/ model drift / risk scoring / AI governance
Model Drift in Supplier Risk Scoring
Risk models need routine checks because supplier documents, fraud patterns, and business rules change.
A supplier risk model can drift even when the code stays the same. New document layouts, new payment patterns, new product categories, and changed analyst rules can make old thresholds less useful.
Track drift at the field level. Name extraction, address matching, certificate expiry detection, and beneficiary comparison may degrade at different times. A single overall accuracy number can hide a critical failure.
Use analyst corrections as early warning. If reviewers keep fixing the same extraction or overriding the same score, the model or rule set needs review. Corrections should feed a monthly quality check.
Refresh test cases with current documents. A model that performs well on clean license images may fail on marketplace screenshots, scanned invoices, or supplier-made PDFs from new regions.
Keep policy separate from model behavior. If the company changes its risk appetite for high-value orders, update workflow rules and reviewer guidance rather than hoping the model will infer the change.
Working checklist
- Monitor critical fields separately.
- Use analyst corrections as signals.
- Refresh test cases.
- Review thresholds monthly.
- Separate business policy from model output.