AI that does half of your recruiter's job

The useful boundary does not separate easy tasks from hard ones; it separates verifiable tasks from decisions about people. Posting a role, acknowledging an application, offering interview slots or spotting that someone is already in the database are actions whose result can be checked at a glance, and automation there returns time at no cost to quality.

Extracting data from a CV, normalizing job titles or summarizing evidence from three interviews is a different category: AI accelerates a great deal, but the output needs a visible source. And rejecting a person, ordering the final shortlist or communicating a decline should not be automated at all.

A system that learns from history reproduces the patterns in that history, including the ones the firm would not want to repeat. Review therefore cannot be a one-off check at go-live: it has to be a routine with a date in someone's calendar.

The reasonable minimum is to compare progression and exclusion rates between groups every quarter, review which fields actually drive the recommendations, and analyse when recruiters override the system and why. A difference does not by itself prove discrimination, but it does require investigation before the use is widened.

A hundred percent acceptance rate for suggestions is a warning sign rather than a success: it means nobody is reviewing. The useful indicators are different and all obtainable from the system itself.

The four that best reflect real improvement are time to a defensible shortlist, profiles considered versus profiles that genuinely fitted, the share of corrections that end up fixing a data point at source, and retention of hires at six months. That last one is the least watched and says the most about process quality.