The Platform PM

Common mistakes

Field Guide

AI governance and model risk

09

Common mistakes

1 min read · Chapter 9 of 10

What usually goes wrong

SymptomLikely causeFirst thing to check
Audit finds AI that isn't in the inventoryIntake forms only; vendor features and personal accounts missedProcurement records and security discovery against the register
People use unapproved tools after a rejection or a long waitIntake too slow, or the low tier too heavyTime from request to decision, by tier
Validation reports say little about the foundation modelBank methods applied to a model you can't inspectBehavior tests on your own data, and what they cover
Answers changed with no release on your sideThe vendor updated or retired the modelPinned versions; vendor change notices; re-test records
Conditional approvals never closeNo owner or date for the conditionsThe exception register and its ages
Human review approves everythingReviewers without time, training or authorityOverride rates by reviewer
A regulator or court asks for a decision you can't reproduceLogs and versions not kept after retirementRetention by rule; the archive
A customer rejects your evidenceA certificate with narrow scope or from an unaccredited bodyThe certificate's scope and issuer
An AI loss isn't coveredGenerative-AI exclusions in the policyExclusion wording; whether dedicated AI cover is needed