GUIDE / AI

ONVUO / RESOURCES

Define AI’s role before choosing a model.

The first question is not which model to use. It is which task needs support and what result would be useful to the person doing it.

Describe the result precisely.

Summarizing a conversation, classifying a request, searching documents and suggesting an action are different tasks. They require different context and criteria. Define the inputs and expected outputs, including what should happen when information is incomplete or unclear.

Relevant information within clear boundaries.

Identify the sources the task needs, how current they must be and who may access them. The existence of information does not mean every user or model should see it. Access should remain connected to the person’s role and the work being performed.

Separate a suggestion from approval.

A model’s suggestion can be useful without being ready to execute. Define where human review enters, what the reviewer needs to see and how an output can be corrected. For sensitive tasks, the boundaries of action are part of the workflow’s design.

Test the cases that make the work difficult.

Do not rely on a single easy example. Use cases that reflect varied language, phrasing, missing information and contradictions. Review the effect of an error on the user, then define where assistance is appropriate and where the workflow should stop or escalate.