability to change price without touching your core packaging. Bundle it into tiers once cost per active account is predictable within a narrow band, usually after two or three quarters, at which point bundling becomes a competitive advantage rather than an unbounded liability.
What is outcome based pricing for AI?
You charge per result rather than per usage or per seat. Intercom prices Fin per resolution, and Zendesk moved to charging for automated resolutions. The appeal to buyers is obvious because they only pay when it works. The risk sits with you: you need an auditable definition of a resolution, and a customer with unusually hard queries can be unprofitable at the list rate.
What happens to AI pricing when model costs drop?
Costs per token have fallen repeatedly and by large multiples since 2023. Plan for it. Write a repricing trigger into your model, for example that a sustained 50 percent cost reduction converts into more included credits rather than a lower headline price. Giving customers more for the same money protects revenue while still passing on value, and is far easier than raising a price you cut.
How do you stop heavy AI users destroying margin?
Not with hard caps, because heavy users are usually your best accounts and cutting them off at 3pm on a Tuesday is a churn event. Use soft alerts at 80 percent of the included pool, an automatic overage rate published in advance, and a commercial conversation at the account level for the top 2 percent of consumers. Monitor cost per active account weekly.
The saas-marketing.net editorial team Research and editorial
We research, write and maintain every page on this site. The library explains marketing decisions through practical frameworks, explicit assumptions and references. Corrections can be requested through the contact page.
Published September 11, 2026. Last updated .