We want to prove our AI feature works. What should we test?

Missing task-specific evaluation

Define success for the real user task before choosing metrics. Build a representative, appropriately protected test set with expected outcomes and meaningful failure cases. Evaluate correctness, unsupported claims, access boundaries, usability, and cost rather than relying only on impressive demonstrations.

What to do now

Define task outcomes and unacceptable failures
Create reviewed representative test examples
Measure results and inspect failure patterns
Set release criteria and retest after changes

Recommended action flow for missing task-specific evaluation: Define task outcomes and unacceptable failures then Create reviewed representative test examples then Measure results and inspect failure patterns then Set release criteria and retest after changes. Escalate when the feature cannot meet agreed quality or safety criteria on relevant cases.

When to escalate

Escalate when

The feature cannot meet agreed quality or safety criteria on relevant cases.

Who decides

The AI service owner and relevant authorized reviewer

Example

An extraction tool is tested on missing fields and contradictory dates, not just clean sample documents.

Mistake to avoid

Selecting only examples on which the demonstration already succeeds.

Prevent a repeat

Re-evaluate after changes to models, prompts, tools or source data; keep important decisions accountable to people.

Suggested ApprovedFlow setup

Tasks
Documents
Approvals

Track evaluation and review tasks, and keep approved evidence with the work; do not claim the platform provides an unverified AI agent or automated diagnosis. See the platform