Pricing model
Services that contain AI cannot be sold safely on the fixed-price, fixed-outcome basis that has traditionally worked for software projects. CUSCOE uses a three-layer structure instead, documented in the Portfolio so that sales, Customer Success and finance all describe it the same way.
Why fixed price fails
Fixed price works for traditional software because the cost of delivery is predictable and support costs after go-live are stable. That assumption fails for AI in three ways.
- There is no natural end point. An AI service needs permanent monitoring and tuning, so there is no point at which the work is finished.
- Infrastructure costs move. The provider usually resells cloud hosting and inference from vendors who can raise prices during a contract.
- End users drive consumption. Token use depends on how long and how complex users' interactions are, which neither the provider nor the vendor controls.
The three layers
| Layer | What it covers | How it is sold |
|---|---|---|
| 1. MVP | The initial build. | Fixed price, but the MVP must be deterministic by design. AI can be used to decide the next best action, but that action must then hand off to a directed flow or to a human. |
| 2. Operational outcomes | Ongoing monitoring, evaluation, tuning and retraining, and further AI components with agreed outcomes once the MVP is live. | A recurring service agreement, normally a monthly retainer. |
| 3. Consumption | Cloud hosting and inference. | Billed on usage at cost plus a defined margin. |
Layer 2 is where most of the skilled work sits after go-live, and it is what pays for the Centres of Excellence to keep the service healthy. Layer 3 passes through the cost risk that the provider cannot control.
Who needs to understand it
Sales must be able to explain the three layers to a customer and build proposals on them. Customer Success must understand them well enough to manage margin on the account. Finance must forecast and recognise revenue on the same basis. See Sales and Finance.
Evidence status
Evaluated against an illustrative scenario. Not yet tested in a live organisation.