Over the past eighteen months the question we have been asked most often is this: how should an AI product be priced. SaaS pricing has been refined for twenty years. AI pricing has no settled answer yet. But several patterns are already clear.
Below are the four structural differences we have worked out while setting prices with AI startups and with existing SaaS companies adding AI features.
Difference 1: variable cost actually varies
In classic SaaS, variable cost was effectively zero. One more user changed infrastructure cost almost not at all. That is why “100% margin per customer” was the SaaS default.
AI products are different. Every additional call consumes GPU time, tokens, and external API cost. In short:
SaaS: build it once, sell it without limit.
AI: every single response costs money.
This one difference destabilises every pricing decision downstream. Allow unlimited use on a flat fee and heavy users eat your margin. Move to usage-based and customers hesitate because they cannot forecast their own cost.
Difference 2: seat count stops working as a unit
The SaaS standard was per-seat charging. One user, one licence. Simple and predictable.
Introduce an AI agent and the equation breaks.
- One user can push through the work of ten using an agent
- Work can even run in the background with no user present
- Hold per-seat charging in place and the company captures steadily less of the value it creates
The units emerging instead are per resolution, per task, per outcome. Per ticket handled, per report written, per enquiry resolved. Intercom’s Fin and Salesforce’s Agentforce are both moving this way.
Difference 3: price gets dragged along by cost
In classic SaaS, price followed value. Letting cost determine price was treated as a cardinal sin. With AI products it is easy to slide into exactly that:
- Model API prices rise
- Pressure builds to raise your own price
- Value has not changed, only price, and customers leave
- So you cut margin instead
The way out is to keep the charging unit tied to customer value, and to treat cost as a floor rather than a basis. A margin protection line belongs in the structure, not in a spreadsheet nobody looks at.
Difference 4: model performance improves fast, which pushes prices down
Inference costs have fallen sharply and continue to.
- The good news — margin improves on its own
- The bad news — customers know it too, and ask for a price cut
If your price is expressed in units the customer associates with model cost, you inherit that deflation directly.
So how should you price
Four patterns we recommend in practice.
1) Hybrid Floor
A base subscription with overage charging, plus a defined minimum margin line per customer. If usage spikes abnormally, a trigger brings sales in to renegotiate.
2) Outcome-Based with Floor
Charging on results, but with a minimum platform fee even when no result occurs. The customer keeps the incentive to adopt while your operating cost stays covered.
3) Tiered Capacity
Each tier states an AI processing allowance (credits, tasks) with overage charged on top. Customers get cost predictability, and value is still recovered from heavy users.
4) Value Sharing
Taking a percentage of the result the customer gains. The strongest alignment available, and the longest sales cycle. It requires a market mature enough to accept it.
| Pattern | Revenue predictability | Margin stability | Barrier to adoption |
|---|---|---|---|
| Hybrid Floor | High | High | Medium |
| Outcome with Floor | Medium | Medium | Low |
| Tiered Capacity | High | Medium | Low |
| Value Sharing | Low | Variable | Lowest |
Three traps to avoid
- Do not promise unlimited. In an AI product, unlimited is the most expensive promise you can make. Heavy users take the margin.
- Do not expose tokens as the pricing unit. Customers do not think in tokens. Abstract up to a value unit: tasks handled, documents produced.
- Do not promise a price frozen for a year. Cost and value both move quickly here. Put the possibility of quarterly adjustment into the contract.
Closing
AI-era pricing is SaaS-era pricing with two new variables added on top: cost volatility and measurability of outcome. No single model handles both. A hybrid that watches value, cost, and operability at the same time is not optional.
The most important part is not treating price as something you set once. For an AI product, price is an operational area.
📩 If you want to rebuild the pricing structure of an AI product, start with the Pricing Check.