Six-week pilot
A working assistant, tested on real support questions before rollout.
In a blind comparison with a custom-built assistant on a leading general-purpose model, it won 19 of 20 questions. This describes that comparison, not a general accuracy rate.
