87. An invoice-matching model worked well at launch, but exception rates have doubled. What would you investigate?
MediumAI & data governance
Technical
Show model answer
Compare recent suppliers, document formats, currencies, business rules and system changes with the launch baseline. Distinguish deteriorating model performance from genuine changes in transaction quality or the way exceptions are recorded.
Analyse error impact, validate proposed fixes and use fallback review where risk warrants it. Keep versioned test cases and monitoring so a retrained model cannot be declared improved solely because it fits the latest failed examples.
What interviewers listen for
- Check workflow and data changes
- Separate recording from performance effects
- Validate fixes with retained tests