An AI product is rarely just a model. A useful system usually contains several layers.
01
Experience
What the user sees
02
Application
Context tools and controls
03
Model
Prediction engine
A complete map
| Layer | Job |
|---|---|
| Experience | Interface, workflow and user feedback |
| Application | Instructions, context assembly, routing and permissions |
| Model | Generates or predicts from an input |
| Data | Source documents, records, examples and metadata |
| Tools | External capabilities the application may call |
| Controls | Identity, approvals, logging, limits and recovery |
| Evaluation | Tests whether the system meets its intended standard |
Changing a model can change quality, speed and cost. It does not automatically repair poor data, vague instructions or excessive permissions.
Vocabulary habit
When someone says "the AI knows our policies", translate it into system terms: the application may retrieve policy documents and place selected passages into the model context. The policy source remains outside the model.