Any implementation should use this as a supplement only, not replacement for humans. For example a human watching 10 checkouts and quickly manually reviewing scenes flagged as suspicious. Any implementation should also account for the fact that loss will still occur and account for that in the RoI.
The fact that the technology is very error prone is not surprising (to anyone that actually builds AI tech). The issue here sounds like poor implementation and unrealistic expectations from management on what AI truly can and can’t do.