T-Mobile's AI Got Worse With More Data

T-Mobile's AI Got Worse With More Data
T-Mobile's early attempt to improve AI customer service by loading large amounts of support documentation into agents backfired. Conflicting instructions caused agents to follow wrong paths, and a subsequent knowledge graph was too slow and hard to navigate. The company now uses a simple lookup table that maps common customer phrasings to the correct agent paths. T-Mobile also runs a second language model on every session to flag errors and customer frustration. That judge model unexpectedly revealed agents were hallucinating hold music when pausing to look up information.
Read the original article →