AI resume screening — has anyone actually quantified bias or accuracy at scale?

Our HRIS is pushing an AI resume-screening add-on. They cite "40% time savings" and "reduced bias." The bias claim makes me nervous (most of the bias literature on AI screening is negative) and the time-savings claim seems suspect.

For people who've deployed this: did you measure the actual accuracy vs. human screeners on the same job reqs? And how did you handle the EEOC / disparate-impact angle if HR challenged the model output?