The short answer
Conventional software costs roughly the same whether a customer uses it heavily or barely at all. AI features do not, and that breaks the instinct to price flat and encourage usage.
Measure cost per customer before pricing, and build in a limit even if you never intend to enforce it aggressively.
Where flat pricing goes wrong
- A small number of heavy users consume most of the cost
- Usage is not proportional to willingness to pay
- Automated or scripted usage arrives without warning
- Costs rise with model changes you do not control
- The most enthusiastic customers become the least profitable
The last one is the uncomfortable part. In conventional software your heaviest users are your best advocates; here they may be losing you money.
Structures that hold up
| Model | Works when |
|---|---|
| Flat with a fair use limit | Usage is predictable and the limit is generous |
| Tiered by included volume | Usage correlates with customer size |
| Flat plus overage | Customers accept variable bills |
| Pure usage based | Value is clearly per unit of work |
| Per seat with per seat limits | Usage genuinely tracks headcount |
Tiered with included volume is the common landing point. It gives customers predictability and gives you a ceiling.
Instrument before you price
- Record cost per request, attributed to a customer.
- Look at the distribution, not the average, because it is usually skewed.
- Identify your most expensive customers and what they do differently.
- Model what happens if usage doubles at your current price.
- Set limits with headroom above normal use.
Averages hide the problem entirely. The mean cost per customer can look comfortable while the top few percent are unprofitable.
Reduce the cost, not just the price
Caching repeated work, using smaller models where they suffice, and avoiding recomputation all move the economics more reliably than raising prices.
Many AI products send the same or near-identical requests repeatedly. That is the cheapest saving available and it is frequently untouched.