Test oversight and assurance
Sign in to save your workDesign meaningful human oversight, then test whether a key control actually works. A documented control is not proof that it operates.
Meaningful human oversight
Before which consequential decision? What does the reviewer see beyond the AI output?
Who can override or pause? How are reasons recorded?
Training, time, workload, sampling of accepted outputs
Notice, explanation, review and redress
Answers save automatically and sync to your account when you are signed in.
SIU review: every referred claim is reviewed and decided by an investigator within 10 working days.
The SIU Manager says 'we review every case promptly'.
| Case | Outcome | Working days held | Customer updated? | Reason for referral recorded? |
|---|---|---|---|---|
| S-01 | Genuine | 8 | No | Score only |
| S-02 | Fraud | 12 | n/a | Yes |
| S-03 | Genuine | 27 | No | Score only |
| S-04 | Genuine | 6 | Yes | Yes |
| S-05 | Genuine | 34 | No | Score only |
| S-06 | Fraud | 9 | n/a | Yes |
| S-07 | Genuine | 19 | No | Score only |
| S-08 | Genuine | 11 | No | Score only |
| S-09 | Genuine | 5 | Yes | Yes |
| S-10 | Genuine | 22 | No | Score only |
Your test: How many cases met the 10-day target? What does 'Score only' tell you about the quality of human review?
Assurance result
Is the control designed well enough to address the risk? What is missing from its design?
Show your count from the sample and what it means
Condition, criteria, cause, effect; then the action, owner and date
Answers save automatically and sync to your account when you are signed in.
Auto-decline proposal
The Board meets on Thursday. I want approval to auto-decline claims above 0.95 from next month. Investigators agree with the model at that score 'almost every time'.
What would you tell the Board, and what evidence would you need before any automated decline?
Your response to the new development
What would you say, recommend and record?
Answers save automatically and sync to your account when you are signed in.