False positives dominate
75% of flagged payments were genuine: 3,000 of 4,000. That is about 3% of the 99,000 genuine payments; the two measures answer different questions.
Step 6 of 12Risk flags
Compare with the red-team review
Sign in to save your workCompare your analysis with the red-team review. Each flag is a line of enquiry, not a conclusion.
Unknown false negatives
Missed fraud is not measured, so neither the benefit nor the harm of the model can be balanced properly.
Essential-payment harm
Blocks can leave people without accommodation, food or medicine; aggregate figures hide severity.
Inaccessible verification
Phone-only verification excludes customers with hearing impairments and others.
Control not operating
'Verify' sometimes arrives after a decline, so the designed safeguard does not work in practice.
Silent harm
Customers who do not complain are never reviewed; repeated friction and lost customers go unseen.
Analyst judgement unchecked
Releases are based on whether a customer 'sounds genuine', with no sampling or quality check.
Unowned model change
Quarterly supplier updates have no assigned approver, testing or rollback plan.
Unverified supplier claim
'95% accurate' and 'up to 40%' are not evidenced for Northbridge's customers.
Data not used
A travel notice existed but was not an input; data governance decides what the model can see.
What did you find, and what did you miss?
Which of these did your earlier analysis already cover?
Which did you miss? What does each change about your assessment?
Answers save automatically and sync to your account when you are signed in.