AI pricing in SaaS: credits, usage and outcome-based models compared
How SaaS vendors charge for AI: bundled seats, add-ons, credits, per-action and per-outcome fees, with one product priced three ways and the arithmetic shown.
Your team has shipped an AI feature, and finance wants to know how to charge for it. A flat seat price ignores that every AI request costs money to run. A pure meter scares buyers who want a predictable bill. You need to know the models vendors use, and what each one does to your customer’s invoice and to your margin.
AI pricing is how a SaaS vendor charges for features that run on machine-learning models: assistants, agents and generators. It sits on top of the subscription basics in What is SaaS?. The general price structures (seat, tier, usage, freemium) are covered in SaaS pricing models. This guide covers only the AI layer.
AI pricing in plain terms
Think of it like a phone plan. Some plans include calls in the monthly fee. Others give you a bundle of minutes and charge for extra. A few charge only when the call connects. The network costs the carrier money every minute either way. The plan only decides who carries that cost.
AI features work the same way. Each request to a model has a running cost, called inference. A credit is a prepaid unit that the vendor converts into actions at a published rate. An outcome is a finished result, such as a resolved support conversation, that the vendor defines and counts.
The five models vendors use
Every example below comes from the vendor’s own page, checked on October 17, 2026. AI prices change often, so check the date before you rely on one.
| Model | What you pay for | Public example | Who carries the usage risk |
|---|---|---|---|
| Bundled in the seat | The per-user price, AI included | Slack lists AI conversation summaries, recaps and AI search inside its per-user plans, with “Advanced AI” on Business+ (Slack) | Vendor |
| Per-user add-on | A fixed monthly fee per person on top of a base plan | Salesforce lists Agentforce add-ons at $125 per user per month (Salesforce); Microsoft sells Copilot as an add-on to eligible Microsoft 365 plans (Microsoft Learn) | Vendor, for heavy users |
| Allowance plus credits | A monthly allowance, then a price per extra credit | Atlassian Rovo includes 25 credits per user a month on Jira and Confluence Standard, then $0.01 per credit (Atlassian); GitHub Copilot Pro at $10 a month lists $15 of AI Credits ($10 base plus $5 flex) at $0.01 each (GitHub) | Shared |
| Per action | Each step an agent takes | Salesforce Flex Credits: $500 per 100,000 credits, 20 credits per standard action, which works out to $0.10 an action | Customer |
| Per outcome | Each result the vendor counts as done | Intercom prices Fin at $0.99 per outcome (Intercom) | Customer pays per result; vendor absorbs failed attempts |
The rows run from most predictable for the buyer to most closely tied to use. Salesforce shows that one vendor can sell several at once: the same page lists per-user add-ons, Flex Credits and a $2 per conversation option.
What each model means on an invoice
Bundled and per-user prices give your customer a fixed bill. The vendor pays for every heavy user out of the same fee. That works while average use stays low.
Credits split the risk. The included allowance covers normal use, and the meter starts above it. Two details decide how this feels in practice. First, whether allowances pool: Atlassian combines an organization’s allowances “into a single shared pool”. Second, what happens at the limit: with extra usage turned off, Atlassian pauses credit-consuming features until the monthly reset. GitHub lets Business and Enterprise admins decide whether paid usage beyond the allowance is allowed; if it is not, Copilot pauses until the next cycle.
Per-action pricing bills the agent’s work, not its success. An agent that takes eight steps to do a three-step job costs more than twice as much.
Per-outcome pricing bills results, so the definition of a result is the contract. Intercom counts an outcome when the customer confirms the issue is resolved, when they do not ask for more help after Fin responds, or when Fin completes a workflow, including handoffs. It charges once per conversation. Its option for teams on another help desk adds a minimum monthly commitment, “e.g. 50 outcomes”.
Worked example: one product priced three ways
Example: a labeled hypothetical. A help-desk vendor sells an AI agent to a customer with 20 support agents. The AI handles 3,000 conversations a month and resolves 1,800 of them, or 60%. Each conversation takes 5 model actions on average, and each action costs the vendor $0.04 in inference.
- Vendor’s inference cost: 3,000 x 5 x $0.04 = $600 a month, in every option.
- Option A, per-user add-on at $30: 20 x $30 = $600. AI margin: $600 - $600 = $0.
- Option B, per action at $0.10: 3,000 x 5 = 15,000 actions x $0.10 = $1,500. Margin: $1,500 - $600 = $900, or 60%.
- Option C, per outcome at $0.90: 1,800 x $0.90 = $1,620. Margin: $1,620 - $600 = $1,020, or 63%.
Now change one input at a time.
| Scenario | A: per user | B: per action | C: per outcome | Vendor cost |
|---|---|---|---|---|
| Base: 3,000 conversations, 60% resolved | $600 | $1,500 | $1,620 | $600 |
| Volume doubles: 6,000 conversations, 60% resolved | $600 | $3,000 | 3,600 x $0.90 = $3,240 | $1,200 |
| Quality drops: 3,000 conversations, 30% resolved | $600 | $1,500 | 900 x $0.90 = $810 | $600 |
Read the rows. When volume doubles, option A loses $600 a month ($600 - $1,200), because the seat price does not move with use. When quality drops, option B still bills $1,500, while option C falls to $810 and leaves the vendor $210. For the buyer, the cost per resolved conversation in the base case is $600 / 1,800 = $0.33 under A, $1,500 / 1,800 = $0.83 under B and $0.90 under C.
So the seat model protects the buyer’s budget and exposes the vendor. Per-action pricing exposes the buyer to an inefficient agent. Per-outcome pricing ties the vendor’s revenue to the agent’s quality, but only as far as the outcome definition is honest.
Questions to ask before you buy or set an AI price
- What exactly is one unit: a credit, an action, a conversation or an outcome? Get the definition in writing.
- How many credits does one task use? Salesforce charges 20 credits for a standard action and 30 for a voice action.
- Do included allowances pool across users, or does each person keep their own?
- At the limit, does the feature pause or bill overage? Who decides?
- Is there a minimum commitment?
- Can the conversion rate change? Figma’s pricing page says “Included AI credits are subject to change”, and Salesforce marks its page as subject to change.
For how these meters fit into renewals and proration, see SaaS subscription.
Common mistakes
- Bundling AI into the seat without measuring use. In the example, doubling volume turns a break-even add-on into a $600 monthly loss. Why it happens: seat prices are easy to sell and easy to forecast. Instead: meter AI use per account internally, even if you do not bill for it.
- Comparing credit prices without the conversion rate. One cent per credit means nothing until you know how many credits a task uses. Why it happens: the credit price is on the pricing page, the conversion is in the docs. Instead: convert every offer to a cost per finished task.
- Accepting an outcome definition you have not read. Intercom’s definition counts a conversation where the customer does not ask for more help, and a completed handoff. Why it happens: “pay only for results” sounds safe. Instead: sample billed outcomes each month and check that they were resolved.
- Ignoring what happens at the limit. A paused agent in the middle of a busy week can cost more than the overage. Why it happens: limits only matter once you hit them. Instead: decide in advance whether a pause or an overage bill hurts more, and set the budget control to match.
Sources
- Intercom, Pricing, checked October 17, 2026.
- Salesforce, Agentforce pricing, checked October 17, 2026.
- Atlassian, Rovo pricing, checked October 17, 2026.
- GitHub, Copilot plans, checked October 17, 2026.
- Slack, Pricing; Figma, Pricing, checked October 17, 2026.
- Microsoft Learn, License options for Microsoft Copilot, checked October 17, 2026.
In three sentences
SaaS vendors charge for AI in five ways: bundled in the seat, as a per-user add-on, as an allowance plus credits, per action, or per outcome. In the example, the same agent earns $600, $1,500 or $1,620 a month depending on the model, and doubling usage turns the seat add-on into a $600 loss. Before you buy or set a price, convert every offer to a cost per finished task and read the unit definition and the limit behavior in writing.
