Three things that fit
- Classification — routing, categorising, prioritising
- Extraction — documents and free text into structured fields
- Drafting — replies and summaries for a person to review
All three produce output a person can check in seconds, which is what makes them safe to add to a system people already depend on.
Call from a task, not a view
Model calls are slow and occasionally fail. Making one during a request means a user waiting and a request that sometimes times out.
- Queue the work, process it, store the result
- The interface shows pending, then the result
- Retries handled by the queue
- A provider outage delays rather than breaks
Store it as data on the model
The output should land as structured fields with a confidence value, not be regenerated each time it is displayed. That makes it queryable, auditable, correctable and cheap.
Add fields for the model version and prompt version too, so an output can be explained months later.
Keep the human path
| Requirement | Why |
|---|---|
| Review queue for uncertain output | Errors reach a person |
| In-place correction | Fixing must be easy |
| Corrections captured as data | They improve the system |
| A switch to disable the AI step | When something goes wrong |
| Original input preserved | Reprocessing and disputes |
Controls
A hard daily cost cap, a per-user cap, cost per item tracked, and prompts in configuration rather than in code so they can be changed and rolled back.
None of that is difficult, and all of it is what makes an AI feature safe to run in a system the business depends on.