Evaluation should begin with the intended decision and user outcome. A model benchmark may be useful, but it does not establish whether the complete workflow is fit for its actual purpose.
A practical evaluation covers:
Task quality
Does the workflow produce the required output with sufficient accuracy, completeness and usefulness?
Source and data quality
Does it use the right information and preserve access boundaries?
Tool behaviour
Does it call the permitted tools with the correct parameters and stop before unauthorised action?
Human oversight
Are uncertain and consequential cases routed to the right person with enough context?
Failure handling
Does the system fail visibly, safely and recoverably?
User experience
Can the intended user understand limitations, review the output and challenge it?
Evidence
Does the workflow retain enough information to reconstruct the decision without storing unnecessary sensitive content?
Acceptance criteria should be written before final testing. Failed criteria should produce remediation or a decision not to deploy.
DDAI view: deployment should be a documented decision based on the whole workflow, not enthusiasm for selected outputs.