AI Feature Pricing: Seats, Usage, Credits or Outcomes
AI feature pricing combines a familiar base, such as a seat or plan, with charges for usage, credits or outcomes. Pick one by how much cost varies per user.
Why is AI pricing different from traditional software pricing?
In classic software, serving one extra user costs almost nothing, so a flat price per seat has good margins. With AI, every query costs tokens and compute. A user who makes ten queries costs very differently from one who makes a hundred, even though both pay the same. That imbalance is the central problem of AI pricing. To understand why cost varies, read cost per token.
What are the pricing models?
Stigg describes six models: five appear in the table and the sixth, freemium, is covered below it. The table summarizes their tradeoffs with real product examples. Prices change often, so confirm the current one on each product’s official page before quoting it.
| Model | Main tradeoff | Example cited |
|---|---|---|
| Tiered plans | Fixed plans with feature and usage limits. Each tier is a set of entitlements your system must enforce | Claude Pro (20 dollars a month billed monthly) |
| Usage-based | Cost follows tokens or calls, but varies with every long response and retry | Anthropic API, billed per million tokens, with different prices for input and output |
| Credit pools | Simple to present, but heavy usage drains the balance fast; you need a ledger | Cursor Pro, described by Stigg as including 20 dollars of credits a month in 2026 |
| Outcome-based | Cost follows results; hard to attribute when one result uses several calls | Intercom Fin charges 0.99 dollars per outcome |
| Seat plus AI add-on | Cost and usage diverge: a few heavy users can drive most of the cost | Notion sold AI as a per-user add-on before bundling it into its Business plan, according to Growthoptix |
On Intercom: according to its pricing page, an outcome counts when the customer confirms the issue is resolved, stops asking for help after Fin responds, or Fin completes a workflow, and each conversation is charged only once. It is a good reminder that “outcome” needs a written definition.
Stigg also covers freemium, which carries its own risk: serving free users can be expensive and unpredictable, especially if few of them convert to a paid plan.
Which model should you choose?
The sources agree on one thing: the hybrid model, a monthly fee plus consumption charges, is where most AI companies land according to Growthoptix (April 2026). A frequent pattern keeps the base customers already know, such as the seat or plan, and adds credits or a usage charge for AI. Another is selling AI as a per-user add-on, with the risk that few people buy it.
A Cruxy survey of 300 B2B SaaS CEOs in the UK and US, published by SecurityBrief on April 16, 2026, found that 97% consider it likely they will retire seat-based pricing within two years, yet 94% said seat pricing still reflects their product’s value today. It is a useful reminder that changing the model takes time and is not always the priority. When you do change the model, plan a migration: tell current customers what changes, keep their existing balance or plan for a transition period, and check how the new charges look on their most common tasks before you switch on the new rules.
How do you design a credit system people understand?
- Show the balance before it is spent. A visible credit counter prevents surprises.
- Charge differently per task, and explain it. If a short query costs 1 credit and a long analysis costs 20, say so in the interface.
- Warn before service stops. A notice at 80% of the balance gives people time to decide.
- Offer a cap. A per-user or per-workspace limit protects both the customer and the margin.
- Record every deduction. Without an auditable record, the balance cannot be explained when someone asks.
Example
Illustrative example: a design product charges 15 dollars per seat. Its text generation feature costs about 0.40 dollars per user a month on average, but 5% of users generate 60% of the usage. The team keeps the seat price and adds 500 monthly credits per user, with a shared team pool and a notice at 80%. Heavy users buy extra credits, and the margin no longer depends on the goodwill of everyone else. The figures are invented to illustrate the logic.
How do you choose a model step by step?
- Measure AI cost per user and its distribution. If 5% of users explain most of the cost, a flat price will not work.
- Decide whether value can be measured per outcome. If it cannot, do not force that model.
- Start with a familiar base and add the variable charge for AI.
- Define credits or per-task charges in a public table.
- Review margin per customer every month, not just total revenue.
Common mistakes
- Charging a flat price for a feature with high variable cost.
- Changing the pricing model without warning current customers.
- Selling “unlimited” without a usage cap.
- Creating credits nobody can translate into real value.
- Attributing to one feature a cost that actually comes from several.
When NOT to charge per outcome
When an outcome depends on several calls, retries and human review, attribution is hard, and the customer may feel they are paying for something they cannot verify. It also makes little sense when the feature’s value is diffuse, such as a general experience improvement with no clear number. In those cases hybrid pricing is more honest. To decide whether the feature justifies its price, review AI feature ROI and AI product metrics.