Migrating from Manual Rules to ML-Assisted Routing

Document the active rules, precedence, expected exceptions, and source-specific routing split before adding scoring. Tag known test sessions so later disagreements can be traced to model behavior rather than an undocumented rules change.

Collect scores and reasons while manual policy still controls the destination. Review disagreement samples by source, country, network, and device; aggregate agreement alone can hide a damaging error concentrated in one important segment.

Define which false positives require immediate review and which uncertain visits should remain on the conservative path. Thresholds should reflect the flow's destinations and traffic mix, not a number copied from another campaign.

Enable ML-assisted decisions for one bounded source or score band, keep explicit allowlists intact, and set a rollback owner. Avoid changing features, destinations, and thresholds in the same window because the result becomes impossible to attribute.

Replay acceptance fixtures, compare event samples with the baseline, and verify that operators can explain both manual and learned contributions. Complete migration only when rollback remains tested and routine monitoring can detect drift or a rising disagreement rate.