Historical case study

From HR voice agent to evidence-based assistance

The original prototype conducted recruitment screening calls and produced scores and a recommendation. The pipeline worked technically; that evaluation was also its regulatory risk.

Original scope

A voice agent asked five application questions. A second model scored communication, motivation, and role fit and returned Hire, Maybe, or Reject.

Core finding

A recruitment evaluation does not become low-risk through friendlier wording. Recommendations, rankings, and person scores influence decisions and should be treated as high-risk functionality.

EU AI Act, Annex III ↗

Robust redesign

predefined job-related criteria instead of open personality judgments
transcript correction and source evidence for every extracted claim
abstention when evidence is missing instead of an invented score
no emotion, language, personality, or origin inference
no automatic rejection and a mandatory human decision
versioning, audit logs, quality metrics, and documented risk management

Product decision

The public portfolio demo now uses a different conversation scenario. The active system reviews product concepts. The HR artifacts remain in the monorepo but are isolated from every live route.